Agience Origin

Identity and authority — the OIDC issuer.

Origin is the trust anchor peers verify against, and the service everything else asks who is this, and may they? It mints and verifies tokens, publishes the JWKS that peers check signatures against, and owns the passkey, OTP and account-setup flows.

Running it

An ordinary Python package — no container, and installing it pulls its own pinned dependencies.

pip install agience-origin
KEYS_DIR=/path/to/keys python -m uvicorn origin.main:app --host 127.0.0.1 --port 8080

It applies its own migrations at startup, so the first boot creates the database. Requires Python 3.11 or newer.

Bind loopback and put a reverse proxy in front of anything public — a service on the public interface answers past whatever header and path rules the proxy applies.

KEYS_DIR must already contain origin.private.pem, origin.public.pem, encryption.key and inbound_nonce.secret. Every key loader raises rather than inventing a key it did not write, which is correct for an authority and makes an empty directory a hard stop. Key material comes from a platform installer, a KMS, or a one-shot key-init step; .env.example documents the full set.

Letting a peer verify Origin

Agience peers read their trust inline, from trust_anchors in the authority.manifest.json of their own keyset. Publishing /.well-known/jwks.json is therefore not what makes a peer able to verify an Origin-signed token: until Origin's public JWK is physically present under trust_anchors.origin, two healthy and mutually reachable services answer 401 for every user token, with nothing to log — from the peer's side there is no mismatch, there is simply no such issuer.

A peer's own key init writes only its own anchor, since asserting a public key for a service whose private key is elsewhere is a trust statement rather than a convenience. Origin emits its half:

origin-emit-anchor                                    # the mergeable fragment
origin-emit-anchor --format anchor --uri https://origin.example.com
origin-emit-anchor --format jwks --keys-dir /path/to/keys
python -m origin.scripts.emit_trust_anchor            # straight from a checkout

--format chooses fragment (a mergeable {"trust_anchors": {"origin": …}}, the default), anchor (the value alone, for placing at trust_anchors.origin), or jwks. --keys-dir defaults to $KEYS_DIR and --uri to config.AUTHORITY_ISSUER.

It reads KEYS_DIR/origin.public.pem and produces the JWK through the same get_jwk_public() that serves /.well-known/jwks.json, so what you place is byte-identical to what Origin publishes, kid included. It writes nothing: placement is the operator's decision, and a command that installed trust in itself on a peer would send a trust anchor the one direction it must never travel.

To place it, merge into the peer's manifest:

origin-emit-anchor --format anchor > /tmp/origin-anchor.json
python - <<'EOF'
import json, pathlib
m = pathlib.Path("/path/to/peer/keys/authority.manifest.json")
doc = json.loads(m.read_text())
doc.setdefault("trust_anchors", {})["origin"] = json.load(open("/tmp/origin-anchor.json"))
m.write_text(json.dumps(doc, indent=2) + "\n")
EOF

Then restart the peer. A wrong kid or a re-encoded modulus fails as the same silent 401, which is why the JWK is emitted rather than transcribed.

Configuration

.env.example is the template — copy it to .env. It states the in-code default for every value and marks where an unset variable is itself a decision: KEYS_DIR unset means the process does not boot, and ORIGIN_ALLOWED_ORIGINS unset derives the CORS allow-list from the issuer, ORIGIN_URI and the facet bases.

Layout

web/ lives inside the package, beside the module that serves it: main.py resolves it as Path(__file__).resolve().parent / "web", which gives one answer in both places the code runs — a checkout and an installed distribution. A path that climbs out of the package resolves against the repository layout, which only a checkout has.

Star history

Security issues: email connect@agience.ai rather than opening a public issue.

Dual-licensed — see LICENSE, COMMERCIAL_LICENSE.md, NOTICE and CLA.md.

Declaration of generative AI use

The author used Anthropic's Claude Opus (versions 4.8, 5, and 5.5) in the preparation of this work. Its contribution was to write code, and to generate and validate content. The ideas, the construction and the claims are the author's. No other generative AI tool was used. The author reviewed and edited all output and takes full responsibility for the content of this publication.