Release notes

• KoderFewer MCP sign-in prompts. On macOS, Koder no longer asks you to confirm an MCP server's sign-in, such as with Touch ID, each time it connects to the server in the background. It asks only when you start the sign-in yourself.

• Desktop appNew ways to open a project. From Open project, you can now start a new folder, clone a repository from a link, or import one of your GitHub repositories.
• Desktop appCharts and diagrams in chat. Replies can draw charts and diagrams right in the conversation. You can show a chart as a table, copy its data as CSV, or download it as a PNG or SVG image.
• Desktop appLinks open beside the session. A web link you click opens in a browser next to the conversation, and you can import your sign-ins from another browser. Turn it off under Open links beside the session in Settings > General.
• KoderPlain or technical replies. In Settings > General, set Reply style to Technical or Plain and add Your role. Plain replies use everyday words, put the outcome first and show code only when you ask.
• KoderSetup from Codex, Cursor and opencode. Besides Claude Code, Koder can now read the MCP servers, commands, rules, skills and agents you set up in Codex, Cursor or opencode, where they live, without copying them. Only Claude Code is read by default, so add the others with linked.sources in your global config.
• Desktop appFewer keychain prompts on macOS. The desktop and terminal apps now share Koder's keychain items without asking, and Koder keeps them only in your login keychain. If you had made another keychain your default, Koder asks you once to sign in and approve your trusted folders and helpers again.

• Agent EngineSub-agent steps by execution id. When a team agent delegates to a sub-agent saved as its own agent, the delegate_task step now carries that sub-agent's execution_id instead of a copy of its steps, so fetch those steps with a result request on that id. Other sub-agents still list their steps under steps, as before.
• Agent EngineClearer run warnings. A run that succeeds after reaching a limit, such as the step limit or its budget, now gets a warning worded as a warning instead of "The run stopped ...". diagnostic_error_codes is deprecated: read error_details.code and warnings[].code instead.
• Agent EngineFewer cut-off answers. When you don't set an output limit, agents on aixplain models now send a default limit within each model's maximum, so answers and tool calls are no longer cut off by the platform's short default. A tool call that is cut off anyway goes back to the agent as a failed call it can retry, instead of failing the run.
• Agent EngineRuns get their full time budget. A run's budget.maxDurationSeconds now sets how long it can run, and up to a minute of agent setup no longer counts against that time. A hung tool call that doesn't change data now times out within the run's remaining time, so the agent can still answer instead of the run failing with RUN_TIMEOUT.
• Agent EngineSteadier tool calls. An aixplain tool or model call that fails because the connection dropped before the request reached the server is now sent again instead of failing. Actions marked as changing data are never resent.
• Agent EngineSkills load under heavy use. Skill packages are now shared across the engine's workers, so each skill is downloaded once instead of on every worker, and agents with many skills are less likely to hit the hourly skill download limit. If the limit is reached anyway, the run stops waiting on downloads and reports SKILL_DOWNLOAD_LIMITED, saying skills will load again after the limit resets.
• Agent EngineAttached files work without file tools. Files and folders attached to an agent now appear to it as a read-only agent_files/ folder (formerly reference_files/), and an agent without its own file tools gets read-only tools to list, read and search them. The agent can no longer change or delete the attached files, and two runs of the same agent no longer remove each other's copies.
• Agent EngineSession files carry over. In a session, files uploaded on earlier turns stay readable by the agent on later turns under inputs/, without being attached again.
• Agent EngineFixes. When a run reaches its step or context limit, its closing answer no longer fails on models with a smaller output limit, such as GPT-4o mini. Web page links an agent cites in its answer, such as map links, are no longer returned as attachments.
• Web appChat drafts and jump to latest. Each chat keeps its own unsent draft, and a jump-to-latest button takes you back to the newest message. Long drafts get a footer under the composer, and a message waits for its attachments to finish uploading before it can be sent.

• KoderScan for security bugs. Scan reads a repository for security bugs, explains each one and can propose a fix, from the Security view or with koder scan. It reads the code without running it, so each finding is a likely issue to review. In CI, it can write a SARIF report and fail the job at the severity you set.
• Desktop appEvery project's sessions in one sidebar. Sessions that are running or waiting on you sit at the top, and the rest group by date or project. The context meter now measures the model's whole window and marks where the chat gets summarized.
• OmniBuild agents with Omni. Omni shows agents as cards in the chat. Open one beside the chat to test it and change its setup, or build a new one with Omni, for aixplain agents and agents on this computer. A finished turn now folds its steps into one line, with the answer right under it.
• KoderAsk really asks. Under Ask, commands, file changes and network calls now wait for your approval. The old rules-only behavior is now called Rules, and Full access still asks before pushes, forced deletes, spending and secrets.
• KoderSteadier sessions. A message sent just as a turn ends is now answered, a hanging tool server no longer holds up a turn, and Auto and Full keep going on their own in the desktop app instead of stopping after the first reply.
• Desktop appCrash reports, if you want them. Koder can send crash reports to aixplain, off until you turn them on in Settings > General > Privacy. A report says what kind of error happened and where in Koder's code, never your messages, files or keys.

• Agent EngineAgents finish multi-step tasks. Agents no longer stop and answer after their first tool step, and GPT-5.x models no longer repeat tool calls they already completed in the same request.
• Agent EngineSteadier runs on GPT-5.x. Runs on GPT-5.x models that include image or file attachments given as URLs no longer fail. Team agents with a planner on GPT-5.x now produce a plan instead of answering the request directly.
• Agent EngineInspectors on task and planner teams. Step inspectors now take effect on team agents that follow tasks or use a planner, so an output an inspector blocks never becomes the answer. Abort, edit and rerun verdicts apply before the step's output is passed on.
• Agent EngineFixes. Failed steps now carry an error object with a code and message, and a failed task step names the step and the reason instead of a generic error. Summarizing a long history no longer cuts into a turn with several tool rounds.
• Web appWorkspace usage page. A new usage page replaces the Usage tab in workspace settings and breaks usage down by API key, with each key's owner and masked key in the picker. Credit purchases now note that credits expire 12 months after purchase.
• Web appSteadier chat. A chat turn that fails now shows a readable error and stays in place, and a reply still streaming when you reload picks up where it left off. Chats still running show a spinner in the sidebar.
• Web appHTML preview in chat. HTML blocks in a chat reply open in a Code or Preview view, and the preview runs sandboxed. The user menu also has a download menu for the Desktop app.
• AccountsSign-in and sign-up links. Links with ?auth=login or ?auth=register open the sign-in or sign-up dialog on top of the page.
• MarketplaceOAuth for MCP servers. MCP server connections can now sign in to servers that protect their endpoint with OAuth, and the marketplace MCP server lists allowed values in its tool schemas. The Models page has a Data retention (ZDR) filter, and a model's page lists its actions and functions.
• KnowledgeMore control over knowledge search. A knowledge collection can store its own search and ingest defaults, and reranking can be turned on per collection (it is off by default). Searches can group hits by document and return each hit with its neighboring chunks, and a new Data page lets you browse, search and edit an index.

