MCP for AI Coding Tools

The Model Context Protocol (MCP) is a standard for connecting AI tools to external data sources and actions. 1claw ships an MCP server that exposes vault operations, transaction signing, and more as tools that any MCP-compatible AI assistant can call.

This guide goes beyond the basic MCP setup to cover configuration for specific tools, security best practices, and advanced usage patterns.

Already set up?

If you just need the quick config, see MCP integration. This guide covers deeper integration patterns and per-client configuration.

Supported AI tools

Tool	Transport	Config location
Claude Desktop	HTTP streaming	~/.claude/claude_desktop_config.json
Cursor	HTTP streaming	.cursor/mcp.json (project) or global settings
VS Code (Copilot)	HTTP streaming	.vscode/mcp.json
Claude Code	HTTP streaming	~/.claude/claude_code_config.json
Windsurf	HTTP streaming	.windsurf/mcp.json
Zed	HTTP streaming	settings.json

Automatic setup with the CLI

The fastest path is 1claw setup. It detects installed AI clients and configures MCP for each one:

brew install 1clawAI/tap/oneclaw
1claw setup

This creates an agent, vault, and policy if you do not have them, then writes MCP config files for every detected client. You can target a specific client:

1claw setup --client cursor
1claw setup --client claude-desktop
1claw setup --client vscode

To use an existing agent key:

1claw setup --agent-key ocv_your_existing_key

Manual configuration

Hosted MCP server (recommended)

The hosted server at mcp.1claw.xyz requires no local setup. Configure your AI tool to connect:

Cursor

Create or edit .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "1claw": {
      "url": "https://mcp.1claw.xyz/mcp",
      "headers": {
        "Authorization": "Bearer ocv_your_agent_api_key"
      }
    }
  }
}

Claude Desktop

Edit ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "1claw": {
      "url": "https://mcp.1claw.xyz/mcp",
      "headers": {
        "Authorization": "Bearer ocv_your_agent_api_key"
      }
    }
  }
}

VS Code

Create or edit .vscode/mcp.json:

{
  "servers": {
    "1claw": {
      "url": "https://mcp.1claw.xyz/mcp",
      "headers": {
        "Authorization": "Bearer ocv_your_agent_api_key"
      }
    }
  }
}

Claude Code

claude mcp add 1claw \
  --transport http \
  --url https://mcp.1claw.xyz/mcp \
  --header "Authorization: Bearer ocv_your_agent_api_key"

Local MCP server (stdio)

Run the MCP server locally for lower latency or air-gapped environments:

npm install -g @1claw/mcp

Cursor

{
  "mcpServers": {
    "1claw": {
      "command": "npx",
      "args": ["-y", "@1claw/mcp"],
      "env": {
        "ONECLAW_AGENT_API_KEY": "ocv_your_agent_api_key"
      }
    }
  }
}

Claude Desktop

{
  "mcpServers": {
    "1claw": {
      "command": "npx",
      "args": ["-y", "@1claw/mcp"],
      "env": {
        "ONECLAW_AGENT_API_KEY": "ocv_your_agent_api_key"
      }
    }
  }
}

Local daemon mode (offline-capable)

For fully local operation with encrypted secret storage:

1claw local init         # initialize local vault
1claw daemon start       # start the daemon
1claw setup --local      # configure AI clients for local mode

In local mode, the MCP server connects to the daemon via Unix socket. Secrets are stored in an AES-256-GCM encrypted file at ~/.config/1claw/local-vault.enc. No network calls to the 1claw API unless you explicitly sync.

Available MCP tools

When connected, your AI assistant has access to these tools:

Tool	Description
list_secrets	List secrets in the vault
get_secret	Fetch a secret value
put_secret	Store a secret
delete_secret	Delete a secret
create_vault	Create a new vault
list_vaults	List available vaults
grant_access	Create an access policy
submit_transaction	Sign and broadcast a transaction
sign_transaction	Sign without broadcasting (BYORPC)
simulate_transaction	Run a Tenderly simulation
list_signing_keys	List provisioned signing keys
provision_signing_key	Provision a new signing key
sign_message	EIP-191 personal_sign
sign_typed_data	EIP-712 typed data signing
rotate_generate	Server-side secret rotation
get_env_bundle	Fetch multiple secrets as env vars
inspect_content	Run content through Shroud inspection

Security considerations

Exfiltration protection

The MCP server defaults to block mode for exfiltration protection. If a tool response contains what looks like a secret being exfiltrated (e.g., the AI tries to include a private key in a response to the user), the server blocks it.

To switch to warn mode (logs but does not block):

ONECLAW_MCP_EXFIL_PROTECTION=warn

Secret caching

The MCP server caches secrets for 5 minutes with a 1000-entry LRU limit. This reduces API calls but means recently rotated secrets may be stale for up to 5 minutes.

Agent scope

The MCP server respects the agent's policies. If the agent does not have a policy granting read access to a path, get_secret returns a 403. Configure policies carefully to follow least-privilege.

Rate limiting

The hosted MCP server enforces 60 requests per minute per IP on the HTTP streaming transport. For higher throughput, run the local server.

Usage patterns

Fetch secrets during development

Ask your AI assistant:

"Get the Stripe API key from the vault and use it to list recent charges"

The assistant calls get_secret to fetch the key, then uses it in a subsequent API call. The key stays within the tool context and is not persisted in your code.

Sign transactions from the IDE

"Sign and broadcast a transaction sending 0.001 ETH to 0xdead...beef on Sepolia"

The assistant calls submit_transaction with the parameters. 1claw signs server-side and returns the tx hash.

Rotate credentials

"Rotate the database password in the vault at db/postgres-password and generate a new 32-character alphanumeric value"

The assistant calls rotate_generate with the path and charset parameters.

Further reading

• MCP integration for the basic setup guide
• MCP tools reference for detailed tool documentation
• MCP security for security model details
• CLI guide for 1claw setup options