agent-authority 0.1.6

Authorization for AI agents: verifiable, scoped, revocable capability tokens (mandates) with attenuable delegation, for MCP and A2A. The reference implementation of agent authority — zero-dependency TypeScript & Python.

agent-authority (Python)

Authorization for AI agents — the reference implementation of agent authority (project name: Behalf). Verifiable, scoped, time-bound, revocable capability tokens (mandates) with attenuable, macaroon/biscuit-style delegation, plus MCP and agent-to-agent (A2A) middleware. Capability-based security, OAuth 2.1 on-behalf-of–style grants, and SPIFFE/SVID-style agent identity for multi-agent / LLM systems.

Identical API shape to the TypeScript library, zero dependencies (standard library only). The optional cryptography extra enables a constant-time Ed25519 backend and sealed credentials:

pip install agent-authority          # core, dependency-free
pip install "agent-authority[seal]"  # + sealed credentials / hardened crypto

In plain words

Think of a Mandate as a permission slip for an AI agent.

Imagine you hire an assistant to run errands for you. You don't hand over your wallet and house keys — you write a note: "You may read my calendar and spend up to $50, and only for the next hour." That note is a Mandate.

The five verbs are just the things you can do with that note:

• grant — write the permission slip. "This agent may do X, up to this limit, until this time."
• authorize — check the slip before acting. The agent must show the slip (and prove it's really theirs) before it's allowed to do something.
• attenuate — make a smaller copy for a helper. A sub-agent can only get the same powers or fewer — never more.
• revoke — tear the slip up. Cancel it instantly; every copy handed downstream stops working too.
• audit — the logbook. Every check is written down, so you can see exactly what happened.

Two things make the slip safe: it can't be faked or upgraded (it's signed with cryptography, and a helper can only shrink the powers), and holding the paper isn't enough — an agent must prove it's the rightful holder, so a stolen copy by itself is useless.

The rest of this page shows how to do each of these in code.

The five verbs

from agent_authority import create_behalf

b = create_behalf()

# 1. GRANT
mandate = b.grant(
    principal="alice",
    agent="research-agent",
    can=["read:calendar", "spend:usd<=50"],
    expires_in="1h",
)

# 2. AUTHORIZE — raises AuthorizationError if denied
mandate.authorize("spend:usd=20")

# 3. ATTENUATE — narrow for a sub-agent; can only shrink
child = mandate.attenuate(can=["read:calendar"], expires_in="10m")

# 4. REVOKE — kills the mandate and its downstream chain
b.revoke(mandate.id)

# 5. AUDIT — every authorize() wrote a hash-chained record
trail = b.audit(mandate.id)

MCP / A2A middleware

from agent_authority.mcp import with_behalf

server = with_behalf(
    my_tool_server,  # exposes call_tool(name, args, ctx=None)
    policy={
        "send_email": "write:email",
        "read_calendar": "read:calendar",
        "transfer_funds": lambda args: f"spend:usd<={args['amount']}",
    },
    on_denied="throw",  # or "prompt" with on_prompt=...
)

# Pass the caller's mandate on the context:
server.call_tool("read_calendar", {}, {"mandate": mandate})

agent_authority.mcp.behalf_mcp_tools() returns the three discovery tools (request_mandate, present_mandate, check_authority).

Develop

python3 -m unittest discover -s tests   # 28 tests