Context4 MCP¶

One endpoint, seven business services, and a permission per service. Context4 sits between an assistant and the systems a business actually runs on. The agent searches for a tool, asks for its schema, and executes it against an account the operator has explicitly connected.

Server type: Remote (Streamable HTTP)
Endpoint:    https://context4.me/mcp
Transport:   JSON-RPC 2.0 over HTTP, SSE responses
Auth:        OAuth 2.1, discovery at /.well-known/oauth-protected-resource/mcp
Scope:       mcp:tools (one scope at the door, permissions per connected account)
Tools:       5 stable meta-tools, 164 provider actions behind them
Built by:    Context4 (context4.me)

Why This Matters for Operators¶

The assistant does not need a hundred tools. It needs a reliable way to find the right one and a record of who allowed it.

That is the shape Context4 chose. Rather than exposing every CRM field as a separate MCP tool, the server exposes five meta-tools and resolves the rest at call time from a published action catalog. Two consequences follow. The tool list stays stable as services are added, so an operator's client does not break every time a connector ships. And the permission boundary sits at the connected account, not at the tool name, so revoking a service is one action in a console rather than a client reconfiguration.

What It Actually Does¶

The endpoint was probed live. An unauthenticated POST initialize to https://context4.me/mcp returned HTTP 401 with a structured error, {"type":"urn:context4:error:authentication_required","title":"Authentication required","status":401,"code":"authentication_required","request_id":"..."}, so the route is live and fronted by its own error contract rather than a generic 404.

Discovery answers properly. /.well-known/oauth-protected-resource/mcp returns 200 with resource https://context4.me/mcp, authorization server https://context4.me, scope mcp:tools, bearer in header and a documentation pointer. /.well-known/oauth-authorization-server adds authorize, token, register and revoke, authorization_code plus refresh_token, and token_endpoint_auth_methods_supported: ["none"], which is a public client with PKCE, so nothing secret is stored on the operator's machine.

The five meta-tools are documented with full JSON Schemas and strict input validation, which rejects unknown fields and caps arguments at 65,536 serialized characters:

Meta-tool	What it does
`CONTEXT4_SEARCH_TOOLS`	Finds provider actions for a task, returns connection readiness, filters on tags such as `read-only`, `destructive`, `idempotent`
`CONTEXT4_GET_TOOL_SCHEMAS`	Returns the exact input and output schema for a discovered action
`CONTEXT4_MANAGE_CONNECTIONS`	Starts and manages the connection to a provider account
`CONTEXT4_WAIT_FOR_CONNECTIONS`	Blocks until a consent flow finishes
`CONTEXT4_MULTI_EXECUTE_TOOL`	Runs one or more discovered actions against connected accounts

The provider actions come from the vendor's own catalog file, https://context4.me/docs/connectors.json, read in this sweep: seven connectors, all marked beta, 164 actions in total.

Connector	Version	Auth	Actions
amoCRM	2.2.0	OAuth 2	21
Bitrix24	1.1.0	OAuth 2	23
GitHub	1.1.0	OAuth 2	24
Grafana	1.2.0	API key	38
Linear	1.4.0	OAuth 2	18
Pipedrive	1.0.0	OAuth 2	15
Sentry	1.1.0	API key	25

A parallel HTTP API exists at https://context4.me/v1, and a Composio-compatible base URL at https://context4.me/compat/context4/v1 names where the execution layer comes from: the provider calls are proxied, which is why one endpoint can carry seven services.

Installation¶

Add the remote server to any client that supports streamable HTTP and let OAuth discovery do the rest.

claude mcp add --transport http context4 https://context4.me/mcp

The vendor publishes client guides for ChatGPT, Claude, Codex, Claude Code, Hermes and OpenClaw, plus an llms.txt and an agent index at https://context4.me/docs/agent-index.json. Accounts are connected in the console at https://console.context4.me/integrations, and access can be revoked per service there.

Business Relevance¶

Ask a question that spans two systems and the answer needs both. "Which deals did the client mention in last week's support incident?" is a CRM record plus an error report. "Which open issues are blocking the billing migration?" is Linear plus GitHub.

An assistant given read access to a CRM can answer those. An assistant given the action layer can also file the follow-up task in the same breath. What this catalog cares about is the middle: does the operator know, and control, what each service allowed. Here the answer is yes, per connected account, with a single scope at the door and a per-service switch in the console.

Integration with CorpusIQ¶

The services Context4 fronts are the same class CorpusIQ connects directly: CRM, project tracking and engineering telemetry as read-only sources, with the operator keeping the account. It is catalogued as a comparison point and as a route for teams already standardised on amoCRM or Bitrix24, which this catalog had no entry for.

Limitations¶

No tool call was executed in this sweep. Every action needs an OAuth session, a workspace and a connected provider account, none of which this pass creates. The 401, the two discovery documents and the published catalog are the evidence; the executed path is the gap.
No entry in the official MCP registry: registry.modelcontextprotocol.io search for context4 returned zero servers on October 25, 2026.
All seven connectors are marked beta, which is stated by the vendor rather than inferred.
Documentation and the console are in Russian. The MCP contract is machine-readable and English-keyed, but an English-only operator is reading vendor pages through translation.
The vendor's homepage illustrations show services that are not in the published catalog, including Wazzup and FinTablo, and the markup labels those as illustrations rather than live data. Treat connectors.json as the real list.
No public price list was found on the pages read. The site advertises a free start, and plan limits sit inside the tools rather than at the door.
The catalog already carries CRM servers at product level, including OnePageCRM, Perfex CRM and onvela CRM. What none of them carry is Pipedrive, Bitrix24 or amoCRM, or a single gateway over several business services with per-account consent, which is the slot this entry fills.

FAQ¶

Do I need a paid plan?¶

No price list was published on the pages read in this sweep, and the site advertises a free start. Plan limits are enforced inside the tools, so a capped call returns an error rather than refusing the connection.

Which services can an agent reach?¶

Seven, from the published connector catalog: amoCRM, Bitrix24, Pipedrive, Linear, GitHub, Sentry and Grafana. All are marked beta. Anything else is not reachable until it appears in that catalog.

Can the agent change data?¶

Yes, and that is the point of the action layer. Individual actions carry tags including read-only and destructive, so a client can search only for reads, and each connected account is separately revocable in the console.

Where does the credential live?¶

Nowhere on the operator's machine. The client registers as a public OAuth client with PKCE, and provider tokens stay server-side against the connected account.