Charon
Self-hostable control plane for managing the full lifecycle of non-human identities (AI agents), with short-lived credential issuance, attestation, and an MCP authorization gateway for per-tool access control.
README
Charon — NHI Lifecycle Engine for AI Agents
A self-hostable control plane that manages the full lifecycle of non-human identities (AI agents, workloads, automated processes): attestation, short-lived credential issuance, gated lifecycle transitions, per-tool authorization for the Model Context Protocol (MCP), delegation with provenance, proof-of-possession, automated decommissioning, and a tamper-evident audit trail — built on SPIFFE naming, JWT-SVIDs, and established OAuth RFCs.
Human identity management is a mature, solved market. Non-human identity — the API keys, service accounts, and AI agents that act without a person in the loop — is not. Enterprise NHI products are closed and priced for enterprises, and the open-source layer (SPIFFE/SPIRE, Vault, Teleport) provides excellent identity plumbing but leaves the lifecycle and governance layer to you. Charon is that layer, focused on the AI-agent frontier, and it builds on the plumbing rather than reinventing it.
Built on real standards: SPIFFE naming and JWT-SVIDs, RFC 8693 (OAuth 2.0 Token Exchange) for multi-hop delegation, RFC 9449 (DPoP) for proof-of-possession, and RFC 7638 for key thumbprints. 84 tests passing, including an adversarial suite; four runnable demos and a live dashboard.
Capabilities
Attested issuance, not bearer secrets. A workload must prove itself before it
gets a credential. JoinTokenAttestor issues single-use, expiring tokens;
K8sServiceAccountAttestor verifies a projected Kubernetes service-account JWT
against the cluster's keys and extracts namespace/service-account selectors.
Attestation carries a TTL, so identities must periodically re-prove themselves,
and issuance is refused without a fresh attestation on record. Selectors can be
bound so a credential for one workload identity can't be obtained by another.
Short-lived, scoped credentials. Every credential is a SPIFFE JWT-SVID
(spiffe://<trust-domain>/agent/<id>) with an explicit scope claim and a short
TTL — no static, long-lived API keys. Rotation issues a fresh credential and
revokes the prior one; the signing key can be rotated while older tokens remain
verifiable until they expire; revoked credentials are rejected at verification.
A real lifecycle. Identities move through
PROVISIONED → ACTIVE → IDLE → REVOKED → DECOMMISSIONED via a state machine that
rejects illegal transitions and gates activation on attestation.
Per-tool authorization for MCP. The gateway fronts MCP servers and authorizes
every individual tools/call against the credential's scopes and argument-level
constraints (path confinement, amount caps) — closing MCP's structural
all-or-nothing access gap. tools/list advertises only the tools a credential
can actually invoke. Policy runs in a dependency-free embedded engine by default,
or in OPA via policies/authz.rego.
Delegation with provenance. Using RFC 8693 token exchange, one agent can act
on behalf of another while preserving the originating human as sub and nesting
each actor in the act claim. Authority only ever narrows down a chain. Any
action — even three hops deep — can be traced back to the human who authorized it.
Proof-of-possession (DPoP). Credentials can be bound to a client key via a
cnf.jkt claim; the gateway then requires a fresh, method- and URL-bound,
single-use DPoP proof on every call. A stolen token is useless without the key,
and a captured proof cannot be replayed.
Automated cleanup. The reaper moves inactive agents to IDLE, decommissions
long-idle ones, revokes-and-decommissions orphaned agents whose owner has
departed, and flags privilege drift (granted scopes never exercised).
Tamper-evident auditing. Every registration, attestation, transition, issuance, rotation, revocation, and authorization decision is written to a hash-chained audit log that detects any modification of historical records.
Dashboard. A single page showing lifecycle counts, the identity inventory, the delegation graph, a reaper preview, and the live audit feed with chain-integrity status.
Architecture
flowchart LR
H([human owner]) -->|registers / authorizes| CP
subgraph CP[Charon control plane]
REG[Registry + lifecycle state machine]
ATT[Attestation: join-token / k8s SA]
CA[Credential Authority - JWT-SVIDs + DPoP cnf]
DEL[Delegation - RFC 8693 act-chains]
REAP[Reaper - idle / orphan / drift]
AUD[(Hash-chained audit log)]
REG --- ATT --- CA --- DEL --- REAP
REG -.writes.-> AUD
end
CA -->|short-lived JWT-SVID| AG([AI agent / workload])
AG -->|JWT-SVID + DPoP proof| GW
subgraph GWX[ ]
GW[MCP Authorization Gateway] -->|per-tool decision| POL[Policy engine - embedded / OPA]
end
GW -->|allowed calls only| MCP[(MCP servers: filesystem / payments / email)]
GW -.authz decisions.-> AUD
DASH[Dashboard at /] -.reads.-> CP
SPIRE[[SPIRE - production CA swap]] -.verifies SVIDs.-> GW
The security-critical logic depends only on PyJWT and cryptography and is
fully unit-tested in isolation. The web framework and database sit at the edges as
thin adapters, so they can be swapped (SQLite → Postgres, or the self-signed CA →
SPIRE) without touching the security core.
