- Python 94.4%
- HTML 3%
- PowerShell 1.1%
- Shell 0.9%
- JavaScript 0.6%
|
Some checks failed
tests / test (macos-latest, 3.11) (push) Has been cancelled
tests / test (macos-latest, 3.14) (push) Has been cancelled
tests / test (ubuntu-latest, 3.11) (push) Has been cancelled
tests / test (ubuntu-latest, 3.14) (push) Has been cancelled
tests / test (windows-latest, 3.11) (push) Has been cancelled
tests / test (windows-latest, 3.14) (push) Has been cancelled
tests / dependency-audit (push) Has been cancelled
|
||
|---|---|---|
| .github | ||
| docs | ||
| requirements | ||
| scripts | ||
| src/agent_authd | ||
| tests | ||
| .gitignore | ||
| CHANGELOG.md | ||
| CONTRIBUTING.md | ||
| Install agent-authd.command | ||
| Install-agent-authd.cmd | ||
| LICENSE | ||
| MANIFEST.in | ||
| pyproject.toml | ||
| README.md | ||
| SECURITY.md | ||
| Start agent-authd.command | ||
| uv.lock | ||
agent-authd
Give your agents access without handing every agent every secret.
Store API keys, passwords and SSH keys on your computer. Choose which agents can use them, connect a server, and see what happened in the activity log.
0.9.0b1 · Local beta · Python 3.11+ · MIT
Install
Download/extract a source archive or clone the repository. Install Python 3.11+ first. From the extracted folder:
macOS / Linux
sh scripts/install.sh
Windows — PowerShell
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\install.ps1
Windows also has a double-click Install-agent-authd.cmd entry point; macOS has
Install agent-authd.command. The installer builds a private runtime, verifies
its dependencies, sets up login startup and opens the dashboard. It preserves
existing credentials when upgrading. Browser automation is optional; ordinary
credentials and SSH do not require Chromium.
The default dashboard is http://127.0.0.1:8765, on your computer only. Open the
installed launcher or run authctl open to sign in. Use authctl status or
authctl doctor when something needs attention. See installation details
for paths, prerequisites, upgrades and uninstalling.
| Platform | Implementation | Verification |
|---|---|---|
| macOS | Keychain, Unix socket, login service | Local integration and installer checks |
| Linux desktop | Secret Service, Unix socket, systemd user | Native CI configured; desktop keyring/login acceptance remains required |
| Windows 11 x64 | User-scoped DPAPI, private named pipe, login scheduled task | Implemented; native CI and desktop acceptance remain required |
Windows is native; WSL is not required. This is not yet a signed, bundled-Python installer. Do not treat a configured CI workflow as a passed Windows test run.
First credential
- Click Add credential and enter a name and value.
- Choose one agent, a restricted group, All agents, or Private.
- Under Connect an agent, select the identity and copy its generated settings into your MCP client. Reopen that client.
- Ask it to call
auth_credentials.
Shared credentials can be provisioned into applications through a temporary private file. Values are not returned in MCP responses. The receiving local program can read the file, so provisioning deliberately grants access to the secret. Copies installed elsewhere cannot be recalled by this broker.
All agents includes future registered agents. Tasks using the same client token share the same identity. Create separate agents to give them different access.
Connect an SSH server
- Click Add SSH server.
- Enter a host, port, username and private key—or choose an existing key record.
- Choose an agent and whether each command needs your approval.
- Click Check server identity and compare the fingerprint with your server's trusted details. Confirm it, then click Test & connect.
The connection is tested before saving. New SSH keys stay private: permission to run commands does not automatically let the agent export the key.
The agent discovers the connection with auth_services and uses ssh_execute.
The default profile is shell. Approval, when enabled, covers the exact command.
An owner can also configure exact named commands in advanced profiles.
SSH supports encrypted private keys, pinned host keys, bounded output and per-command deadlines. This release runs non-interactive commands with a maximum 45-second deadline. It does not include SFTP, interactive terminals, forwarding or a job cancellation endpoint. A timeout closes the connection but cannot prove that a detached remote process stopped. SSH guide.
What you control
| Access | What it permits |
|---|---|
| Credential sharing | Export a credential to a short-lived private file for application setup |
| Connection permissions | Use a credential through brokered SSH or API requests |
| Owner approval | Approve an exact request before execution |
| Agent reset/removal | Revoke a client token or remove that agent's access |
Use Reset connection after a client token is lost. Reopen the agent client to load the replacement. Remove agent removes its identity and group memberships. API profiles and sharing grants remain separate, shown beside the relevant records.
Other connection types include fixed Basic/Bearer HTTPS GET operations, OVHcloud signed requests, Google OAuth and optional managed browser sessions.
Activity and recovery
The audit log records the authenticated actor, time, broker machine, observed source identity, action, target and result. It does not record credential values. Local socket/pipe calls have no source IP. Later actions using exported copies require the destination application's logs. Audit details.
Encrypted recovery export/import is available from the trusted terminal while the broker is stopped. The passphrase is entered privately; no plaintext backup file is created. Restore into a new empty state directory. See recovery.
Security boundary
This is a broker for a trusted local owner, not a sandbox for hostile code running as your OS user. Such code may read local files or access your credential store. It does not run while your computer is asleep, and startup follows your user login. Remote hosting and multi-user server deployment are not supported.
Read the security model and security reporting policy before connecting important accounts. Treat API, website and SSH output as untrusted.
Documentation and development
- Install, upgrade and troubleshoot
- SSH connections · Credential sharing · Custom records
- OVHcloud · Google · Browser sessions · Fixed APIs
- Audit · Recovery · Architecture
- Verification · Release gates · Changelog
For development:
uv sync --frozen --extra all
uv run python -m playwright install chromium
uv run python -m unittest discover -s tests -v
uv run authctl demo
Tests use synthetic credentials and local fixtures. The demo needs no real account. Contributions are welcome; see CONTRIBUTING.md. Licensed under MIT.
