getpaseo/paseo

Paseo is a local daemon that lets you run Claude Code, Codex, GitHub Copilot, OpenCode, and Pi side by side, controlled through a single interface that spans iOS, Android, desktop, web, and CLI. The daemon manages agent processes on your own machine; clients connect to it over WebSocket, TCP, a Tailscale tunnel, or an end-to-end encrypted relay for phone pairing. Nothing runs on Paseo's servers by design — no telemetry, no forced accounts.

The multi-agent angle is the real pitch. You can fire agents in parallel against different worktrees, stream their output, and send follow-up instructions mid-run. The CLI makes this scriptable: paseo run accepts a provider string like codex/gpt-5.5, a worktree flag, and a prompt, and you can target a remote daemon by passing a host flag. That last bit matters for anyone who wants to kick off long-running agent work on a beefy workstation from a laptop or phone without leaving a terminal session open.

The Skills feature is the most interesting design choice. Rather than building orchestration logic into Paseo itself, it exposes slash commands — /paseo-handoff, /paseo-advisor, /paseo-committee — that agents can call to delegate, consult, or convene a two-agent review panel. The README describes a concrete workflow: plan with Claude, then hand off implementation to Codex. That pattern is simple enough to be credible and specific enough to be useful, which is more than most multi-agent orchestration pitches manage.

The TypeScript SDK (@getpaseo/client) connects over WebSocket and exposes a straightforward create-wait-read pattern for building dashboards or issue integrations on top of the daemon. The monorepo structure is clear: a Node server for orchestration, an Expo app for mobile and web, Electron for desktop, and a separate Elixir relay service in its own repo.

Docker support is present and practical — a single docker run command starts the daemon and a self-hosted web UI on port 6767, with a persistent home volume for credentials and a workspace mount. The base image expects you to layer in whichever agent CLIs you need, which keeps the image lean but adds a setup step.

The AGPL-3.0 license is worth noting for anyone considering embedding this in a commercial product. Voice control is listed as a feature but the README says nothing about how it works under the hood — whether it's device-native speech-to-text or something else is undocumented here.