• Desktop appA redesigned Koder. The desktop app and the terminal now share one visual system. Every screen, menu, dialog and settings page was rebuilt for light, dark and narrow windows, and for right-to-left languages.
• KoderA faster, livelier terminal. Code, diffs, tools and todos are color-coded, and wide terminals show a side column with context use, tokens, spend, speed and live subagents. Typed keys appear about three times faster, even while a reply streams.
• KoderSource control built in. Connect GitHub, GitLab, Forgejo/Gitea, Bitbucket or Azure DevOps to clone and publish repositories, open pull requests and start a chat from a pull request's review comments. Each chat can work in its own worktree.
• KoderKoder as an MCP server, and more agents. Other agents, such as Claude Code and Codex, can start and read Koder sessions after pairing with a code, held to a scope they can't exceed. You can also add Gemini CLI, Grok Build, OpenCode, Pi and others from the agent registry.
• Desktop appYour phone as a remote. Pair a phone to follow chats, answer questions and approve steps from an installable web app, with notifications. A phone sees only what its pairing allows and can never raise a chat to Full access.
• OmniOmni does more. Omni now makes slides, documents and spreadsheets, runs on folder and script triggers, and drives a browser with its own sign-ins. Its steps read in plain language and open into detail.

• KoderStarts in about a second. Koder no longer waits on the optional Codex, Claude and Cursor sign-in checks at launch: they run in the background, so the prompt is ready right away. First launch went from about 25 seconds to about 1.
• KoderA model picker that helps you choose. Browse and compare models instead of guessing at a model id.
• KoderKnow when Koder needs you. The terminal rings and the window title shows a badge when a run needs your input, or finishes while you're in another window.
• KoderPrivate logs. Prompts, commands and answers are no longer written to the log, log files are readable only by you, and koder debug config hides API keys. Model and tool output can no longer fake buttons or hide text in the transcript and approval prompts.
• KoderPermissions that ask when they should. With Codex, reading outside the workspace, secret files or browser data each asks first, and commands stay inside the workspace. Safe Auto asks before changes, as intended.
• KoderFixes. Sign-in opens your browser reliably on Mac and always shows the link with a Copy button. Exporting a transcript works, a provider's content filter gets a plain explanation that keeps the partial reply, and Cursor uses the model you pick.

• Agent EngineFaster agent setup. Tool schemas, tool details and skill metadata are now cached and shared across runs, so agents start faster on a first run and after an edit. Edits to tools and skills still show on the next run.
• Agent EngineNo more hanging setup. Calls made while setting up an agent and its tools now have deadlines, so one stalled service can no longer hold a run for about 15 minutes before any step runs. Short bursts of server errors during setup are retried.
• Agent EngineClearer setup errors. When none of an agent's tools can start, the run error now gives the reason code and names the tools, for example "None of the agent's 3 tools could be initialized (Company search)", instead of "Something went wrong while running the agent." Tool warnings carry their own codes.
• Web appRedesigned chat. The chat composer is now a pill with a "+" menu, the chat home has a new greeting and a Ready for you row of apps to connect, and the sidebar is one collapsible nav with side columns per page.
• Web appNew welcome. A new welcome replaces the old onboarding, including a card for connecting WhatsApp or Telegram.
• MarketplaceGoogle Places search and Veo pricing. A new Google Places text search tool is available, and Veo 3.1 video generation is now priced per second.

• KoderAthena is the default model. New users start on Athena, aixplain's meta-model, which picks the right model for each request. If your workspace's catalog doesn't include Athena, Koder starts you on GLM 5.3.
• Desktop appSpeech-to-text dictation. Dictate prompts with local Whisper or aixplain's cloud speech models. Cloud speech follows your workspace's residency and provider policy.
• Desktop appPreview, download and version generated files. Files Omni and Koder create are listed, and HTML and images preview beside the chat. You can download a file, show it in Finder or Explorer, or go back to an earlier version.
• Desktop appUpdates arrive ready to install. The desktop app checks for updates at launch and downloads them in the background. On macOS, "Restart and install" is now quick.
• KoderLarge MCP catalogs cost fewer tokens. When your MCP servers expose many tools, Koder searches them on demand instead of sending every schema on every request. That's about 25–28% fewer input tokens per step, with the same solve rate.
• KoderAsk mode asks before every change. Shell commands, file edits, code runs, package installs, MCP and connected-app actions, schedules, worktrees and workflows all ask first. Full access still asks before an edit deletes a file.

• Agent EngineSubagent steps inside delegations. In a team agent's step trace, each delegate_task step now carries the subagent's own steps under steps, updated while it runs, plus the subagent's execution_id. You can open a delegation and see what the subagent did, including for parallel delegations.
• Agent EngineFewer failed tool calls. Tool-call arguments that are close but mistyped, such as a number or true/false sent as text, or a list sent as a JSON string, are now converted to the tool's schema before the call. Before, the call was rejected and the agent spent a step retrying.
• Web appBrowse as a guest. You can browse the web app without signing in. Private pages and actions open a sign-in or email-verification dialog, and the marketplace stays visible to guests.
• Web appReuse agents as sub-agents. In the builder you can add an existing agent as a sub-agent, and an attached sub-agent keeps its own settings instead of taking on the main agent's.
• Web appTrigger and sign-out fixes. Trigger creation lists only tools that can fire events, and signing out now clears your account data from the browser.
• KnowledgeSearch with several queries. Knowledge search now accepts a list of queries in one call.

• Desktop appIntroducing Design. Design sits beside Koder and Omni: ask for a landing page, an onboarding flow or a settings screen, and it builds a working prototype. It then renders the prototype, studies its own screenshots and fixes what it finds.
• Desktop appLive preview and point-to-edit. Watch changes land in iPhone, Pixel, tablet and desktop frames, select any element to say what should change, and adjust colors, fonts and sizes live before saving them to the code.
• KoderMeet Athena. Athena is aixplain's meta LLM: for every request, it chooses how to answer from several models. Where it's available, Athena is the default and the first choice in the model picker.
• KoderEvery step has a name. Steps read like "Read cart.js" or "Ran npm test", a failure shows its exit code and last lines of output, and a finished turn folds into "Worked for 2m". It works the same in the desktop app, the terminal and koder run.
• KoderScheduled runs. Put Koder and Omni on a schedule: run a task every morning, see its history, pause it or run it now. Unattended runs keep every deny rule, and a new triggers skill schedules or event-triggers your aixplain agents.
• OmniA redesigned Omni. The composer adapts as you type, Omni has a character you can shape, and an Ask, Auto and Full access switch sits beside Send.

• Desktop appThe desktop app is now aixplain. The app shows up as aixplain in Applications, Spotlight, Launchpad, the Dock and the Start menu. Your settings, sessions and sign-ins carry over.
• Desktop appUsage shows what you were billed. The Usage page starts with what aixplain billed per day and per model, plus your balance, and refreshes every 30 seconds. Koder's own cost estimate now matches your wallet; before, it could show about four times the real amount.
• KoderModel picker starts short. The picker opens on top coding models, models on a free offer and your recent picks, each with aixplain's real per-million-token prices. The hover card shows data retention, limits, price and host.
• OmniOmni shows its work. A line under Omni's icon shows what it's doing, a checklist ticks off multi-step tasks, and a files button lists everything Omni made in the chat. Drop an Excel file and Omni can read and edit the workbook.
• Desktop appFixes. The desktop app works without the Koder CLI installed, Kimi K3 no longer fails after a PDF the provider rejects, and opening Manage models from Settings no longer crashes the app. On Windows, the koder.exe CLI is now signed.

