legant-sdk 0.1.1

Offline verifier + authorizer for Legant delegation tokens (RFC 8693 sub/act), with Tier-B revocation-feed support.

legant-sdk (Python)

Offline verifier + authorizer for Legant delegation tokens. Its only dependency is cryptography (Python ≥ 3.9).

from legant_sdk import fetch_jwks, Verifier, Action, fetch_revocation_feed

issuer = "https://auth.example.com"
keys = fetch_jwks(f"{issuer}/.well-known/jwks.json")

# Tier B (optional): reject revoked tokens offline, refreshed in the background.
feed = fetch_revocation_feed(f"{issuer}/.well-known/revoked", issuer, keys)
feed.start_polling(10.0, on_error=print)

verifier = Verifier(issuer, "https://my-api.example/", keys, feed=feed)

# Per request:
claims = verifier.verify(bearer_token)  # raises VerifyError / RevokedError on failure
claims.authorize(Action(scope="expenses:submit", amount=120, category="travel"))  # raises AuthorizeError on 403
print(claims.provenance())  # "user:alice -> agent:assistant"

verify raises VerifyError (or RevokedError when the token is in the feed); authorize raises AuthorizeError when a scope or constraint is denied. Catch them to return 401 / 403.

Guard an agent's tools — any framework

AgentGuard wraps any tool callable so every invocation is authorized against the agent's delegation token — offline, no callback. The wrapped function is a plain callable, so it drops into LangChain, CrewAI, LlamaIndex, AutoGen, or your own loop unchanged. A prompt-injected or buggy agent cannot exceed the scoped, revocable slice the token carries.

from legant_sdk import Verifier, AgentGuard

verifier = Verifier(issuer, "https://my-api.example/", keys, feed=feed)
guard = AgentGuard(verifier, token=agent_delegation_token)  # token may be a callable, to refresh

@guard.tool("expenses:submit", amount_arg="amount", category_arg="category")
def submit_expense(amount: float, category: str) -> str:
    ...  # only runs if the token permits this scope, amount, and category — else AuthorizeError

# LangChain: from langchain_core.tools import tool;  lc_tool = tool(submit_expense)
# CrewAI:    @tool("Submit expense")  def submit_expense(...): ...  then wrap with @guard.tool(...)

Or check inline without the decorator: guard.authorize("scope", amount=…) (raises) or guard.allowed("scope", amount=…) (returns a bool).

Test

python3 -m unittest discover -s tests    # runs the shared conformance vectors (see ../conformance)