Skip to content

v0.55.0: learning projects, file uploads from AI clients, 93 new connectors

Choose a tag to compare

@helpcodeai helpcodeai released this 16 Nov 07:00
dc6a366

AnythingMCP 0.55.0: learning projects, file uploads from AI clients, 93 new connectors

This release collects everything merged since 0.54.0. The main addition is learning projects: learning turned on for one member on one MCP server, so that what an experienced colleague teaches the AI becomes skills for the whole team. It also lets an AI client hand a file it created to a connector tool, repairs the automatic skill generation, adds an optional one-tap setup for Telegram bots, checks catalog arguments before a call is sent, keeps database passwords out of connector URLs, and grows the catalog to 825 connectors.

Learning projects

  • Keep what one colleague teaches the AI. An admin starts a project for one member on one MCP server. When that member corrects the AI or explains how the work is really done, the lesson is kept, and the lessons of a topic become a skill for everyone on that server (#1708, #1709, #1711, #1712). Guide: docs/learning-projects.md.
  • The member's own AI client writes the lesson. Three tools appear on /mcp/<serverId> for the project's member only: learning_record (a rule, fact, correction or worked example, filed under a topic), learning_review and learning_confirm. The model is told when to call them in the server's instructions, in the tool description, and in one line after every tool that may change something. No AI key is needed on the instance.
  • One skill per topic, built from lessons. The skill text is put together from the topic's lessons, so a new lesson never rewrites the others, and the same lesson said twice is counted once.
  • A draft is served to its member only. They see the model improve at once and correct it again when needed. A person publishes a topic, and only confirmed lessons reach other users. A published skill is an ordinary AI skill with versions, a folder and usage counts, labelled from a learning project.
  • Consent first. Nothing is kept before the member accepts under Settings → Profile, and the notice they accepted is stored. They can stop at any time and erase everything that was not published. Their opt-out of skill learning also stops every project. The member's name is shown to admins only, unless the project is set to show it.
  • Dashboard: Learning projects (new, under Intelligence) to start and manage projects, review lessons, preview and publish topics. English, Italian and German.
  • API: api/learning/*, also available as amcp api operations. Security events LEARNING_PROJECT_CREATED, LEARNING_PROJECT_CONSENT, LEARNING_PROJECT_STATUS, LEARNING_PUBLISHED, LEARNING_ERASED.
  • Plans: one project at a time on Cloud Team (and the Cloud trial), no limit on Enterprise (Cloud and self-hosted), one in the self-hosted 30-day evaluation. Not included in Cloud Starter and self-hosted Community. A downgrade removes nothing: the oldest projects keep working up to the plan's number.

File uploads from AI clients

  • A file an AI client created in its own sandbox can reach a connector tool (a listing photo, a ZIP for a digital product). On servers that have a tool with a file parameter, file_upload_link gives a one-time upload link (15 minutes) and file_public_url a signed public link (one hour; images, videos and PDFs) for providers that download files themselves. A file parameter accepts the reference amcp-file:<id> (#1717). Guide: docs/connectors/file-uploads.md.
  • A reference only works for tools of the workspace that uploaded the file. Files are deleted 8 hours after upload.
  • On by default in every edition, Cloud and self-hosted. It uses the File Share storage (the file_share_data volume, or FILES_STORAGE=s3). FILE_UPLOADS_ENABLED=false turns it off.
  • Limits: on Cloud per plan (25 MB per file on the trial, 50 MB on Starter, 100 MB on Team and Business, 200 MB on Enterprise); on a self-hosted instance by environment variables (100 MB per file unless FILE_UPLOADS_MAX_FILE_BYTES says otherwise).
  • For uploads from Claude's code execution, allow your AnythingMCP domain in Claude's network settings (Settings → Capabilities).

AI skills

  • Skill generation no longer repeats itself. A run tells the model which skills a scope already has and does not file a suggestion that repeats one you applied or dismissed (#1697).
  • Generated skills always wait for review. Auto-apply no longer acts on the confidence a model gives its own suggestion. The switch now covers skills learned from an error that was fixed the same way at least 3 times (#1697). If you had the switch on, see the upgrade notes.
  • The Scheduled AI extension switch is shown only where the scheduled job is turned on (KG_LLM_CRON_ENABLED=true).