• Desktop appLocal models in the desktop app. Koder detects your hardware and suggests a model that fits your RAM and free disk. It sets up Ollama, adopts a running llama-server or LM Studio, and checks that a model answers before offering it.
• KoderMore document formats. The desktop app, the terminal and the read tool read Excel (including .xlsm and templates), Word, PowerPoint, OpenDocument, EPUB and RTF. Older .xls, .doc and .ppt files get a clear message to save a modern copy.
• KoderMath rendering. The desktop renders LaTeX, including chemistry via \ce{…}, and the terminal shows formulas as readable Unicode. A new built-in latex skill compiles documents to PDF with Tectonic and checks equations with SymPy.
• KoderUsage from everywhere. The usage page shows Koder, Claude Code and Codex usage together, with plan limits.
• KoderEdit approvals show the diff. The permission card shows the change before you allow an edit.
• KoderSecurity and reliability fixes. Tool calls from connected native agents go through Koder's permission prompts, shared chats stay bound to the account and server that created them, and local engines accept only loopback hosts. The snapshot store no longer re-captures files that keep growing, such as model downloads.

• KoderUse plans you already pay for. Claude Code, Codex and Cursor subscriptions now work in Koder, which runs each vendor's own CLI with the sign-in you already have. Codex lists the models your ChatGPT plan offers.
• KoderSubscription status at a glance. Settings → Providers → Your subscriptions, or koder providers subs, shows each one as Ready, Not signed in or Not installed, with your account, plan and the one command that fixes it.
• OmniOmni acts in your connected accounts. Read your inbox, check your calendar, draft and send email, and work with Slack, Jira and Linear through aixplain connections. Every send, create or delete shows the exact content and asks first.
• OmniOmni knows your aixplain. When you're signed in, Omni sees your team's agents, models, connections, knowledge bases, tools and pipelines so it can pick the right one. Memory can be reviewed, edited or forgotten, and koder omni runs it in the terminal.
• KoderZero data retention filter. Filter to models the catalog reports as zero data retention, in the app or with koder models --zdr. Every model shows its maker's logo, grouped by who serves it, with aixplain first.

• Agent EngineShared Memory attaches as memory. A single-user Shared Memory connection on an agent is now treated as memory: its contents are added to every run and the agent can save to it. Before, the agent saw it as a plain tool with raw get and insert actions.
• Agent EngineUsage tied to the right agent. Model and tool calls made during an agent's run are now attributed to that agent, so usage by agent adds up correctly.
• AccountsSafer social sign-in. Social sign-in now checks that credentials were issued for aixplain, and sign-in only matches your exact email address.
• MarketplaceLower cost for cached prompts. Cached prompt tokens are now billed at the model's cache-read price.

• KoderTeam billing. See your selected team's credits and saved-card summaries in Settings → Account, with /billing in the terminal, or with koder billing. Top-ups and new payment methods go through Stripe's hosted pages in your browser.
• KoderReliability fixes. The "Worker event exceeds size limit" crash is fixed, terminal mouse selection and dialog dismissal work again, and desktop messages, todos and permissions keep flowing during file activity.
• Desktop appCleaner stops and quits. Abandoned background work no longer restarts stopped sessions, and the app waits for its server and terminal to clean up before quitting.
• Desktop appDownloads for every platform. Signed and notarized macOS desktop builds for Apple silicon and Intel, Linux installers and a Windows x64 installer. The Windows desktop build is unsigned in this patch, so Windows may show an unknown-publisher warning.

• Agent EngineSeveral memories on one agent. An agent can now carry more than one memory connection, such as a per-user memory next to a shared one. Each memory gets its own labeled block in the agent's context, and a save that would go over a memory's size limit is refused instead of quietly dropping the oldest facts.
• Agent EngineBudget limits enforced. Budget limits set on an agent now stop the run when they are reached. Before, they were not applied and the run went on as normal.
• Agent EngineFailed task steps reported. On team agents that follow tasks, a failed step now fails the run instead of ending in an empty successful answer, and a failed run is no longer restarted from the beginning, so earlier steps don't repeat their actions. A code tool whose code raised an error now returns that error instead of a result that looks normal.
• Agent EngineStructured run warnings. Run responses now list warnings as {code, message} objects and give a failed run's cause as error_details ({code, message}). If you read warnings, each item is now an object instead of a string, and an empty list is [] instead of null.
• Web appNo sign-out when the service is unreachable. When aixplain is briefly unavailable, the web app shows an unreachable screen instead of signing you out.
• Web appName tools per agent. In the builder you can give a tool its own display name on an agent, and run steps show it as displayName next to the asset name. Two connections of the same integration, such as a sales Slack and a support Slack, are now easy to tell apart.
• Web appCapability library updates. Integrations show a description, a Custom toggle gathers the integrations you build yourself, and Connect new moves to the footer. The trigger creation dialog now offers the default agent.
• AccountsUpdated Google sign-in. Google sign-in now uses Google Identity Services. If Google doesn't show its prompt, the page shows an error instead of a button that does nothing.

• KoderAccurate usage and cost. The context budget shows on both CLI and desktop, multi-step turns count their full token usage, and cost is marked as an estimate everywhere.
• KoderBigger `/review`. Dedicated security, performance and concurrency specialists join the existing reviewers.
• KoderOffice and PDF attachments. Read and attach Excel, Word and PowerPoint on CLI and desktop. On a text-only model, PDFs are pointed at pdftotext instead of dead-ending.
• KoderBuilt-in local engine. Run on-disk GGUF models with no separate daemon. The full live OpenRouter model list is also available.
• KoderNew tools. multiedit makes batched atomic edits to a single file, repo_map gives a fast structural map of the codebase, and LSP navigation is on by default.
• KoderSteer mid-turn. See and control messages queued during a turn, with a steer-now option. MCP tool calls are bounded to 5 minutes by default, so a hung server can't stall a turn.

• KoderRead and attach Office documents. The agent can read Excel, Word and PowerPoint files, and you can attach them on paste: spreadsheets as CSV per sheet, documents and decks as text. Works on both the CLI and desktop.
• KoderPaste text and source files. Pasting a text or source file attaches its contents, with a size cap, instead of just inserting the path.
• KoderImages no longer fail the turn. An image that can't be processed is replaced with a clear placeholder on the model request, and the turn continues. The saved transcript keeps the original.
• KoderMulti-session memory leaks fixed. Long or multi-session runs in the terminal no longer grow memory without bound, and session history stays intact.
• KoderAudit trail. Sessions are bracketed with session.start and session.end events, and per-call tool.start and tool.end events can be recorded with config.audit.record_activity.

• AccountsSign in from another device. Device sign-in is now on: an app can show a short code that you confirm on app.aixplain.com to sign it in.
• AccountsVerified emails for social sign-in. Social sign-in now only accepts verified email addresses.
• MarketplaceDescriptions in listings. An asset without its own description now shows its function's description in listings.

• Web appCapability library. The tool library is rebuilt as a single capability library that lists your tools and integrations together.
• Web appWorkspace and credit tweaks. The create-workspace page has a back control, and the top-up dialog offers updated credit amounts.

• Desktop appSettings crash fixed. Opening Settings → General no longer crashes the desktop and web app, a regression from v0.3.8.
• KoderRecommended models first. The catalog marks and ranks the models aixplain recommends, with a ★ in koder models and a "Recommended" tag in the picker. The same order applies in the CLI, the terminal picker and the desktop and web picker.

• Agent EngineNo empty answers on provider failure. When a model provider fails partway through a response, the run now fails with a supplier error instead of returning an empty successful answer or running a cut-off tool call. Failed run responses and webhooks also keep the steps recorded before the failure.
• Web appTop up credits. You can now top up credits from the billing pages in the web app.
• Web appAgent dashboard filters. Agent runs can be filtered by session, the dashboard can open an agent-scoped tab, and asset lists show status badges. Saved max tokens and output format are no longer overridden by run defaults.
• MarketplaceExa search and model fixes. A new Exa web search tool is available. Bedrock tool-call arguments now come back as the model wrote them, and a user message that happens to be valid JSON is no longer misread.

