Skip to content

Platform Compatibility

Crow uses the open Model Context Protocol (MCP) standard. Any MCP-compatible AI client can connect to Crow's gateway — no vendor-specific extensions are used.

Compatibility Matrix

Local clients work today; cloud web clients cannot reach a private Crow

Clients that run on your own machine (desktop apps, CLIs, IDEs) reach your self-hosted Crow directly over LAN/tailnet — they work out of the box. Clients that run in a vendor's cloud (claude.ai web/mobile, ChatGPT, Grok on the web) must reach your Crow from the public internet, and a self-hosted Crow is deliberately not exposed there. Cloud web-client access is planned as a follow-on after v1, behind a dedicated security review.

PlatformRunsTransportAuthStatus for a private Crow
Claude Web & MobileVendor cloudStreamable HTTPOAuth 2.1Not reachable — planned follow-on
Claude DesktopYour machinestdioN/A (local)Fully tested
Claude Code (CLI)Your machinestdio / HTTPOAuth 2.1Fully tested
ChatGPTVendor cloudSSE / HTTPOAuth 2.1Not reachable — planned follow-on
GeminiYour machine (CLI)stdio / HTTPOAuth 2.1Compatible
Grok (xAI)Vendor cloudStreamable HTTPBearer tokenNot reachable — planned follow-on
CursorYour machinestdio / HTTPVariesCompatible
WindsurfYour machinestdio / HTTPVariesCompatible
ClineYour machinestdio / HTTPVariesCompatible
Qwen CodeYour machinestdio / HTTPOAuth 2.1Compatible

Mobile Apps

The table above covers MCP clients (AI assistants that connect to Crow's tools). Crow also has dedicated mobile access to the Crow's Nest dashboard itself:

PlatformInstall methodStatus
Android AppNative APK, or PWA via ChromeFully tested
iPhone (PWA)Installable web app via Safari (no App Store)Fully tested

MCP Endpoints

Every path is relative to your gateway URL (e.g. http://crow:3001). Each server is available over Streamable HTTP at <prefix>/mcp and over legacy SSE at <prefix>/sse + <prefix>/messages:

PrefixServerNotes
/routerCategory router (recommended)10 consolidated tools instead of the full 126+ raw surface — see Context & Performance
/memoryMemoryThe bare /mcp path is a compatibility alias for this server
/projectsProjects/research is a legacy alias — same server, older name
/sharingSharing
/storageStorageAvailable only when MinIO is configured
/blog-mcpBlog
/toolsExternal tool proxyIntegrations (GitHub, Trello, …) aggregated into one endpoint

Naming aliases

The projects server was previously called research. Old configs using /research/mcp or the crow_research router tool keep working — they are aliases for /projects/mcp and crow_projects.

Transport Types

Crow's gateway supports two MCP transport protocols:

  • Protocol version: 2025-03-26
  • Endpoints: <prefix>/mcp from the table above
  • Used by: Claude, Gemini, Grok, Cursor, Windsurf, Cline, Claude Code

SSE (Legacy)

  • Protocol version: 2024-11-05
  • Endpoints: <prefix>/sse + <prefix>/messages from the table above
  • Used by: ChatGPT

stdio (Local Only)

  • Direct process communication, no network
  • Used by: Claude Desktop, Claude Code (local), Gemini CLI (local), Qwen Coder CLI (local), Cursor (local), Windsurf (local), Cline (local)

Authentication

The gateway uses OAuth 2.1 with Dynamic Client Registration. When you connect a new client, it automatically:

  1. Discovers the OAuth metadata at /.well-known/oauth-authorization-server
  2. Registers itself as a client via /register
  3. Redirects you to authorize at /authorize
  4. Receives an access token via /token

This is the same standard flow used by most OAuth providers. No manual token management needed for platforms that support OAuth discovery.

For platforms that don't support OAuth discovery (like Grok), you can use the /introspect endpoint or configure bearer tokens manually.

Cross-Platform Context (crow.md)

Crow goes beyond shared data — it also shares behavioral context across platforms. The crow.md document defines how Crow behaves: identity, memory protocols, transparency rules, and your customizations.

Automatic delivery: When any AI connects to Crow, it receives a condensed version of your behavioral context during the MCP handshake — before any tool calls happen. The AI immediately knows how to use memory, follow session protocols, and respect transparency rules. No user action required.

On-demand guidance: For detailed workflow instructions, the AI can request MCP prompts like session-start, crow-guide, research-guide, blog-guide, or sharing-guide. These provide comprehensive guidance without consuming context window space upfront.

Manual access: Use the crow_get_context tool (any MCP platform) or GET /crow.md (HTTP endpoint) for the full document.

See the Cross-Platform Guide for a complete walkthrough.

Released under the MIT License.