Docs

MCP Integration Guide

One MCP endpoint, many host clients. AIS exposes https://ais.agentsandswarms.ai/mcp as a remote MCP server. Every tool-using agent that speaks MCP can read and write memory with no API key — OAuth Device Flow on first connect.

Choose your host

Skip directly to your IDE / agent host. If yours isn't listed, jump to stdio fallback — it works with any MCP-compatible client.

Claude Code Cursor Zed Claude Desktop Codex (OpenAI) OpenCode Continue.dev Claude.ai (web) Other (stdio)

Claude Code

One command in your terminal. Claude Code registers the remote server and handles the OAuth Device Flow on first run.

Terminal

claude mcp add --transport http ais-memory https://ais.agentsandswarms.ai/mcp

Cursor

Add to ~/.cursor/mcp.json for global use, or .cursor/mcp.json in a project root for workspace-only. Cursor infers the HTTP transport from the url field — there is no separate transport key.

~/.cursor/mcp.json

{
  "mcpServers": {
    "ais-memory": {
      "url": "https://ais.agentsandswarms.ai/mcp"
    }
  }
}

Cursor ≥ 0.39 also supports a one-click cursor://anysphere.cursor-deeplink/mcp/install deeplink — see the Get Connected buttons on the homepage.

Zed

Add to ~/.config/zed/settings.json. Zed calls these context servers rather than MCP servers, but they are the same MCP protocol underneath.

~/.config/zed/settings.json

{
  "context_servers": {
    "ais-memory": {
      "url": "https://ais.agentsandswarms.ai/mcp"
    }
  }
}

Claude Desktop

Claude Desktop's claude_desktop_config.json only speaks stdio — it can't talk to remote HTTP MCP servers directly. Use the published aismemory npm package; it authenticates via OAuth Device Flow on first run and surfaces 7 agent-native tools.

claude_desktop_config.json

{
  "mcpServers": {
    "ais-memory": {
      "command": "npx",
      "args": ["-y", "aismemory"]
    }
  }
}

File location: ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows.

Want the full 87-tool HTTP surface? Use Claude.ai Custom Connector (web — no local config), or wait on CORBOT-85716A29 which adds OAuth Authorization Server Metadata to AIS so mcp-remote bridging will work.

Codex (OpenAI)

Add to ~/.codex/config.toml. The Codex CLI uses TOML for MCP server entries. We use the aismemory npm package directly (stdio + OAuth Device Flow) — same 7-tool surface as Claude Desktop.

~/.codex/config.toml

[mcp_servers.ais-memory]
command = "npx"
args = ["-y", "aismemory"]

If your Codex version supports remote HTTP MCP servers natively and you don't want to go through the npm bridge, you can use url = "https://ais.agentsandswarms.ai/mcp" directly — but auth handling is up to your client.

OpenCode

Add to ~/.opencode/mcp.json (or the project-local .opencode/mcp.json). OpenCode uses the same shape as Cursor's mcp.json.

~/.opencode/mcp.json

{
  "mcpServers": {
    "ais-memory": {
      "url": "https://ais.agentsandswarms.ai/mcp"
    }
  }
}

Continue.dev

Add to ~/.continue/config.json (or your workspace .continue/config.json). Continue currently routes MCP through its experimental section and prefers stdio transport, so we use the aismemory npm package directly (same as Claude Desktop and Codex).

~/.continue/config.json

{
  "experimental": {
    "modelContextProtocolServers": [
      {
        "transport": {
          "type": "stdio",
          "command": "npx",
          "args": ["-y", "aismemory"]
        }
      }
    ]
  }
}

Verify against the current Continue docs — MCP support has moved location across Continue versions.

Claude.ai Custom Connector (web)

No local config required. Add AIS as a Custom Connector in your Claude.ai account — Claude.ai handles auth and tool surfacing across web, mobile, and desktop sessions tied to that account.

1. Open claude.ai/connectors in a logged-in browser.
2. Click Add custom connector.
3. Paste the MCP URL: https://ais.agentsandswarms.ai/mcp
4. Approve the OAuth bond. Done — every Claude.ai chat on that account now has remember / recall / whoami.

Custom Connectors in Claude.ai require a Pro / Team / Enterprise plan. Free Claude.ai accounts can still use AIS via Claude Desktop or Claude Code (above).

Other MCP clients (stdio fallback)

For MCP clients that only speak stdio and don't support remote HTTP servers, install the published aismemory package. It authenticates via OAuth Device Flow on first run and proxies every tool call to the HTTP endpoint.

MCP config (stdio)

{
  "mcpServers": {
    "ais-memory": {
      "command": "npx",
      "args": ["-y", "aismemory"]
    }
  }
}

Works with any MCP client: Anthropic clients, third-party editors, custom agent frameworks. The stdio package surfaces 7 agent-native tools; the HTTP endpoint above surfaces all 87.

Troubleshooting

"Server not found" or 404 on https://ais.agentsandswarms.ai/mcp

Make sure your client supports remote HTTP MCP servers. If it doesn't, use the stdio fallback — that path works everywhere.

OAuth flow opens but never returns

Some hosts intercept the browser callback. After approving in the browser tab, return to your IDE — the server polls for approval and finishes connection within 5 seconds.

Tools work but recall returns nothing

You're probably querying a different agent than the one your memories are bound to. Run whoami to confirm identity, or set AIS_AGENT_NAME in your MCP config env block to pin the session.

Source of truth

The canonical reference for tool surfaces, counts, and version drift lives alongside the AIS source code. If something here ever diverges from what the API actually exposes, the source file wins.

HTTP MCP source: packages/agent-identity-service/src/mcp/tools/*.ts in the AIS repo. 87 tools across 13 categories.
npm stdio source: packages/agentmemory/src/index.ts. 7 tools.
Live API: https://ais.agentsandswarms.ai/mcp — call tools/list to enumerate.