• Desktop appDesktop security fixes. On Windows with WSL, paths no longer go through a shell, so a path can't run commands. Links in the desktop and web app open only http, https and mailto addresses, and deep links are validated before they reach the app.
• Desktop appVerified Windows updates. Windows desktop updates verify the installer's Authenticode signature when the build is signed. The hosted web app now sends baseline security headers.
• Desktop appFull menu on Windows and Linux. New Window, Reload, Check for Updates, Settings and navigation are now reachable outside macOS. "Open with editor" works on Windows, and a startup failure shows an error dialog instead of a blank window.
• Desktop appClearer onboarding errors. Onboarding warns when a pasted API key is the wrong shape, and a mistyped session ID says "Invalid session ID". A fresh install no longer shows a database migration message when there's nothing to migrate.
• KoderWeb app connection fix. The hosted web app no longer connects to the wrong origin and silently fails every request.
• OmniOmni SDK runtime and bounded autonomy. Omni gets a managed aixplain SDK runtime, and session autonomy is now bounded.

• KoderCLI-only release. Builds for macOS arm64 and x64, Linux arm64 and x64, and Windows x64, under both koder and the legacy aixplain-code names. Existing users can run koder upgrade.
• OmniOmni keeps its MCP tools. Omni retains configured MCP tools and their permission checks, and includes bundled aixplain skills that work outside the source checkout.
• KoderNarrower delegated tasks. Delegated tasks can pick a model and use narrower permissions, while keeping the parent's restrictions and agent permission ceilings.
• KoderReliable reverts and shell output. Session reverts preserve Unicode and special filenames and report restore failures. Shell commands keep their output at exit and during cancellation, and Windows installs preserve existing PATH entries.
• OmniCredentials stay on your server. Omni HTTP helpers keep credentials on the configured server, don't automatically retry paid writes, and save result files atomically.

• Agents SDKAgent files and skill files. With the unified File asset you can attach files to an agent once, and they are available on every run instead of being uploaded each time. Skill.update() updates a single file in a skill without recreating it, and list_files() lists the files a skill holds.
• Agents SDKSDK v1 deprecation notice. Importing SDK v1 now shows a deprecation notice. v1 will be removed on February 1, 2027, so move to from aixplain import Aixplain using the migration guide. Set AIXPLAIN_SUPPRESS_V1_DEPRECATION=1 to hide the notice.
• Agents SDKClaude Code plugin. The SDK repository now ships an aixplain plugin for Claude Code, with skills for searching the marketplace and for creating, deploying, testing and updating agents with the SDK.
• Agents SDKSafer, steadier requests. The SDK now requires HTTPS for its endpoints, warns when one isn't an aixplain host, and drops your API key when a request is redirected. Every request has a default timeout, and retries no longer submit and bill the same run twice.
• Agents SDKFixes. Updated skills now persist on save(), trigger schedule fields (timezone, time, run_at, days, schedule_type) load instead of reading null, and APIError messages include the provider's error detail.
• Web appFixes. File downloads are named after the file, scheduled triggers no longer run twice, and per-API-key budgets are now enforced.
• MarketplaceBrave Search and billing fixes. A new Brave Search tool is available, and the search tools now take locale and date filters. Bedrock models are priced from their real price table, bad audio is no longer billed as 900 seconds, a cached result is only reused for a call with the same settings, and client errors are no longer reported as server errors.
• KnowledgeIngestion fixes. Documents with no readable text are now reported instead of being added empty, and embeddings stay matched to their chunks when ingesting in batches.

• Agent EngineFixes. Tools with multi-value inputs, such as several weekdays or days of the month, now receive the full list. When a tool returns nothing, the agent's answer is now kept in the session, so the next turn no longer re-answers or contradicts it, and a tool failure with a specific cause, such as QUOTA_EXCEEDED, reports only that code instead of also adding TOOL_FAILED.
• Web appUnsaved agent drafts. The builder has an Unsaved tab for agent drafts saved in your browser.
• Web appFixes. Marketplace landing search keeps results in relevance order, searching by asset id works, and a stray Backspace no longer deletes an agent node in the builder.
• MarketplaceData retention on models. The marketplace shows each model's data retention and marks zero-retention models.

• KoderAutonomous goal rounds. A session with an armed goal keeps working toward it after each turn, up to a configurable round cap, and stops when it stalls. It's gated behind KODER_EXPERIMENTAL_GOALS.
• KoderUpdate-available toast. The first time a new version is found, Koder shows a one-time toast alongside the status-line badge. Run /update to upgrade.
• KoderConfigurable review depth. review.effort, review.verify_threshold and review.comment steer the multi-specialist review pipeline.
• Koder`permission.ask` plugin hook. Plugins can allow or deny a permission before you're prompted, or fall through to the prompt. It runs after your configured rules, so it can never override a deny.
• KoderFixes. Crashed workflow runs show as paused instead of live in /workflows, session sync no longer applies a stale workspace snapshot, and shell commands render with clear IN/OUT labels.

• Web appRun traces in chat. Chat messages now carry the trace of the run behind them, so you can open the steps that produced an answer from the conversation.

• KoderSidebar working indicator is back. With several sessions running, the session list again shows which ones are busy. A pending permission still takes precedence.
• KoderCLI builds for every platform. Downloads cover macOS (Apple silicon and Intel), Linux (x64 and arm64) and Windows x64, under both koder and the legacy aixplain-code names. This release supersedes v0.3.2.

• KoderTool calls that arrive as text. Some OpenAI-compatible servers, such as vLLM or SGLang started without a tool-call parser, stream a tool call as plain text, so it never runs. Koder can now read the call back out and run it, behind KODER_LIFT_TEXT_TOOL_CALLS (off by default), with permissions applied as usual.
• KoderMore options in grep and read. grep gains output modes, context lines, a head limit, and case-insensitive and multiline search. read renders Jupyter notebooks cell by cell with outputs, and glob takes a limit.
• KoderClaude subscription login fixed. The Claude subscription bridge no longer rejects valid logins. Your own Claude Code hooks and CLAUDE.md now apply in that subprocess, and Koder warns you when it starts.
• KoderSubagent spend in cost readouts. Subagent spend is included in cost readouts and updates live, and forks start at zero. aixplain usage is priced by what it actually charges.
• KoderCleaner app and terminal. Context usage moved to the composer, redundant spinners are gone, and slash commands match between the terminal and desktop. Subagents have visible stop controls.

• Agent EngineJSON messages no longer fail. When a message's text is itself a JSON object or array, such as a webhook payload passed straight through as the first message, the run now goes ahead instead of failing with a model-unavailable error.
• Agent EngineClearer trigger schedules. Setting days of the week on a trigger that isn't weekly now returns an error asking for a weekly schedule, instead of quietly running every day.
• Web appManage agent channels. An agent's channel connections are now listed in one place, and you can revoke each one on its own. Agent pages also show the skills attached to the agent.
• KnowledgeSteadier knowledge ingestion. Re-ingesting a document now writes the new chunks before removing the old ones, deleting a document removes only that document, and searches stop at a real time limit. Each ingest request is now capped at 100 documents.
• KnowledgeSafer knowledge bases. File downloads and document processing for knowledge bases are now better protected.

