Claude Web & Mobile

Not available for self-hosted Crow (planned as a follow-on)

claude.ai and the Claude mobile app connect to MCP servers from Anthropic's cloud, so they can only reach a server that is publicly accessible on the internet. A self-hosted Crow is deliberately not reachable from the internet — that privacy posture is the product's core promise, not a missing feature. There is currently no supported way to connect claude.ai or the Claude mobile app to a private Crow. Public cloud-client access is planned as a follow-on after v1, behind a dedicated security review.

Use instead: Claude Desktop or Claude Code — both run on your own machine, reach your private Crow directly, and share the same memories.

The instructions below apply only to the unsupported, advanced case where you have made your gateway publicly reachable yourself (your own domain and reverse proxy — read SECURITY.md first). Tailscale Funnel is not such a path: Crow's gateway rejects funneled requests to MCP and OAuth routes by design.

Prerequisites

• Crow gateway deployed, healthy, and publicly reachable (unsupported — see above)
• A Claude plan that supports custom connectors (Free currently includes one connector; paid plans allow more)

Setup Steps

1. Go to claude.ai/settings → Integrations
2. Click Add Custom Integration
3. Enter a name (e.g., "Crow Memory")
4. Paste your gateway URL:

   https://your-crow-server/memory/mcp

5. Click Save — Claude will initiate the OAuth flow
6. Authorize the connection when prompted

Repeat for each server you want to connect:

Server	URL
Memory	https://your-crow-server/memory/mcp
Projects	https://your-crow-server/projects/mcp
Sharing	https://your-crow-server/sharing/mcp
Storage	https://your-crow-server/storage/mcp
Blog	https://your-crow-server/blog-mcp/mcp
External Tools	https://your-crow-server/tools/mcp

Note: The storage endpoint requires MinIO to be configured. See the Storage Guide for setup.

Transport

• Type: Streamable HTTP
• Protocol: 2025-03-26
• Auth: OAuth 2.1 (automatic)

Self-Hosted Crow

A standard self-hosted Crow cannot be connected to claude.ai — see the notice at the top of this page. In particular, Tailscale Funnel does not work for this: the gateway rejects funneled requests to every MCP and OAuth route (only the blog and a few public files are Funnel-safe). Earlier versions of this page documented a Funnel-based setup; that flow has never worked on current Crow and the instructions were removed.

Your self-hosted Crow works today with the local clients on the platforms index — Claude Desktop and Claude Code give you the same Claude models with full access to your private Crow.

Verification

After connecting, try asking Claude:

"Store a memory that Crow is now connected."

If it works, you'll see Crow's memory tools in action. You can verify stored memories by asking:

"What do you remember?"

Troubleshooting OAuth

If the connector fails to connect or tools don't appear:

• Leave OAuth Client ID and Client Secret blank. Crow uses Dynamic Client Registration (RFC 7591) — Claude discovers the OAuth endpoints automatically from the .well-known metadata. No pre-configured credentials needed.
• Verify your gateway's OAuth metadata is reachable. Visit https://your-crow-server/.well-known/oauth-authorization-server in your browser — you should see a JSON response with authorization_endpoint, token_endpoint, and registration_endpoint.
• Check CROW_GATEWAY_URL in your .env file. This must match your actual public URL exactly (including https://).
• Start a new conversation after gateway restarts. OAuth sessions are tied to the gateway process. When the gateway restarts, existing Claude.ai sessions become invalid — start a fresh conversation to re-authenticate.
• Check gateway logs for OAuth errors. If running via systemd: journalctl -u crow-gateway -f. Look for errors in the /register, /authorize, or /token endpoints.

Cross-Platform Context

Crow automatically delivers behavioral context during the MCP connection handshake. When Claude connects to any Crow server, it receives Crow's identity, memory protocols, session protocol, transparency rules, and capability reference — no user action required.

This means Claude knows to recall relevant memories at the start of each conversation, store important information, and follow transparency rules from the very first message.

For deeper guidance, Claude can request MCP prompts:

• session-start — Detailed session start/end protocol
• crow-guide — Full crow.md document with all behavioral sections
• research-guide / blog-guide / sharing-guide — Workflow guidance for specific features

These prompts serve as skill equivalents, giving Claude detailed workflow instructions on demand without consuming context window space upfront.

You can also manually load the full context:

"Load your crow.md context"

See the Cross-Platform Guide for more details.

Context Optimization

Claude Code supports toolListChanged notifications, making crow-core a good fit for local use — it activates servers on demand rather than loading all tools upfront.

For Claude via the gateway, the /router/mcp endpoint reduces the tool count from 126+ to 10 consolidated category tools, significantly reducing context window usage. Instead of connecting each server individually, you can connect a single router endpoint:

https://your-crow-server/router/mcp

See the Context & Performance guide for more details.

Shared with Claude Code CLI

Custom Integrations you add here are also available in Claude Code CLI sessions — they share the same connector configuration within Anthropic's ecosystem. If you set up Crow on claude.ai, it works in Claude Code without additional setup. This cross-platform sharing is specific to Claude; other platforms (ChatGPT, Gemini) manage their MCP connections independently.

Tips

• You can connect all five servers (memory, projects, sharing, storage, blog) plus external tools simultaneously
• The mobile app uses the same Custom Integrations as the web
• Tools from external integrations (GitHub, Slack, etc.) appear through the /tools/mcp endpoint
• Memories stored here are instantly accessible from ChatGPT, Gemini, or any other connected platform