Connectors and catalog

  • 825 connectors in the catalog (from 732), among them seller connectors for more than 40 retailer marketplaces, Walmart, Target Plus, Temu, SHEIN, Lazada, Miravia and AliExpress seller, Fnac Darty, Joom, Ceneo, Skroutz, OpenCart on the shop database, USPS, Microsoft Advertising, HORNBACH and a read-only Strava adapter (#1673, #1679).
  • A catalog tool's argument that does not match its declared pattern is refused before the request is sent, with the value, the argument's description and the pattern in the message (#1702).
  • A connector variable can offer a fixed list of choices (envVarMeta.<VAR>.options), shown as a select on the setup page and described to the model in the chat setup (#1678).
  • Adapter engine: a per-request ${uuid} placeholder in login-token headers; fixed arguments on every tool call of an MCP adapter; XML logins and login tokens inside the request body or read from a response header; OAuth2 tokens read from an envelope; plain digest signatures inside a JSON body; authorization and token renewal at Alibaba IOP gateways; 64-bit ids kept exact.
  • DATABASE adapters guard variables that are used as SQL identifiers.
  • Searches for a retailer go to its own connector, not to the generic Mirakl adapter.

Setup

  • Telegram: create the bot in one tap (optional, off by default). With a manager bot configured (TELEGRAM_MANAGER_BOT_TOKEN, TELEGRAM_MANAGER_WEBHOOK_SECRET, see .env.example), a user creates their own bot from the page after approving an AI client, without a chat with @Botfather or a pasted token. The page that offers it exists on Cloud only today (#1693, #1696, #1715).
  • Etsy: the setup opens Etsy's seller app form, and an approved app is followed up until the user has signed in (#1694, #1703). Shopify: the setup is guided and Shopify's refusals are named (#1707).
  • The setup page says why Authorize cannot start when the app keys are not saved (#1663).
  • Continue after approving an AI client no longer opens a chat tab on a phone, which broke the connection there (#1690).
  • Telegram reads that return nothing are explained, and close Odoo field names are suggested (#1701).

Security and reliability

  • Database passwords are kept out of the connector's base URL. A DATABASE connector saved as postgresql://user:password@host/db now stores the username and password in the encrypted authConfig; a password still in a stored URL is masked in every connector response and in exports for anyone but an admin (#1677). After upgrading: a database connector saved before this version keeps its password in the stored URL until it is saved again. Open each database connector, choose Edit and save it once, then use "Test connection". Before this version the full connection string was shown to every member who could open the connector, so consider changing those database passwords.
  • A refused API key is noticed sooner and its owner is told. Three refusals in a row mark the connector (down from five), counted across restarts when Redis is available, and the owner gets an email (#1704).
  • WSDL downloads that are not XML are refused, and unquoted OpenAPI versions are read (#1687).
  • "Never used" no longer follows from old calls having been pruned: stored first-success markers are used (#1706).

Cloud

  • The paywall, the trial page and the trial banner are translated into all seven website languages and say "payment method" (#1685, #1688).
  • A workspace cannot be deleted while Stripe still retries its subscription (#1676).
  • The sign-up trial is asked for once, then again in the background (#1684). The steps of connecting an AI client are recorded (#1692, #1698).

CLI 0.1.7

  • amcp api --list and the coverage test know the 16 learning project operations. Install or update with npm i -g @anythingmcp/cli.

Upgrade notes

  • Two database migrations, both of which only add: 20261115160000_learning_projects (three new tables and four columns on kg_skill_suggestions) and 20261115210000_file_uploads (one new table). They run by themselves at start, and the previous version runs against the migrated schema, so a rollback to 0.54.0 leaves the schema as it is.
  • No new required settings. Optional: the three TELEGRAM_* variables above and the FILE_UPLOADS_* limits.
  • File uploads are on after the upgrade, in every edition. Set FILE_UPLOADS_ENABLED=false if you do not want them.
  • Auto-apply changed. If Auto-apply was on, skills the model generates now stay under Suggested until you apply them, and skills learned from fixed errors seen at least 3 times are applied automatically. Turn the switch off if you want to review those too.
  • Learning projects on a self-hosted instance need an Enterprise licence key, or the 30-day evaluation (one project).
  • After a member accepts a learning project, they refresh the connector's tool list in their AI client (in claude.ai: Customize → Connectors → the connector → ⋮ → Refresh tools list).