• Agent EngineScheduled triggers stay on. A trigger created without an enabled value now starts live instead of being saved switched off, and renaming or updating a trigger no longer turns it off. Sending enabled: false still switches it off.
• Web appMove to the new app. The previous studio now shows a banner inviting you to switch to the new web app at app.aixplain.com.
• MarketplaceCache-read pricing. Model pages now show the cached-input price next to the input and output prices.
• MarketplaceTool schemas kept intact. Tools with an input named like a JSON Schema keyword, such as items, type or properties, no longer have their schema rewritten, so model calls to them no longer fail with an invalid-schema error.

• Agent EngineGraph edge conditions in JSON. Edges in a static graph can now carry a condition written as JSON: a state key, a comparison (eq, ne, gt, gte, lt or lte) and a value, checked directly against graph state. Conditions written as Python import paths are no longer accepted.
• Agent EngineClearer quota and attachment errors. Reaching a daily usage limit or sending a malformed image attachment now reports the real cause, instead of showing the model as unavailable.
• Web appSkills on the marketplace. Skills published by other teams now appear on the marketplace and in the agent builder's picker, and you can download a skill as a zip from its page. Skills you don't own no longer offer edit actions.
• AccountsRedesigned sign-in. Sign-in and sign-up pages have a new split-screen layout, and password reset has a simpler page. Sessions also refresh more reliably, so you are no longer signed out unexpectedly while working.
• MarketplaceMore complete search. Marketplace search now includes skills and files, and deleted assets no longer show up in results.

• Web appRecurring API key credit limits. An API key's credit limit can now reset daily, weekly or monthly instead of only being a lifetime cap, and limits can include fractions of a credit. Usage counts only the current period, and the full allowance returns when it rolls over.
• Web appSmall app updates. Files pages have a Download action, starting a new chat keeps the agent you were using, and the collapsed sidebar shows settings icons.
• MarketplaceGemini 3.1 Flash Image. Gemini 3.1 Flash Image is now available for text-to-image generation and for editing with a reference image, at 1K, 2K and 4K.
• MarketplaceWeb search citations stay intact. Web search now returns its reference links in a format that keeps them clean, so cited links no longer end in a stray dash and break.

• KoderSelf-updating repaired. Earlier releases misdetected how Koder was installed, so koder upgrade failed and background updates never ran. Both work again for installs made by the shell installer.
• KoderReinstall once from v0.2.4 or v0.3.0. Those builds can't upgrade themselves to this release. Run the installer once with curl -fsSL https://github.com/aixplain/koder/releases/latest/download/install | bash; your configuration, credentials and session history are untouched.
• Koder`koder uninstall` removes everything. It now removes the whole install, including command aliases, the PATH entry and other installed versions, with a stricter guard against touching anything outside Koder's own directory.

• KoderOllama works out of the box. If Ollama is running, Koder finds it with no config and asks which models can chat, call tools and accept images, and how much context each has. Embedding models are left out of the picker.
• KoderWork that outlives a turn. koder spec turns a request into requirements, a plan and tasks run in concurrent waves, /workflows gives named runs you can pause, resume and stop, and /loop schedules a recurring prompt. Corrections you make are recorded and re-applied in later sessions.
• KoderDormant MCP servers. MCP servers can stay dormant until a keyword in the conversation wakes them, so connecting many tools no longer means paying for all of them at startup. Koder also reads Claude Code agents, hooks, plugin manifests and marketplaces.
• KoderControls you can verify. koder inspect shows what a configuration permits, koder trust which directories may contribute config, koder compliance every network destination it allows, and koder audit a tamper-evident record of permission decisions.
• KoderBrowser automation, opt-in. Drive a real browser over CDP: navigate, read the accessibility tree, manage tabs, detect login walls and record sessions for replay. Turn it on with KODER_EXPERIMENTAL_BROWSER=1.
• KoderSecurity fixes. Opening an untrusted repository can no longer turn off the sandbox, links in model output can't navigate the desktop app away from its own page, and sessions, schedules and workflow runs stay within their own project. An administrator's deny rules now hold everywhere a permission is decided.

• Agent EngineAttach files and folders to agents. You can now give an agent persistent files and folders that it reads on demand during a run, rather than having them loaded into the prompt. The agent can list, read, glob and grep these reference files, folders keep their nested structure, and a file it opens stays available for the rest of that run.
• Agent EngineLong conversations stay within limits. When a session nears the model's input limit, Agent Engine now condenses older conversation before each model call instead of running out of context, and the stored session history stays complete. Set contextOverflowStrategy on an agent, a sub-agent or a single request to truncate (the default), summarize or none.
• Agent EngineReasoning models keep their answer. Some aixplain-hosted reasoning models mixed their thinking into the answer and lost the actual response. The engine now separates the two, so the final answer comes through.
• Agent EngineMore reliable tool calls. When a model writes a tool call as XML-style tags instead of the expected JSON, the engine now recovers the call and runs it, rather than returning the raw tag text as the answer.
• Web appAgent builder keeps your drafts. The agent builder now saves your work locally and warns you before you leave with unsaved changes. The agent drawer has a new Channels tab, and the chat box grows as you type, with a toggle to expand it.
• Web appOut-of-credits warning. When you have no credits left, the app now blocks new runs and shows a warning instead of letting them fail.
• Web appGet aixplain Desktop from the web app. A sidebar card and an account menu item offer the Desktop download, and a setup guide at /guide/desktop walks you through it.
• MarketplacePDF input for text models. Text-generation models now accept PDFs and other documents in one standard file content block, as a file id, base64 data or a URL, and each supplier receives it in its own format. Documents at a URL are downloaded and passed along for suppliers that can't fetch URLs.

• KoderImport sessions from other tools. koder import brings conversation history from opencode, Claude Code and Cursor into Koder. Re-running adds no duplicates, and each source is opened read-only.
• KoderOptional OS-keychain credentials. Store API credentials in the macOS Keychain, Windows Credential Manager or Linux libsecret instead of a file. It's opt-in; the file remains the default.
• KoderSecurity hardening. OAuth tokens are no longer left in a world-readable database, your credentials no longer reach proxied upstreams, and remote-fetched code paths require HTTPS and explicit consent. Stalled streams now fail fast by default.
• KoderApp improvements. A "Session changes" review mode, a Stop button for individual subagents, a per-message fork, and show/hide toggles for API-key inputs.
• KoderCost labeled as an estimate. Displayed cost is labeled as an estimate, and turns with no usage no longer show $0.00.

• KoderCommand-line release. Koder installs on macOS, Linux and Windows from the command line with curl -fsSL https://github.com/aixplain/koder/releases/latest/download/install | bash. Open a new terminal afterwards and run koder.
• KoderSigned macOS binaries. macOS binaries are Developer ID signed and notarized, so they run without a Gatekeeper prompt. The previous command name, aixplain-code, still works.

• Web appRename and delete chats. Chat sessions now have rename and delete actions, and the default agent is now called the workspace agent.
• Web appSteadier chat attachments. Files attached to a chat session are now read and placed correctly, with clearer errors when one can't be read.

• Web appWorkspace settings for members. Members can now open workspace settings, with admin-only actions hidden, and the agent library lists all agents.

• Agent EngineFewer timeouts before a run starts. Some runs, even a plain greeting, failed with a timeout while the agent's tools and models were still loading. Setup now keeps the run alive, and slow lookups give up and retry quickly, so the run gets through.
• Agent EngineTools that fail to load are reported. If some of an agent's tools fail to initialize, the run still succeeds and returns warnings naming each one, instead of the agent quietly answering without them. If none of the configured tools load, the run fails with a message saying how many failed.
• Web appEasier browsing and run details. Model and tool lists now load more as you scroll, run traces show the date with each run time and can open the raw JSON, and model pages show the host's logo next to the developer's.

