MCP server

Hosted endpoint#

Stitch AI is published in the official Model Context Protocol server registry, which is how clients discover it without a hardcoded list. The machine readable descriptor lives at /api/mcp/manifest and carries the server identity, chain details, and the JSON schema of every tool below.

The server is read only. It never asks for a private key and never signs anything. Buying access is a wallet action you take yourself, covered in the SDK guide.

Connect a client#

In Claude Code, one command registers the server. Run it in any project and the three tools are available on the next turn.

claude mcp add --transport http stitch https://stitch-ai.net/api/mcp

Claude Desktop reads claude_desktop_config.json. On macOS that is ~/Library/Application Support/Claude/claude_desktop_config.json, on Windows %APPDATA%\Claude\claude_desktop_config.json. Add the stitch entry in its remote form and restart the app. Any client with native HTTP transport takes the same object.

{
  "mcpServers": {
    "stitch": {
      "type": "http",
      "url": "https://stitch-ai.net/api/mcp"
    }
  }
}

Fallback for clients without HTTP transport#

Some older clients still only speak stdio. For those, the mcp-remote bridge proxies stdio to our endpoint. It is a client side shim, not a Stitch server: you are still talking to the hosted one.

{
  "mcpServers": {
    "stitch": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://stitch-ai.net/api/mcp"]
    }
  }
}

Verify it#

The endpoint answers plain JSON-RPC, so you can check it before touching any client config.

curl -s https://stitch-ai.net/api/mcp \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize"}'

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "protocolVersion": "2025-06-18",
    "capabilities": { "tools": { "listChanged": false } },
    "serverInfo": {
      "name": "stitch-memory",
      "title": "Stitch AI Memory",
      "version": "1.0.0"
    },
    "instructions": "Stitch AI is the decentralized knowledge hub for AI agents. ..."
  }
}

Protocol#

Every call is a POST of a JSON-RPC 2.0 message to https://stitch-ai.net/api/mcp. Four methods are implemented: initialize returns the server info above, ping returns an empty result, tools/list returns the three tool definitions, and tools/call runs one.

Arrays are treated as batches and answered with an array of results. Messages without an id are notifications and produce no response entry, so a batch of only notifications comes back as HTTP 202 with no body.

[
  { "jsonrpc": "2.0", "method": "notifications/initialized" },
  { "jsonrpc": "2.0", "id": 9, "method": "ping" }
]

[{"jsonrpc":"2.0","id":9,"result":{}}]

A GET on the same URL returns the server description plus a transport hint rather than a 405, which makes it easy to probe from a browser.

Tools#

Every tools/call result is { content: [{ type: "text", text }], isError }. The text is a plain-text pack listing meant to be read by a model, not a JSON object to parse.

stitch_search_memory#

The tool an agent reaches for on its own. The search is case-insensitive and runs over pack name, category, and content URI, so it finds the pack rather than the passage. Read the pack payload once you know which one you want.

Example call:

{
  "jsonrpc": "2.0",
  "id": 6,
  "method": "tools/call",
  "params": {
    "name": "stitch_search_memory",
    "arguments": { "query": "stock token", "limit": 2 }
  }
}

The text field of the result, rendered:

1 of 8 memory packs matched "stock token".

#2 Stock Token Mechanics & Oracles
  category: onchain
  price: 0.006 ETH
  version: v2
  purchases: 1
  uri: https://pub-dd7413f0f0a94cbfa2cbeb26c62dc9ae.r2.dev/stitch/packs/2/v2.json
  creator: 0xa7c7F04164db3fB97f6D48916F00A8BE52a6C602
  active: yes

stitch_get_pack#

A direct read of one pack from the registry contract, with the human readable category label and ISO timestamps appended to the same block.

Full response for pack 1:

{
  "jsonrpc": "2.0",
  "id": 3,
  "result": {
    "content": [
      {
        "type": "text",
        "text": "#1 Robinhood Chain RWA Primitives\n  category: onchain\n  price: 0.004 ETH\n  version: v2\n  purchases: 1\n  uri: https://pub-dd7413f0f0a94cbfa2cbeb26c62dc9ae.r2.dev/stitch/packs/1/v2.json\n  creator: 0xa7c7F04164db3fB97f6D48916F00A8BE52a6C602\n  active: yes\n  label: On-chain\n  published: 2026-08-31T03:53:58.000Z\n  updated: 2026-08-31T04:17:20.000Z"
      }
    ],
    "isError": false
  }
}

stitch_list_packs#

Pages the catalogue newest first, matching the ordering of the listPacks view on the contract. Use it to let a model choose its own memory rather than hardcoding pack ids.

Result text for a call with a limit of 2:

Showing 2 of 8 memory packs (offset 0).

#8 Arbitrum Orbit Deploy Checklist
  category: developer
  price: free
  version: v2
  purchases: 0
  uri: https://pub-dd7413f0f0a94cbfa2cbeb26c62dc9ae.r2.dev/stitch/packs/8/v2.json
  creator: 0xa7c7F04164db3fB97f6D48916F00A8BE52a6C602
  active: yes

#7 MCP Server Integration Notes
  category: developer
  price: free
  version: v2
  purchases: 0
  uri: https://pub-dd7413f0f0a94cbfa2cbeb26c62dc9ae.r2.dev/stitch/packs/7/v2.json
  creator: 0xa7c7F04164db3fB97f6D48916F00A8BE52a6C602
  active: yes

Errors#

There are two failure layers. A tool that runs but finds nothing returns a normal result with isError set to true, so the model reads the reason and moves on:

{
  "jsonrpc": "2.0",
  "id": 8,
  "result": {
    "content": [
      { "type": "text", "text": "No memory pack with id 99 exists on the Stitch registry." }
    ],
    "isError": true
  }
}

A malformed request never reaches a tool and comes back as a JSON-RPC error instead:

{"jsonrpc":"2.0","id":4,"error":{"code":-32602,"message":"\"id\" is required"}}