Skip to content

AnythingMCP

The connector problem is that every system a company runs has its own idea of identity, and the agent is the only participant with none. A customer exists in SAP, in Odoo, on a marketplace, and in a spreadsheet on a file server, with three identifiers and no relationship. AnythingMCP is an open-source, self-hosted gateway that answers both halves of that: it turns 700-plus named systems and any REST, SOAP, GraphQL, OData, SQL or mail source into MCP tools at runtime, and it keeps a knowledge graph of how those systems connect so an agent can link the customer to the orders without being told the join.

Server type:  self-hosted gateway (Docker) or hosted cloud, Streamable HTTP
Hosted:       https://cloud.anythingmcp.com/mcp  -> 401 keyless, OAuth metadata present
Self-host:    http://localhost:4000/mcp behind a first-account bootstrap
Server:       registry record io.github.HelpCode-ai/anythingmcp is at 0.10.0
              repository server.json on main declares 0.53.0
Licence:      AGPL-3.0 (Community), Enterprise for SSO, SCIM, extended File Search
Image:        helpcodeai/anythingmcp:latest, 258.7 MB, amd64 only, 60,677 pulls
Install:      measured on this host and it does not start on arm64 (see below)

The three connector counts, and which one is a measurement

There is no single number, and the gap between them is the honest finding:

README header       732 connectors, 12,000+ tools, 42 need no API key
README feature table 696 catalog connectors, plus your own sources
registry record 0.10.0  255 connectors  (published earlier)
repo server.json 0.53.0  "732 connectors (SAP, ERP, SQL), file search, memory and skills"

The registry holds a stale record. Its exact-name lookup resolves io.github.HelpCode-ai/anythingmcp at version 0.10.0 described as "255 connectors + any REST/SOAP/GraphQL/SQL as MCP tools for Claude & ChatGPT. Self-hosted, AGPL." while the repository's server.json on main declares 0.53.0 with the 732-connector description and a declared remote at https://cloud.anythingmcp.com/mcp. One more quirk that will waste a later pass's time: a registry search for the product name anythingmcp returned no records, and only the fully namespaced io.github.HelpCode-ai/anythingmcp resolved it. Searching a registry by the name on the box is not the same query as searching it by the name in the record.

What the hosted endpoint declares

A keyless initialize against https://cloud.anythingmcp.com/mcp answers HTTP 401 with a JSON-RPC body, {"code":-32001,"message":"Authentication required."}, and WWW-Authenticate: Bearer realm="AnythingMCP MCP Server", resource_metadata="https://cloud.anythingmcp.com/.well-known/oauth-protected-resource/mcp".

The two metadata documents disagree in a way a client has to handle:

protected-resource   resource https://cloud.anythingmcp.com/mcp
                     authorization_servers [https://cloud.anythingmcp.com]
                     scopes_supported []            <- empty
                     bearer_methods ["header"], mcp_versions_supported ["2025-06-18"]

authorization-server issuer, /authorize, /token, /register
                     grant_types authorization_code, refresh_token, client_credentials
                     token_endpoint_auth_methods none, client_secret_post, client_secret_basic
                     scopes_supported ["openid", "email"]
                     code_challenge_methods ["S256"]

The resource document declares no scopes at all while the authorization server advertises only openid and email, so an MCP client cannot learn from discovery which capabilities its connection will carry. On a gateway whose whole purpose is per-tool authorization that is the one document worth fixing first: the consent screen is where an administrator is supposed to grant a narrow slice, and today the client has nothing to request. client_credentials in the grant list is also unusual for a per-person connector and worth understanding before an admin approves one.

The self-host path, measured

The quickstart is deliberately short. A compose file pulls a published image, requires JWT_SECRET and ENCRYPTION_KEY (the app refuses to boot on placeholder values, by design), pins both ports to 127.0.0.1, and starts a postgres:17-alpine behind it. The first account to register becomes admin, open registration is off afterwards, and the endpoint is /mcp on the API port with a /health probe.

