ACP UIvsherdr
Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.
ACP UI
Jun Han · Agent harness
OSS
Panel
6.33 spec wins
- Reliability
- 5.8
- Usefulness
- 5.7
- Cost
- 8.3
- Longevity
- 5.2
“There is an Android build, so you can now approve a file write from the bus and regret it before your stop.”
herdr
herdr · Agent harness
OSS
Panel
6.32 spec wins
- Reliability
- 6.2
- Usefulness
- 5.7
- Cost
- 7.5
- Longevity
- 6.0
“A tmux for agents whose detach key is ctrl+b q, so twenty years of muscle memory finally counts as onboarding.”
Spec by spec
| Spec | ACP UI | herdr |
|---|---|---|
| Architecture | ||
| Category | Agent harness | Agent harness |
| Runs | local | local |
| Platforms | macos, linux, windows, web | macos, linux, windows |
| Context window | not documented | not documented |
| Protocols | ||
| MCP client | No | No |
| MCP server | No | No |
| Capabilities | ||
| Runs terminal commands | YesAs the ACP client, ACP UI executes the terminal commands the agent requests and shows their output; the desktop build only, since the browser cannot spawn processes. | Yes |
| Multi-file edits | YesFile reads and writes are performed by ACP UI on the agent's behalf against the host workspace, and only on desktop. | No |
| Git operations | No | No |
| Browser control | No | No |
| Sandboxed execution | No | No |
| Multi-agent orchestration | No | Yes |
| Headless / CI mode | No | No |
| Models | ||
| Backbone | Claude Code, Codex, GitHub Copilot, Gemini CLI, Qwen Code, OpenCode, OpenClaw, Kiro CLI, any ACP-compatible agent | Claude Code, Codex, Cursor, OpenCode, Grok |
| Bring your own model | Yes | No |
| Local models | No | No |
| Cost | ||
| Pricing model | byok | byok |
| Starts at | $0/mo | $0/mo |
| Free tier | Yes | Yes |
| Bring your own key | Yes | No |
| Openness | ||
| Open source | Yes | Yes |
| License | MIT | Apache-2.0 |
| GitHub stars | 491 | 41,859 |
Which one would each critic pick
| Critic | ACP UI | herdr | Pick |
|---|---|---|---|
| El Juez | — | — | not enough reviews |
| El Amigo | 6.5 | 7.0 | herdr — Pick herdr if your Claude Code, Codex, Cursor or OpenCode sessions die when you close the laptop; pick Claude Squad if what you actually need is a worktree per task. |
| El Crítico | 6.0 | 6.0 | no preference |
| El Profesor | 6.5 | 6.8 | herdr — One Rust binary with a socket API, per-pane status tracking, and a wait primitive that returns only when another agent is genuinely blocked; no benchmark, and nothing to benchmark. |
| La Inversora | 5.5 | 4.8 | ACP UI — One author, 461 stars, no company and no revenue line: the asset is a protocol client, and protocol clients get absorbed by whoever owns the protocol. |
| La Jefa | 6.0 | 5.5 | ACP UI — Nothing per seat across sixty desks and nothing to govern: no console, no single sign-on, no audit record, and nothing that runs in a pipeline. |
| El Hacker | 7.0 | 8.0 | herdr — Apache-2.0, one binary via brew or mise, a socket API I can script from anything, no MCP and no model settings because it owns terminals, not agents. |
Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.