the honest comparison
agentproto vs Paseo, Claude Squad, Conductor, Agent Farm
The honest framing first. Paseo is the one real shape-competitor here: also a local daemon, also multi-provider, with polished desktop and mobile clients and voice control. Where we differ is the center of gravity — Paseo is built around a human steering agents from any device; agentproto is built around supervision: per-turn policy gates, commits staged behind a human ack, role-gated nested orchestration, and an open spec family underneath. Apache-2.0 vs AGPL matters too if you want to embed the daemon in your own product.
The other three are not really competitors. Claude Squad, Conductor and Agent Farm are interactive tools a human sits inside of to run several coding-agent sessions at once — and they are good at that. agentproto is a daemon with a programmatic lifecycle you call from code, from another agent, or from cron. The supervision primitives it adds — durable policy gates, multiplexed fan-in monitoring, nested orchestration — do not exist in the other three.
| Dimension | agentproto | Paseo | Claude Squad | Conductor | Agent Farm |
|---|---|---|---|---|---|
| Agents supported | Claude Code, Codex, Hermes, opencode, Mastra (Code + Agent), claude-sdk (Anthropic/Moonshot/OpenRouter), OpenClaw, browser-as-agent — uniform AIP-45 manifests | Claude Code, Codex, Copilot, OpenCode, Pi — plus custom providers (custom binaries, ACP agents, Anthropic-compatible endpoints) | Claude Code, Codex, Gemini, Aider, other local agents | Claude Code, Codex, Cursor | Claude Code only (by design) |
| Interface | Daemon with CLI + HTTP + MCP surfaces; scriptable from code, another agent, or cron — no terminal required | Daemon + desktop, iOS/Android, web, and CLI clients; voice control — built for a human steering from any device | Interactive terminal TUI (human sits in it) | Mac desktop app (human sits in it) | Python script / batch CLI |
| Isolation | Daemon-tracked sessions; git-worktree isolation; sandbox provider family landing (local shipped, e2b microVM in progress) | Git worktree per run (--worktree); official Docker image; remote daemon (--host) | tmux session per agent + git worktree per branch | Isolated workspace per agent | tmux panes + lock-based file coordination |
| Programmatic API | Yes — spawn/prompt/monitor/kill/export over MCP or HTTP; fan-in monitor; usage/cost introspection | Partial — CLI verbs (run/ls/attach/send) + WebSocket API; primarily driven from its own clients | No — TUI app | No — desktop app | Partial — configurable script, not a long-lived API |
| Durable supervision / policy gates | Yes — shell or LLM-judge gate on turn-end, event bus, commit gated on human ack; survives client disconnect | Partial — remote approval of agents' own permission prompts, plus a Ralph-loop skill with optional verifier; no daemon-side turn-end gates or commit staging | No — human watches the TUI | No — human reviews in the app | Partial — workload watchdog (auto-restart), not per-turn gates |
| Nested orchestration | Yes — scoped sub-gateways, executor/supervisor role gating, depth/children caps | Partial — handoff/committee/advisor skills between agents; no role gating or depth caps | No | No | Partial — fans one workload across N agents |
| MCP surface | Yes — the daemon is an MCP server (~90 tools); agents can mount external MCP servers | Yes — daemon MCP server (opt-in config); drivable from Claude Desktop/Code or any MCP client | No | No | No |
| Your own tools, shared with every agent | Yes — author a tool once (TOOL contract + DRIVER implementation), the daemon serves it over MCP to every agent; external MCP servers can be imported and handed to any agent at spawn | No — the daemon injects only its own control tools (agents, terminals, schedules, permissions) into agents | No | No | No |
| Scheduled / recurring runs | Yes — workflows (staged concurrency, schema-validated outputs, cost ceilings) + durable cron on the daemon | Yes — recurring schedules on an agent, with run history, manageable over MCP | No | No | Partial — batch runs, not scheduled |
| License / platform | Apache-2.0 — cross-platform daemon + CLI (macOS + Linux verified) | AGPL-3.0 — cross-platform, desktop + mobile apps; solo-maintained | AGPL-3.0 — cross-platform TUI (requires tmux + gh) | Closed source — macOS only | MIT — cross-platform (Python 3.13+, tmux, Claude Code) |
Facts sourced from each project's own page, 2026-07-07 (Paseo: 2026-07-13). Corrections welcome — file an issue.
When to use which
- Paseo — you want to steer agents from your phone or by voice, across desktop and mobile, with a polished cross-device experience.
- Claude Squad — you want a terminal cockpit for a handful of parallel sessions with git worktrees, reviewing each by hand.
- Conductor — same workflow, polished Mac app, across Claude Code, Codex and Cursor.
- Agent Farm — you have a lint/type-fix queue and want to throw ~50 Claude Code instances at it and get an HTML report.
- agentproto — you want to drive coding agents from code or cron, with gates that survive a client disconnect, fan-in monitoring over N sessions, and children that can be granted their own scoped orchestration gateway.
They can compose
agentproto does not replace the cockpits. A tool like Claude Squad or Conductor could mount the agentproto MCP server and get durable policy gates and multiplexed monitoring for free, instead of hand-rolling a tmux + Redis watchdog around its panes. Interactive cockpit on top, supervised daemon underneath, any adapter either side wants to drive.