Agensis Agent

Open-source Relay host for Agensis workspace agents — connect Claude Code, Codex, Hermes, Grok, and other coding CLIs on a machine that should execute work.

In the product, agent run modes are:

The daemon receives Relay jobs over an authenticated WebSocket, runs the selected coding CLI in the configured working directory, and streams results back to Agensis (web and desktop). The website, backend, and desktop app live in a separate repository; this repo is only the host-side Relay client.

Amp Orb agents use the same connection: the daemon invokes the signed-in local Amp CLI to create or continue an Amp-managed orb thread, then streams that thread back into its Agensis conversation. Amp authentication and billing stay on the host. See the CLI package README for setup and failure behavior.

npm install -g @agensis/agensis-agent
agensis --help

Copy the connection command from an agent profile in Agensis, then run it from the repository the agent should work in:

agensis connect --url https://agensis.io --token aga_... --workspace ... --agent ...

After that first successful connection has saved a local profile, keep the agent running when the terminal or desktop app exits:

agensis service install --profile default
agensis service status --profile default
agensis service logs --profile default

service install creates a per-user macOS LaunchAgent (RunAtLoad + KeepAlive) or Linux systemd user unit (Restart=always). Install the CLI globally first so the recorded executable path remains stable. The service definition contains only the profile name plus executable and log paths; the connection token, workspace, agent id, and working directory remain in the existing mode-0600 profile under ~/.agensis/daemon-profiles.

Use agensis service logs --profile default --follow only when you want a foreground live tail, and agensis service uninstall --profile default to disable and remove that exact profile's service. Windows service installation is not implemented; the command fails without changing the machine.

Lean execution is enabled by default. Claude starts in safe mode, Codex skips user config, project instructions, memories, plugins, hooks, and skill search, and both receive only the Agensis MCP configuration. Use --full-cli-context only when the connected agent intentionally needs local Claude/Codex customizations.

Local Claude memory synchronization is off by default. --sync-memory opts in to uploading the selected project's Claude memory files, including file names, contents, sizes, and the absolute memory-root path, to the connected Agensis workspace. Each file is read-only, restricted to the derived or configured memory root, and capped at 256 KiB.

Skill content synchronization is on by default. The daemon already advertises the NAMES of the skills it finds under ~/.claude/skills, ~/.codex/skills and <project>/.claude/skills; it also uploads each one's SKILL.md so agents elsewhere in the workspace can read a skill while this machine is offline. Reads are confined to those three roots, capped at 64 KiB per skill with any truncation marked in the text, and a file that cannot be read is simply reported as having no body. --no-sync-skills (or AGENSIS_SYNC_SKILLS=0) keeps the bodies local; only the names are then advertised. Opting out stops further uploads — it does not delete text already mirrored into the workspace.

• packages/agensis-cli — readable daemon source
• packages/agensis-agent — npm package and single-file bundle build
• tests — daemon integration and unit tests

npm install
npm run verify

The release package is @agensis/agensis-agent. Its npm name and wire protocol are stable across the repository split.

The daemon executes coding CLIs with access to the selected working directory. Job payloads and CLI results travel through the Agensis backend. Review permission mode and host-folder settings before connecting it. Keep aga_... connection tokens out of logs and rotate any token that is exposed.

Security reports should be submitted through GitHub's private vulnerability reporting for this repository.

MIT