This host is aarch64, and that is where the install stopped:

podman pull helpcodeai/anythingmcp:latest
  -> choosing an image from manifest list: no image found in image index
     for architecture arm64, variant "v8", OS linux

podman pull --arch amd64 helpcodeai/anythingmcp:latest
  -> succeeded, digest 1ccaeb4533d4538069c6b9d09cdb7e1275c22a256745407ba6e63a92fcd64d78

podman run ... helpcodeai/anythingmcp:latest
  -> Exited (255) after 8 seconds
     exec /usr/local/bin/docker-entrypoint.sh: exec format error

The Postgres half is not the problem: postgres:17-alpine started on the same podman network and pg_isready answered accepting connections before the app was attempted. The blocker is that the published image is amd64 only with no arm64 manifest, and this host registers no amd64 binfmt handler, so the forced pull lands and then cannot execute. On an Apple Silicon laptop Docker Desktop ships its own emulation and the vendor documents the same pin working that way; on a Linux arm64 server the practical options are qemu-user-static binfmt registration or an amd64 machine. The repository states the cause plainly itself: arm64 is blocked upstream by a Prisma bug, and the quickstart comment says so.

Nothing was therefore driven: no tools/list, no connector, no knowledge graph, no audit line. What follows is the vendor's published design, read from the repository, and each claim is marked as theirs rather than measured here.

What the vendor says it does

  • Runtime tools from any source. 696 catalog connectors plus your own REST/OpenAPI, SOAP, GraphQL, OData, SQL, IMAP/SMTP or MCP source. Tools register at runtime, which is the difference between a fixed tool list and a system added on a Tuesday.
  • File search with page-level answers. Indexes SFTP, FTP/FTPS, WebDAV (Nextcloud, NAS), S3 or a network share, every PDF page and Excel row with its column names, Outlook mails with attachments, and answers with file, page and date.
  • File share. Expiring, revocable links for exported files, internal or public.
  • Company memory. A fact saved in one AI client is found in another, personal and company scopes, with folders and role limits.
  • AI skills. Written procedures, approved by an admin, learned from real requests and corrected errors.
  • Knowledge graph. How the systems connect, so the customer in one system is linked to their orders in another without being told.
  • Code Mode. Several tool calls collapse into one short program run in a sandbox with no secrets and no network; write access is Enterprise.
  • Governance. Per-role, per-tool permissions, response mapping that drops fields before they reach the model, an audit log of every call, SSO and SCIM from the directory (Enterprise).

Editions as published: Community is free under AGPL-3.0 for commercial use up to 3 active users with unlimited connectors and MCP servers; Enterprise adds users, single sign-on, SCIM, File Search beyond a 30-day trial and support. The hosted cloud is the same code run by the vendor in Frankfurt with a 7-day trial, and the vendor states 16,000+ workspaces and 1.5M+ AI tool calls a month, vendor figures unverified from here. GitHub reads 775 stars, 126 forks, repository created 2026-02-28, pushed 2026-11-11, 41 open issues, AGPL-3.0.

Why it earns a page in this catalogue

Because it is the only entry here that treats the tool surface itself as the product, and an operator running four systems behind one assistant will recognise every part of it. The neighbours make the distinction concrete. Cotlet MCP solves the app-with-no-API problem by putting a real Chrome in the loop; AnythingMCP solves it by wrapping the API where one exists and requiring a connector definition where none does. AgentKeyGate answers "how does the agent use a key it never sees" from the credential side; AnythingMCP answers it from the platform side, storing the credential encrypted and dropping fields from the response before the model reads them. n8n MCP automates a workflow, Tuqui MCP reads one ERP, and mcp-billing-gateway meters what a call costs.

The limits, plainly. No handshake was completed and no tool was called, so every capability above is the vendor's description rather than a measured surface. The image does not run on this architecture, the protected-resource document declares no scopes, the registry record is three minor versions behind the repository, and the connector count reads 732, 696 or 255 depending on which document you open. The starting move for an operator on amd64 is the quickstart compose and a first account, which makes the admin bootstrap and the per-tool permission screen verifiable in about half an hour; on arm64, register the binfmt handler first or do not start.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing