Changelog

Settings now has an API keys page where you can create, copy, revoke and rotate the keys for the workspace you are in. A key only works in the workspace it was created in, so calls made with it can only reach that workspace's functions.

The Invoke tab on a function's page has an API key toggle that puts your workspace key into every example, so the code you copy runs as-is. Examples now include the Notte CLI, alongside cURL, Python, TypeScript and MCP. The key is hidden until you turn the toggle on.

Programs can check a running build and read its saved result after disconnecting, without starting another build. Background builds keep running when the client closes, and their progress is available when it returns.

Chat history now shows conversations created in the web app by default. Use the source filter to show web, CLI, MCP, API, or Notte Console conversations. All sources includes those five sources.

You can now pick Mercury 3 from Inception for a thread, with low, medium or high reasoning effort. It reads text only, so it cannot view screenshots, and a thread started on another provider's model cannot switch to it.

Publishing now explains that other users can download your function's source code and fork their own copies. The notice also clarifies that unpublishing cannot retract copies already made.

When you ask about earlier work in a thread, the agent now sees the commands it ran and what they returned, not just its own summary of them. It no longer tells you it never ran a check that it did run.

When a tool command reaches its time limit, you now see guidance to check its progress before retrying a smaller task or allowing more time. The message no longer says the entire turn has run out of time.

Your agent can now inspect screenshots and images without stopping with an unsupported file error.

MCP clients can now distinguish marketplace lookups from tools that build, run, or publish APIs. Tool descriptions also clarify that running an API can modify its target website.

New threads start on GPT-6.1 Sol, which replaces GPT-6 Sol in the model picker. A default you saved in settings is kept, unless it was GPT-6 Sol, in which case new threads move to GPT-6.1 Sol. Threads already running on GPT-6 Sol carry on with it, and GPT-5.6 Terra is still in the picker.

When a build runs on the OpenRouter key saved under Settings, then Bring your own key, a rejected key, an empty OpenRouter balance, or a model your OpenRouter account is not allowed to use now says so and names the fix. Previously these showed a generic credentials error that blamed Anything.

A thread started on a model that has since left the picker, such as GPT-5.6 Sol, GPT-5.6 Luna or GLM 5.2, now shows that model and keeps its model picker, so you can switch it to another model from the same provider. Before, the picker could disappear and then show a different model from the one the thread was actually running on.

If you saved an OpenRouter key under Settings, then Bring your own key, the build agent's model calls now run on your OpenRouter account instead of your Anything balance. If the saved key cannot be read, the build stops with an error rather than charging your balance.

Jomashop category listings now retry temporary request failures without attempting to solve a nonexistent CAPTCHA. Category navigation can fetch data from the browser session when the page cache is empty. Category listings and navigation also retry once with a fresh browser session if CAPTCHA solving is exhausted during initial page preparation.

You can now pick Sonnet 5.5 for a thread. Sonnet 4.6 and Sonnet 5 are still there, and threads already running on either carry on unchanged.

Marketplace searches that name a well-known provider no longer return an unrelated provider merely because its function mentions that brand in a parameter.

You can now pick Opus 5.5, Grok 4.7 and Gemini 3.8 Flash, while GPT-6 Sol and GPT-6 Luna replace their 5.6 versions and GLM 5.3 replaces GLM 5.2. With a ChatGPT subscription connected, GPT-6 Astra, Sol and Luna can run on it too. Threads already on GPT-5.6 Sol, GPT-5.6 Luna or GLM 5.2 carry on with the model they started on.

A heal that died for a reason outside your endpoint, such as the model provider being unavailable, used to count as giving up on that version and turned away every later heal of it for 30 days. Those attempts now leave the endpoint healable, and a heal on a thread built with a connected ChatGPT plan keeps running when that plan's allowance is spent.

The builder checks that an existing workspace can run commands before starting a turn, retries temporary connection failures, and identifies workspace failures separately from model errors.

Anything API now has a public changelog at anything.notte.cc/changelog, listing what was added, improved and fixed in each release. Open it from Resources in the sidebar. Every entry has its own link, so you can point a teammate straight at the change you mean.

The Self-heal tab now has a model picker, so you can decide what each function's heals cost. By default heals run on a lower-cost model from the same provider the function was built with, for example GPT-5.6 Terra for a function built on GPT-6 Astra. The picker lists only models from that provider.

When a reply fails, for example because your connected ChatGPT plan has run out of its Codex allowance, the thread now keeps the error message. Reloading the page or opening the thread later shows why the reply stopped, instead of a question that looks unanswered.

In an organization workspace, any member can now continue, stop or upload files to a thread a teammate started. Before, only the person who created the thread could.

When the agent finishes building a function, the "Function built" card now offers to list it on the marketplace, every time. The agent's summary (what the function does, its arguments, and how to run it) is always written out above it. If a listing is refused, the agent picks it up straight away to fix what it can.

The model picker now stays in the composer after the first reply, so you can move a thread to a faster or cheaper model, or back to a stronger one. A thread can only switch to models from the same provider it started with, and the picker lists just those.

Each function can now self-heal up to 3 times a day, so one busy workspace no longer runs out of heals for all its functions. If a heal gives up without shipping a fix, that version is not healed again until you deploy a new one. The Self-heal tab shows a heal in progress, how many heals are left today and when the limit resets, and any pause on the current version.

Files you attach to a thread can now be uploaded into forms on the website the agent is working on. A thread no longer ends a turn without showing its question when the agent's first attempt to ask one is malformed. Newly built functions now reliably record the site they target, so they show the right icon and appear under the right site.

Threads start responding sooner. New threads now default to GPT-5.6 Terra, and Fable 5.1 and GLM Flash Latest are available in the model picker. GPT-6 Astra now defaults to medium reasoning effort for quicker short turns, and you can still pick a higher effort in the picker.

A verified marketplace function that fails its scheduled check three days in a row loses verification and is marked Needs repair until it is fixed and reviewed again. Functions with two failed days in a row show At risk 2/3. Both are available as views in the marketplace's showing dropdown.

The marketplace, /api/marketplace/search, and the MCP search tool now return the same functions in the same order, with each site's endpoints grouped together and duplicate listings removed. Requests to change something, such as deleting a post, no longer return read-only scrapers for that site, so your agent moves on to build. A category filter now only narrows results and never adds matches the query itself did not find.

About a thousand marketplace functions now describe their nested response fields in detail, not just the top level. Call spec or open the function page to see exactly what comes back before you write code against it.

Documentation now lives at anything.notte.cc/docs and is linked from the home page and the sidebar. The Setup for agents button on the home page copies a prompt that connects Claude Code, Cursor, or Codex to the MCP server and walks your agent through search, spec, run, and build.

The marketplace and function pages now show listings right away and fill in run counts and health metrics as they arrive. If the catalog fails to load, you see an error instead of an empty marketplace.

You can pick GPT-6 Astra in the model picker, and new threads use it unless you choose another model or have saved your own default in settings.

The MCP tools now have REST equivalents under /api/marketplace: GET /search, GET /spec, POST /run, and POST /build, which streams progress as server-sent events. run and build need Authorization: Bearer <NOTTE_API_KEY> and act in that key's workspace, and the full contract is published at /api/marketplace/openapi.json. Search results now include name, so read that instead of the deprecated label field.

The sidebar now shows your 15 most recent chats and links to a Chats page with search, In progress and Unread filters, and rename, archive, stop and delete actions. Failed, needs-you and done threads now have distinct icons, with the status spelled out on hover.

Private and published functions now open on their own page, with tabs for runs, invoking, schedules, self-healing and versions, and the open tab kept in the URL so you can share it. Open any run to see its arguments, output, traceback and logs, or use the new Activity feed for recent runs across your workspace. Schedules preview their next run times in UTC and your local time, and a rollback redeploys the chosen version as a new one.

When a query names a site the marketplace has no functions for, marketplace and MCP search now return no results instead of functions for other sites, so an AI client knows to build one. Category filters also now show only functions in that category.

When the latest health check could not get a result, spec now returns last_check_status: "not_observed" with a last_check_note, and omits last_checked_at. "failed" still means the function itself errored, so update any client logic that treated "error" as broken.

POST /api/anything/start now defaults version to "v2" and answers version: "v1" with a 400 unsupported_version error. If your integration parsed the old event feed, switch it to the message stream the endpoint now returns, and omit version or pass "v2". detach and thread_id now work without setting a version.

Parameters typed as lists or dicts, including optional ones, were sent as text, so Test endpoint failed on functions that work. They are now sent as real JSON, and a list can also be typed as a comma-separated line. The curl, TypeScript and MCP samples now show these defaults as proper collections too.

A site's endpoints now share one page, with a switcher showing each endpoint's runs, recency and 30-day status above the playground and code samples. The marketplace adds a list view for comparing sites side by side, and the catalog no longer stalls for several seconds on some visits.

