Clavis

Self-Hosted Credential Infrastructure for Autonomous Agents

Your AI agents need API keys, OAuth tokens, and wallet keys.
Hardcoding them = one prompt injection away from disaster.

Clavis stores credentials server-side and injects them at runtime.
Your agents never see the secrets.

Installation

pip install clavis

Quick Start

from clavis import ClavisClient

client = ClavisClient(
    api_key="your-jwt-token",       # from POST /v1/auth/login
    base_url="https://your-clavis-instance.com"
)

# Secure call — credential injected server-side, never touches agent memory
response = await client.call(
    "stripe",
    "GET",
    "https://api.stripe.com/v1/balance"
)
print(response["body"])  # {"object": "balance", ...}

# Check credential health without making external calls
status = await client.check_credentials("stripe")
# {"valid": True, "status": "healthy", "expires_in": 3542}

# Proxy a request (path-based, credential injected server-side)
response = await client.proxy(
    "openai",
    "POST",
    "/v1/chat/completions",
    body={"model": "gpt-4o", "messages": messages}
)

Why Clavis

Security Model

clavis.call() is the core primitive. When your agent calls it:

1. Clavis fetches the credential from its encrypted vault
2. Injects it into the upstream HTTP request server-side
3. Returns only the API response — no credential ever leaves the server

Even a fully compromised agent cannot exfiltrate credentials it never received.

Features

• call() — Server-side credential injection; agent never sees raw secrets
• proxy() — Path-based proxying with automatic auth
• check_credentials() — Dry-run health check; no external calls made
• get_token() — Retrieve a valid token for custom auth flows
• Encrypted credentials at rest (Fernet)
• Automatic OAuth token refresh
• Redis-backed rate limiting (cross-agent)
• Full audit log on every auth event
• MCP server for Claude Desktop integration

Self-Hosting

Clavis is fully self-hostable. See clavisagent.com/docs/mcp.html for setup.

License

MIT