• Web appTool library shows every tool. The tool library now lists all the tools you can use, not only the ones you created.
• Web appMCP under integrations. MCP servers now have their own tab in integrations, with one action for adding an API key.
• Web appEasier chat and invitations. The chat tools control now sits next to the attach button, and links to a chat open it with the right agent selected. Pending workspace invitations now show up across the app.

• Agent EngineRun artifacts in responses. Agent run results now include an artifacts list with what the run produced, such as images from media-generating tools and files or links cited in the final answer. Each artifact carries its category, MIME type, link expiry when known, and whether the answer mentions it.
• Agent EngineFull links in answers. Final answers no longer contain shortened link placeholders in place of real URLs. Agent Engine now expands any that remain when it assembles the result.
• Web appEasier integration setup. The integration connection dialog makes it easier to choose which actions to connect.

• Agent EngineCalling agent sent to tools. Requests from an agent to aixplain and MCP tools now carry an x-agent header with the name of the agent that made the call, next to the caller's x-user-id. In a team, the header names the sub-agent that made the call.
• Agent EngineLimits and costs hold across teams. Run, agent and sub-agent limits now combine instead of overwriting each other, sub-agents receive the run's limits and per-role model parameters, and partial budgets are kept. Run cost now follows the credits the platform reports, and streamed model calls report their real cost, so cost budgets can take effect.
• Agent EngineSteadier tool calls. Agents now detect cut-off tool-call arguments instead of retrying until the run times out, and repair malformed arguments from the model. Tools whose input schemas use union types now work.
• Agent EngineClearer model errors. When a model can't be reached, the run fails with a MODEL_UNAVAILABLE error that names the configured model, and more platform error codes now map to a plain message.
• Agent EngineConversation history kept. Messages saved to the same session at the same time now merge instead of overwriting each other, and planner and graph runs now save the agent's reply to the session history.
• Web appMax iterations in the builder. Set an agent's maximum iterations in the agent builder, within the allowed range. The output template now takes JSON only and is checked before you save.
• Web appIntegration tab on assets. Integrations now open in a shared Integration tab on asset pages, and the integration drawer shows the tools you have connected.

• Agent EngineStep thoughts from model reasoning. Agent step thoughts now come only from the model's own reasoning, so reasoning models show their actual reasoning and other models no longer show generated thoughts.
• Agent EngineNo empty answers after long reasoning. When reasoning uses up the whole output budget, the agent now runs a final turn to give an answer instead of returning nothing. The response flags it with limit_reached and RUN_LIMIT_OUTPUT_TOKENS.
• Agent EngineSteadier planned team runs. The planner now plans work for sub-agents as agent steps, so team runs that use planning hand off to sub-agents correctly.
• Agent EngineAttachments without a message. You can now run an agent with attachments and an empty message.
• Web appClearer tool pages. Trying a tool now has its own tab, and a tool's actions read as documentation.

• Agent EngineCaller identity sent to tools. When you run an agent with an identifier, Agent Engine passes it as an x-user-id header on every aixplain and MCP tool request, so your tools know which end user they serve. MCP servers need the streamable HTTP transport to receive it.
• Agent EngineRuns no longer hang on a stuck tool. Tool calls now have a default execution timeout, so a connector that stops responding can't stall the run. Graph nodes now honor their declared timeout and retry policy.
• Agent EngineFaster progress and webhooks. Progress checks now respond faster, and the completion webhook is sent as soon as the run finishes.
• Agent EngineAttached documents read directly. When the platform has extracted a document's text, the agent reads it on the current turn instead of fetching the file. Later turns refer to the document without replaying its full text.
• Agent EngineInspector reruns work. Runs where an inspector asks for a rerun of tool calls now continue instead of failing.
• Web appCustomizable default agent. The agent library now includes aixplain's default agent, which you can set up with your own tools, integrations and models.
• Web appTidier builder canvas. The agent builder canvas lays out nodes automatically so they don't overlap, and shows alignment guides as you move them.
• Web appDashboard and chat updates. Reorder dashboard tabs, and see recent runs and the last run date on each agent row. Chat now shows error messages and loads older sessions with Show more.

• Agent EngineClearer errors and working output formats. Failed runs now return a plain message naming the cause, such as a rate limit, quota, budget, timeout or invalid model setting, instead of a raw error. An agent's output format and expected output now apply in Agent Engine 3, so JSON agents return structured JSON and Markdown agents return Markdown.
• Agents SDKSkills for agents. Create a skill from a folder with a SKILL.md file plus resources and scripts using aix.Skill(...), then attach it to an agent with skills=[...]. Skill objects and skill IDs both work.
• Agents SDKSessions on the platform. aix.Session manages sessions stored on the platform, with messages, reactions and file uploads, and create_session() replaces the deprecated generate_session_id(). Agents also accept attachments=[...] with images or audio as files, paths or URLs for vision and audio models.
• Agents SDKScheduled and event triggers. aix.Trigger runs an agent on a schedule (once, daily, weekly, monthly or at an interval) or when an external event arrives, such as a new email in a connected Gmail account.
• Agents SDKRun budgets and tool overrides. Set agent.budget = Budget(max_cost=..., max_duration_seconds=..., max_iterations=...) to cap a run, and save a tool's parameter values at create time (tool.actions.X.inputs.Y = value) that any run() call can override.
• Agents SDKGuardrails from the marketplace. Find guardrails with aix.Inspector.search(), fetch one by path with aix.Inspector.get("aws/sensitive-information-guardrail/aws"), and attach it with inspectors=[...]. This replaces the PrebuiltInspector class.
• Web appRedesigned web app. The agent builder now puts chat next to Canvas, Code and Tracer tabs with a live tracer, and asset pages, quick search and the home page are redesigned. Prices now show in credits, and dashboard tabs keep their agents, date range and pins.
• Web appChat updates. Recent chats now appear in the sidebar, you can connect integrations from the chat landing page, and you can attach files when the agent supports them.
• AccountsWorkspace invitation page. A new invitation page lets you view and accept a workspace invite.
• MarketplaceImages and reasoning for LLMs. LLMs now accept images, and reasoning effort works the same way across the major LLMs, including aixplain-hosted ones. Python script tools accept much longer code, SQLite tools no longer fail on long descriptions, the File Manager handles non-English file names, and Deepgram Nova-3 transcripts no longer drop words.

• Agent EngineTools with many actions. Agents now see every action and parameter of a tool, with no cap of 20 actions or 250 parameters. The prompt lists tools briefly, and the agent looks up a tool's full parameters before it calls it.
• Agent EngineSelect actions for connected tools. A run now stops with a clear message when a connected integration tool has no actions selected. Select at least one action for each connected tool.
• Agents SDKAPI key limits take a model. APIKey limit methods and limit objects now take model instead of model_id, and accept a model object as well as an ID. The key's access_key field is now key.
• Agents SDKSign-in link for OAuth tools. Connecting an OAuth integration now returns the tool with tool.redirect_url set to the page where you finish sign-in before using it.
• Agents SDKTool calls in model streams. Streamed model chunks now include tool_calls, usage and finish_reason when the model sends them.
• Agents SDKNew default agent model. v2 agents created without an llm now use GPT-5 Mini instead of GPT-4o.
• MarketplaceStructured output on Bedrock. LLMs hosted on Bedrock now honor response_format, so a request for JSON output gets JSON.
• MarketplaceTurn off integration triggers. Integration connections that support triggers can now disable or delete a trigger by its ID, with the disable_trigger and delete_trigger actions.