Notes the agent records while building a function, such as typical run time, known failure modes and which parameters matter, now appear above the parameters in the run dialog. The MCP spec tool returns them in an instructions field, so AI clients can read them before calling run.

A build no longer reads as a log of every command. Consecutive tool steps collapse behind one expandable line that shows what is running live, and a finished build shows a card with the function's name, version, description and the builder's notes.

Every thread now runs on a more resilient build agent, including threads you started earlier, which keep their full history. A build keeps going if you close the tab or lose your connection, and the thread picks it back up when you return. Files you upload stay available to follow-up messages, so you no longer need to send them again, and the limit of three running threads is gone.

Connect your ChatGPT plan under Settings > Codex subscription to use its GPT-5.6 models in threads, billed to your plan instead of per token. When the plan hits its usage limit, the thread tells you so and offers to start a new thread on the standard version of the model.

Asking a question in an existing thread, such as what the agent said last, now gets a direct reply. Only follow-ups that ask for a change trigger a rebuild.

The composer lets you pick from a wider set of models, including Sonnet 5, GPT-5.6, Grok 4.6 and Kimi K3, and set the reasoning effort the model supports. Your choice is remembered for new chats, and a thread keeps the model it started with.

When a function built in a thread fails a run, the thread that built it now wakes up in the background to diagnose the failure and redeploy a fix. Each function gets 3 attempts a week, and a workspace runs one repair at a time, up to 10 a day across all its functions. Repair runs appear as collapsed cards in the thread, and each one opens a page with the full transcript.

In the marketplace playground and the run dialog, JSON responses are now colored so keys, values and nesting are easy to tell apart. A failed run shows its traceback trimmed to your own script, so the line that broke is no longer buried under runner output.

The marketplace front page and the MCP search tool now list only functions that passed review. A newly published function works at its link straight away and becomes searchable once reviewed. If review turns it down, its page shows you the reason, and shipping a new version puts it back up for review.

Each listing page now opens with a 90-day strip showing, day by day, whether the function ran without erroring, backed by a daily check. Listings with no runs show no strip rather than an empty one.

Publishing a function whose target site appears only in a parameter default was refused, while the thread still said the listing was live. These functions now publish, and when a publish does fail, the error is shown clearly and the agent explains it instead of announcing success.

Marketplace search now matches whole words, ranks results that cover more of your query higher, and understands countries in a site's domain, so a search like "grocery store belgium" finds Belgian shops. Queries with no real match return nothing, which lets an MCP agent move on to build, and browsing with no query lists popular sites first.

Publishing no longer fails with "This function needs a website" for functions whose site cannot be guessed from their name, and listings are tagged with the site the code actually uses. The agent's recap after publishing now links to the real listing page.

/api/marketplace/search now honours base_url and category the same way the MCP search tool does, instead of silently ignoring them. An unknown category now returns a 400, so check the value if a request that used to succeed starts failing.

The MCP search tool now returns only what an agent needs to pick a function: function_id, name, description, tagline, domain, categories, run_count and variables. Fields such as response_format and required_secrets are no longer included, so call spec with the chosen function_id if you need them. All MCP tools also answer in compact JSON.

When the agent asks you a question, it now stops and waits for your answer instead of continuing to write and burying the question card. A question still waiting on you is highlighted so it stands out from earlier steps.

POST /api/anything/start accepts "detach": true to return right away with the thread_id and a url to watch the run in the app. Pass a previous thread_id to continue that thread, with the earlier turn's files and work still available. Both options require "version": "v2" in the request body and return a 400 otherwise.

The marketplace and MCP search previously saw only part of the catalog, so some listed functions could not be found. Every listing is now included, and the marketplace loads results page by page, grouped by site, so it stays fast as the catalog grows.

The browser the agent drives now appears as a card in the chat, next to the step that opened it, instead of in a separate panel. Finished sessions stay in place as replays, including when you reopen the thread later.

Threads in the sidebar are grouped into Today, Yesterday, Previous 7 days, Previous 30 days and by month. Marketplace and API Functions sit in the top navigation, while settings, theme, support links and sign out live in a Resources menu and an account card.

Functions you delete disappear from your function list and can no longer be opened by id.

Your own listings carry a Yours badge in the marketplace, and a thread whose function is already listed offers view, update and delete instead of a second publish. When the agent finishes building a private function, it asks whether you want to list it.

Switching workspace now reloads the app into the one you picked, so the sidebar and pages no longer show another workspace's threads. Requests made with an organization's Notte API key, including /api/anything/start, now run in that organization's workspace instead of your personal one.