Security posture
Security posture for the Coven daemon: same-user local IPC trust, daemon-owned enforcement, no stored provider credentials, remote access guidance, and explicit destructive confirmations.
4 min read
Coven's daemon security model is local-first and deliberately narrow.
- The Rust daemon is the authority.
- The platform local IPC endpoint is same-user trust.
- Ordinary harness authentication stays with the provider CLI; experimental Runtime Authority has a separate explicit credential boundary.
- Destructive operations require explicit confirmation.
For the broader product safety framing, see Safety Model and Authentication and local access.
Daemon authority
Clients can ask. The daemon decides.
The daemon validates:
- Project roots.
- Working directories inside project roots.
- Harness adapter ids.
- Session ids.
- Session liveness before input or kill.
- Control-plane action ids.
- API route version compatibility.
Client-side validation is useful for better forms and friendlier errors, but it is not the trust boundary.
Same-user local IPC trust
The daemon does not use OAuth, JWTs, bearer tokens, API keys, or browser cookies. Access uses:
$COVEN_HOME/coven.sockon Unix-like hosts; or- an owner-only named pipe reported by health and
coven daemon statuson Windows.
Anyone who can connect to that endpoint can act as a local Coven client. Keep $COVEN_HOME and the endpoint private to the user who owns the daemon.
Optional loopback TCP
On Unix-like hosts, coven daemon serve --tcp <addr> can also bind a TCP listener. The daemon refuses non-loopback bind addresses and checks the Host and Origin headers to reduce browser risk. Each --allow-host <host> entry trusts one proxied hostname, and a browser Origin for it must be exactly https://<host[:port]>; plain-HTTP or different-port origins are rejected.
TCP does not prove that the caller owns the daemon. Over TCP, the daemon refuses owner-only operations with HTTP 403, including:
- automation mutations and automation reads that return prompts, run logs, receipts, or event checkpoints (
AUTHORITY_REQUIRED); - restricted session launches and the
launchPolicyor ClaudewriteScopelaunch fields (forbidden); - any request that carries a
contextAdmissionmember.
Health over TCP reports sessionLaunchPolicy: false, sessionClaudeWriteScope: false, and an empty sessionPolicyContracts list. Send owner-only requests over the local IPC endpoint.
Provider credentials
Ordinary harness sessions keep using the provider CLI's own login and auth stores. The experimental Runtime Authority scoped-read path is separate: it can use an owner-supplied Anthropic API credential held in private local state. That path does not reuse a harness login or silently obtain a key, and its hidden engineering controls are not a supported provider-onboarding workflow. Do not infer production qualification from the presence of the adapter or its credential store.
That boundary matters:
codexauth stays with Codex.- Claude Code auth stays with Claude Code.
- Gemini CLI auth stays with Gemini CLI.
- Local model configuration stays with the harness or provider runtime.
Coven records session metadata and output for replay and audit, so $COVEN_HOME is still sensitive even though provider credentials are not copied into the daemon.
Remote access
Coven does not bind TCP by default. If you need a remote daemon, put an explicit authenticated transport in front of the platform local IPC endpoint and treat that transport as the auth layer.
Reasonable owner-operated patterns:
- SSH local forwarding to a machine you control.
- Tailscale between your own devices.
- A private gateway that adds authentication before touching the socket.
Avoid exposing the daemon's local IPC endpoint directly to browsers, public networks, shared machines, or unauthenticated TCP bridges.
Experimental agent filesystem boundary
The experimental Agent filesystem crate has an opt-in NFSv3 export for engineering validation. It is not a Coven daemon route or a supported CLI workflow. Its example binds 127.0.0.1 only and uses an ephemeral port by default, but loopback NFS has no authentication: any local process that can reach the port can make unauthenticated requests to the export. Operating-system policy can still restrict a particular client's mounted-volume operations.
Do not treat the NFS spike as a sandbox or an access-control boundary. A mounted view does not stop a process from writing outside that view. A human-operated macOS Terminal pass confirmed loopback read/write behavior for one consented client, but the macOS privacy policy remains process-specific and other client processes may still be denied without explicit consent. Loopback exposure, access control, platform sandboxing, and recovery must be validated before a mount becomes a supported default.
Destructive confirmations
Destructive operations must remain explicit even when wrapped by another client:
| Operation | Confirmation posture |
|---|---|
coven sacrifice <id> --yes | Permanent delete requires --yes. |
| Process/cache relief commands | Require explicit confirmation flags. |
| OpenClaw patch flows | May generate patches, but v0 should not silently commit or push. |
Every local client must surface the same confirmations instead of hiding them behind a prettier button.
Related
Last updated on