• Agent EngineFaster agent startup and better tool inputs. Agent runs start faster because the tool, model and agent lookups made at the start of a run are now cached. Tool descriptions now include each parameter's type details and allowed values, so agents fill tool inputs more accurately.
• Agent EngineFixes. Tools with no input parameters now work, same-named actions in different tools are told apart, and every step now reports an end time.
• Agents SDKAPI keys in SDK v2. The new APIKey resource creates, lists, retrieves and deletes API keys, reads their usage, and sets request and token limits per day or per minute, for all models or one.
• Agents SDKStreaming for models. v2 models that support streaming offer run_stream(), which returns response chunks as they arrive.
• Agents SDKGet models by name. ModelFactory.get() and IntegrationFactory.get() now accept name= as well as an ID.
• Agents SDKv2 works without environment keys. Using the v2 SDK with Aixplain(api_key=...) no longer crashes when TEAM_API_KEY isn't set in the environment.
• KnowledgeAny file when you connect aiR. When you connect the aiR integration with files, files at any web link are now added and parsed into text, not only plain-text files. Files are split into sentence chunks by default.

• Agents SDKLive agent progress. Agent runs can show progress in the terminal or in Jupyter and Colab notebooks, as a single status line or a timeline. Set it with progress_format, progress_verbosity and progress_truncate, which replace the show_progress flag.
• Agents SDKInstruction variables. Pass variables to an agent run to fill {{variable}} placeholders in its instructions and description.
• Agents SDKDebugger meta-agent. Call response.debug() on an agent result, or aix.Debugger().run(...) on its own, to get an analysis of what the agent did. You can pass a question, such as why a run took so long.
• Agents SDKToken limits by type. API key limits take a token_type of input, output or total tokens, so you can cap one kind of usage.
• Agents SDKInspectors can edit. Inspectors can now edit content as well as continue, rerun or abort. Team agents built with the v1 SDK no longer support inspectors; use the v2 Inspector instead.
• Agents SDKResponse data changes. Agent response data now returns steps only in steps, with intermediate_steps removed, and execution_stats is now filled in correctly.
• MarketplaceFind assets by path. Models, tools and integrations can now be looked up by their path, and searches can filter by developer and host. In the v2 SDK, pass path, developer or host to search().
• MarketplaceRename Python script tools. You can now change the name of a Python script tool after you create it. The new name must not match another asset of your team.

• Agent EngineOptional response generation. An agent run can now skip the final response-generation step and return the raw output of its work, for when you format it yourself. Set runResponseGeneration to false in the run request.
• Agent EngineAction details in steps. Steps from integration tools such as Slack or Google Calendar now include the action's name and description, so you can see which action the agent called.
• Agent EngineText output stays text. When you ask for text output, responses no longer come back wrapped in JSON or markdown. In team agents, JSON output from the response step and its max tokens setting now work as set.
• Agents SDKNew `run_response_generation` parameter. run() and run_async() take run_response_generation (default True) to switch the final response step on or off. It stays on when you set output_format or expected_output, and the SDK warns you.
• Agents SDKFixes. run_async() now works with draft agents, and connection tools no longer fail validation when you pass action and data.
• MarketplaceBrowse by category and developer. Marketplace assets can now be filtered by category and by developer. An asset's path now updates when you rename it or change its developer.
• MarketplaceFixes. Models that take JSON or object inputs now receive their data correctly.
• KnowledgeFuller index search results. Index search no longer drops results that score below 0.5, so a search returns the number of results you ask for. Retrieval with Jina embedding models works again.

• Agent EngineFailed tool steps recorded. When a tool call fails validation, it now shows up in the agent's steps. When the same tool runs more than once, each step gets the right timestamp.
• Agent EngineLinks come through intact. Links in agent answers now come back complete and correct, including in long runs and team agents.
• Agents SDKScript connection tools fixed. Script connection tools now work with nested functions, and they pick the right function from the code you provide.
• Agents SDKMCP connections without auth. authentication_schema is now optional for MCP connections, so MCP servers that need no authentication connect without it. Connection tools also handle actions that have no inputs.
• Web appWeb app updates. The web app shows your free-trial details, pinned integrations appear first in lists, and removing an inspector from an existing agent now works.
• MarketplacePostgres and easier integration setup. You can connect a Postgres database as a tool. When you connect an integration without choosing an auth scheme, the right one is now detected, and tools created from integrations show the correct developer.
• MarketplaceCleaner script tool output. Script tools now return only their result. Add debug to the input to also get stdout and stderr.
• KnowledgeDocument parsing in indexes. Files added to an index by link are now parsed into text before indexing. Parsing is on by default, and you can choose the parser or turn it off with parsing_params.

• Agent EngineWebhooks carry the result. Webhook calls now include the response data and the request ID, so your endpoint gets the agent's output without a second call.
• Agent EngineTool steps while the agent runs. Tool steps now appear as each tool runs, not only in the final response, and the final response doesn't repeat them. Model steps show the reason the model gave instead of a generic "LLM call" label.
• Agent EngineEvolution keeps your description. Instruction tuning now changes only an agent's instructions and leaves its description as you wrote it.
• Agents SDKErrors for unknown parameters. Passing a misspelled or unsupported parameter when you create an agent or team agent now raises an error instead of being silently ignored.
• Web appAgent builder (beta). You can build an agent in the web app: add integrations and pick their actions, add tools, inspectors and a responder, and set max iterations and max tokens. Validation then walks you through each step and shows the agent's progress.
• MarketplaceAgents through chat completions. You can call an agent through the OpenAI-compatible chat completions API by passing agent- followed by the agent ID as the model. Conversation history is supported.
• MarketplaceMore capable script tools. Python script tools can load their code from a URL or a file ID, take dictionary inputs and create files. Your aixplain API key is passed to the script, so it can call aixplain.
• MarketplaceSQLite integration. You can connect a SQLite database as a tool.
• KnowledgeIndexes through integrations. You can create an index through the aiR integration and add text to it when you connect. When embedding fails, you now get a clear error.

• Agent EngineLive execution status. Agents and team agents now stream their status while they run: the current stage, the tool in use and the reason for each step.
• Agent EngineModel defaults for temperature. If you don't set temperature or top_p, agents now use the model's own defaults instead of a fixed temperature of 0. Setting them yourself, including to 0, works as before.
• Agent EngineOutput format respected. Agents now return the output format you asked for. Before, responses could come back as JSON whatever the setting.
• Agent EnginePrivate agents stay private. Private agents are now visible only to the team that owns them.
• Agents SDKProgress display in runs. Agent.run() and TeamAgent.run() show progress inline, with a step counter for team agents. Set progress_verbosity to "compact" (the default), "full" for tool inputs and outputs, or None to turn it off.
• Agents SDKResponse format for models. Running a language model now accepts a response_format parameter, so you can request structured output straight from the model.
• Agents SDKFewer agent setup errors. temperature and top_p are now optional, and creating an agent no longer hits a TypeError from the SDK's class checks.
• KnowledgeSearch shows your team's assets. Marketplace search now takes your team into account, so your team's own assets show up in results.

SDK build 0.2.34

• Agents SDKReleased Evolver v1 (beta).

SDK build 0.2.33

• Agents SDKImproved session ID creation and management.