charon/
spiffe.py SPIFFE ID construction / parsing
lifecycle.py state machine + gated transitions (pure, no I/O)
models.py domain dataclasses (persistence-independent)
audit.py hash-chained, tamper-evident audit log
ca.py Ed25519 signing authority + trust bundle (persistable, rotatable)
credentials.py Credential Authority: issue / verify / rotate / revoke JWT-SVIDs
attestation.py pluggable attestors: join-token, k8s SA JWT, dev
delegation.py RFC 8693 token exchange + act-claim chains + provenance
dpop.py DPoP proof-of-possession: RFC 9449 + RFC 7638 thumbprints
spire.py SPIRE integration adapter (py-spiffe) — production CA swap
reaper.py idle / orphan / drift detection + auto-decommission
policy.py authorization engines: embedded (default) + OPA-backed
repository.py Repository interface + stdlib-sqlite3 implementation (Postgres-ready)
service.py Registry: orchestrates lifecycle + attestation + delegation + audit
mcp/servers.py example MCP servers: filesystem, payments, email
mcp/gateway.py MCP authorization gateway: per-tool authz enforcement point
mcp/stdio_server.py real MCP-SDK entrypoint wrapping the gateway
api/main.py FastAPI HTTP layer (thin adapter over service.py)
api/dashboard.py single-page dashboard: inventory + lifecycle + delegation graph
policies/authz.rego Rego policy mirroring the embedded engine (for OPA)
Quick start
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Run the tests:
python -m unittest discover -s tests # or: pytest
# Self-contained walkthroughs (no DB or network needed):
python demo.py # lifecycle: attest, issue, rotate, revoke, audit
python demo_gateway.py # per-tool authorization (read-only agent blocked from payments)
python demo_delegation.py # delegation provenance (human -> A -> B -> C) + reaper sweep
python demo_hardening.py # DPoP defeats token theft + replay
# Run the HTTP control-plane API + dashboard:
uvicorn charon.api.main:app --reload
# then open http://127.0.0.1:8000/ (dashboard)
# and http://127.0.0.1:8000/docs (interactive API)
State persists in charon.db and the signing key in charon_signing_key.pem
(override paths with CHARON_DB / CHARON_SIGNING_KEY), so issued credentials
survive a restart. The core (tests and all four demos) runs with only PyJWT and
cryptography; fastapi/uvicorn are needed only for the HTTP layer.
Optional integrations: install mcp and run python -m charon.mcp.stdio_server
to expose the gateway as a real MCP server; run opa run --server policies/ and
construct MCPGateway(..., policy=OpaPolicyEngine()) to use OPA; run SPIRE and pass
charon.spire.SpireJwtVerifier as the gateway's verifier to make SPIRE the issuer.
Highlighted demos
- Least privilege (
demo_gateway.py): an agent scoped tofs:readreads files under/databut is deniedpayments.charge(missing scope), denied/etc/shadow(path escape), and never even sees the tools it can't call. - Provenance (
demo_delegation.py): ahuman -> A -> B -> Cchain, traced from the final agent's credential all the way back to the originating human. - Proof-of-possession (
demo_hardening.py): the same stolen token is allowed for the key-holder and denied for everyone else (no proof / wrong key / replay).
HTTP API
| Method | Path | Purpose |
|---|---|---|
| POST | /agents |
register an identity |
| GET | /agents / /agents/{id} |
inventory |
| POST | /agents/{id}/attest |
attest a workload |
| POST | /agents/{id}/transition |
gated lifecycle transition |
| POST | /agents/{id}/credentials |
issue a JWT-SVID (optionally DPoP-bound) |
| POST | /agents/{id}/credentials/rotate |
rotate |
| POST | /agents/{id}/credentials/revoke |
revoke |
| POST | /credentials/verify |
verify a token |
| POST | /delegation/begin |
start a chain (agent on behalf of a human) |
| POST | /delegation/exchange |
RFC 8693 token exchange (further hops) |
| POST | /delegation/trace |
reconstruct a credential's provenance path |
| GET | /delegations |
delegation edges (for the graph) |
| POST | /reaper/run |
run the reaper (apply or dry-run) |
| POST | /mcp/tools |
list tools the credential may call (via the gateway) |
| POST | /mcp/call |
authorize + forward a single tool call |
| GET | / |
dashboard |
| GET | /.well-known/charon/trust-bundle |
public keys for verifiers |
| GET | /audit |
audit trail + integrity status |
Security model
Charon assumes the control plane is trusted and treats everything outside it —
agents, networks, MCP servers — as potentially hostile. Nothing crosses the trust
boundary without attestation plus a valid, unrevoked, unexpired credential whose
scope permits the action. It deliberately does not address risks that live at
the model and tool layer rather than the identity layer (tool-description
poisoning, prompt injection, context over-sharing) or runtime behavioral
monitoring of a compromised agent — those pair with an identity engine rather than
belonging inside it. See docs/THREAT_MODEL.md for the full analysis, including a
mapping to the OWASP MCP Top 10.
Documentation
docs/DESIGN.md— design and landscape analysisdocs/THREAT_MODEL.md— assets, adversaries, and the OWASP MCP Top 10 / NIST mappingdocs/BLOG.md— where Charon fits against the emerging agent-identity standards
License
MIT — see the LICENSE file. The OWASP MCP Top 10 material referenced in
docs/THREAT_MODEL.md is CC BY-NC-SA 4.0 and is only cited, not reproduced.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.