Models

• MarketplaceClaude 4.5 Sonnet, DeepSeek V3.1, DeepSeek V3.1 Terminus and 1 more (4 new assets)

• Agent EngineMore focused inspector reviews. Inspectors now judge only the latest run of a step, without earlier critiques mixed in. A step that ran several times appears once, with its final output.
• Agent EngineSteadier team runs. Team agents now follow the dependencies between the steps in their plan, and integration tools with several actions each run against the right connection. A run that hits the model's max tokens now says so.
• Agent EngineAgents follow their description. An agent's description and instructions are now part of its system prompt, and agents in a team receive their own instructions.
• Agent EngineUnique draft names. Creating a draft agent with the name of an agent you already have now returns an error.
• Agent EngineMax tokens respected. The max tokens you set on an agent or team agent now reaches the model, including the agents in a team.

• Agents SDKCustom inspector policies (private beta). An inspector's policy can now be a Python function as well as WARN, ABORT or ADAPTIVE. You can also use a tool, your own Python check, as an inspector, next to guardrail and text-generation models.
• Agents SDKIndex files directly. upsert() on an index now takes a file path, such as a text or PDF file, and parses the file into records for you.

SDK build 0.2.34

• Agents SDKEnabled specifying agent JSON outputs.
• Agents SDKReleased Inspector v1 (beta).

SDK build 0.2.33

• Agents SDKAdded Integrations and Tools.

SDK build 0.2.32

• Agents SDKAdded prompt benchmarking service.
• Agents SDKEnabled creating aiR indexes with any embedding models.

• Web appConnected and updated Console, centralized aixplain account management system.
• Web appRenamed “utility” to “tool”.
• Web appAdded Integrations and Tool creation.

• Agent EngineEnabled access to team agents and individual agents.
• Web appEnabled uploading images.

SDK build 0.2.31

• Agents SDKEnabled support for LLM streaming.
• Agents SDKEnabled setting instructions for Mentalist.

SDK build 0.2.29

• Agents SDKAdded Vectara and ZeroEntropy indexing.
• Agents SDKAdded GraphRAG.

SDK build 0.2.28

• Agents SDKAdded an SQL tool for SQLite files and CSVs.
• Agents SDKAdded image-based indexing.
• Agents SDKEnabled getting and deleting records from an index.
• Agents SDKAdded save and draft methods to Tools.

SDK build 0.2.26

• Agents SDKAdded Tasks for Teams.

• Web appAdded wallet auto top-up.

• Agents SDKAdded a dedicated “instructions” field for an Agent’s role.
• Agents SDKEnabled setting default parameter values for Agents.
• Agents SDKEnabled setting default parameter values for Pipelines.

• Web appUpdated the Discover search engine.
• Web appRevamped the search layout in Discover.
• Web appMoved tool models into a new Tool asset.
• Web appAdded sorting by relevance when searching in Discover.
• Web appSimplified the subtransactions display.
• Web appEnabled renaming API keys and extending their expiration.
• Web appEnabled transferring Agents and Tools between teams.

Models

• MarketplaceGemini 2.5 Pro (Experimental), Deepseek V3, Gemma 3 27B and 5 more (8 new assets)

• MarketplaceStarted providing a fallback LLM (the same model but on another host) for LLM requests.

SDK build 0.2.26

• Agents SDKAdded a Python code interpreter tool for agents to run code (Beta).

• Web appUpdated Integrations tabs UI.
• Web appEnabled renaming and extending the expiration of API keys.

Models

• MarketplaceDeepSeek R1 Distill Llama 70B, DeepSeek R1, o3-mini and 17 more (20 new assets)

Tools

• MarketplaceDocling

SDK build 0.2.24

• Agents SDKEnabled deploying custom tools.
• Agents SDKEnabled adding custom code directly into agents.
• Agents SDKAdded a markdown display option for agent responses.
• Agents SDKAdded indexing and retrieval through IndexFactory.
• Agents SDKAdded agent building capability to the Swift SDK.
• Agents SDKSimplified script nodes.

• Web appAdded an initialize button for aixplain-hosted models.
• Web appEnabled tryout for agents.

• Web appAdded a stop/cancel mechanism.

Models

• MarketplaceLlama 3.3 70B, Llama 3.3 70B Instruct, Llama 3.3 70B Instruct Turbo and 8 more (11 new assets)

Tools

• MarketplaceScrape (Firecrawl), Crawl (Firecrawl) and Search API (Tavily) (3 new assets)

• Web appAdded a Copy Asset ID button.

• Web appAdded cost information to run assets.

SDK build 0.2.23

• Agents SDKEnabled draft agents and explicit deployment.
• Agents SDKEnabled agent name and description.
• Agents SDKAdded variables to agent prompts.
• Agents SDKRenamed TEAM_API_KEY to AIXPLAIN_API_KEY.

• MarketplaceLlama 3.2 1B, Llama 3.2 3B Instruct, Llama 3.2 3B Instruct Lite and 53 more (56 new assets)

• Web appEnabled Agents in Discover.
• Web appAdded suggested tags for filtering models in Discover.
• Web appAdded list view.

• Web appAdded suggested actions upon creating a new session.

SDK build 0.2.22

• Agents SDKObjectified model response.
• Agents SDKImproved agent error logs.
• Agents SDKAdded rate limiting.

• MarketplaceSabia-3, Llama 3.1 Nemotron 70B instruct, Hume AI Image Expression Measurement and 1 more (4 new assets)
• MarketplaceMeta Llama 3.1 8B (Model – Fireworks AI, Together AI, SambaNova)
• MarketplaceMeta Llama 3.1 70B (Model – Fireworks AI, Together AI, SambaNova)

• Agents SDKAdded programmatic pipelines.
• Agents SDKAdded multi-agent systems (Team agent).
• Agents SDKEnabled adding prompts to agents.
• Agents SDKEnabled agent debugging, including intermediate steps.

• MarketplaceGemini 1.0 Pro, Gemini 1.5 Pro, Meta Llama 3.2 1B and 10 more (13 new assets)
• Marketplaceaixplain Search 2.0 (Model – Index and retrieval engine)

• Web appDeprecated GPT-3.5 Turbo.

• Web appEnabled adding credits.
• Web appActivated live credit updates.
• Web appImproved Bel Esprit response formatting.

• Web appReleased aixplain Chat iPhone application.

SDK build 0.2.18

• Agents SDKAdded input and output attributes to model and pipeline objects.
• Agents SDKIncreased stability of FineTune.

• MarketplaceEnabled complete hosted model self onboarding (Text-to-Text LLMs).

• Web appActivated new Bel Esprit experience in Studio.

SDK build 0.2.17

• Agents SDKEnabled getting team’s credits.
• Agents SDKEnabled tool validation when creating Agents.
• Agents SDKEnabled referencing additional Agent inputs in prompt.
• Agents SDKAdded planning capability to Agents.

• MarketplaceResemble.AI TTS and GPT-4o mini (2 new assets)

• MarketplaceStarted displaying all I/O parameters in model specification pages.

SDK build 0.2.16

• Agents SDKReduced the number of imports to create agents.
• Agents SDKEnabled adding local paths as additional inputs for agents.

• Agent EngineFixed agent outputs when using pipelines that don’t output text.
• Web appFixed text and audio reconstructors not providing outputs.
• Web appFixed filters not updating when navigating within Discover.
• Web appFixed error notice when attempting to clone an audio that is too short.

• MarketplaceALLaM 1, Mistral 7B Instruct v0.3 and Llama 3.1 405B, 70B, 8B (3 new assets)

• Agent EngineIntroducing Agentic framework — This SDK update allows you to create your own AI agents.

• Web appImproved edge display UI in Design.
• Web appReplaced old documentation on Studio with external aixplain Docs.
• Web appUpdated small window size warning.

• Web appFixed hypothesis and source dropdown placeholders.
• Web appFixed expirable credits appearing only after refreshing the page.