protocol agentproto.shcli cli.agentproto.shpanel /panel

runs them · watches them · checks their work

Let your coding agents work without you watching.

agentproto runs Claude Code, Codex, and any other coding agent in the background, shows them all in one place, and checks each one's work — with your tests or a reviewer model — before anything is committed. No more babysitting terminal windows.

npm i -g @agentproto/cli
Get startedGitHub

9 adapters · Apache-2.0 · npm @agentproto/cli

agentproto · gateway up · http://127.0.0.1:18790
SESSIONADAPTERMODELSTATUS
  • apiclaude-codesonnet-5running
  • └─ testscodexgpt-5.5-codexrunning
  • reviewhermesz-ai/glm-5.2turn-end
  • lint+testpolicygatepassed
  • commitstagedawaiting ackack?

any agent · any vendor · any model

Mix vendors. Route by cost.

Claude for the plan, Codex for the refactor, a cheap open model for the mechanical work — same commands, same list, same checks. When a better or cheaper model ships, it's an adapter away, not a migration. No vendor decides what your fleet looks like.

the problem

Coding agents made you the babysitter.

without agentproto

  • One terminal per agent — you rotate between them all day.
  • You read every diff yourself. Or nobody does.
  • Close the laptop and everything dies mid-task.

with agentproto

  • Agents run in the background — one list shows them all.
  • Your tests or a reviewer model check every step.
  • Commits wait for a green check and your OK. Work continues without you.

quickstart

Three commands to a running, watchable agent

Install, start, delegate. The session shows up in one list from the first second. Attach a check to it and walk away — the work continues, watched and gated, without you.

See how it works inside →
zsh — 30 seconds to a supervised agent
$ npm i -g @agentproto/cli
+ @agentproto/cli — 9 adapters detected
$ agentproto serve
● gateway up · http://127.0.0.1:18790 · panel ready
$ agentproto sessions start claude-code --prompt "fix the flaky test"
● s_7f2k · claude-code · running — sessions --watch

your gateway, on any screen

One board for the whole fleet — terminal or web

The daemon is one gateway with three surfaces: CLI, HTTP, and MCP. Watch the session board in your terminal, open it as a live panel inside your MCP client, or expose the gateway through a named tunnel and check on your agents — and ack that staged commit — from another machine. The fleet is wherever you are.

terminal

agentproto sessions --watch — the live board where you already live.

mcp panel

Session and fleet views render as MCP App panels inside clients that support them.

tunnel

A named tunnel gives the gateway a stable HTTPS address — reach your fleet from anywhere you trust.

orchestration

One agent can manage the others

agentproto plugs into the agent you already use, as tools. Ask Claude Code to spin up a Codex session for a fix — or three cheap open-model sessions for the grunt work — and the whole tree is real, visible sessions: who spawned whom, what each one costs, which gate it's behind. That's a team, run from your prompt.

  • Roles keep delegation sane

    Executors do the work and can't spawn; supervisors delegate, with depth and children caps enforced by the daemon. No runaway agent pyramids.

  • One call waits on the whole fan-out

    sessions wait blocks until any of N sessions finishes a step — no polling loops, in the CLI or over HTTP/MCP.

  • Workflows and cron, on the daemon

    Ordered stages of concurrent steps, output validated against a schema, per-run cost ceilings — scheduled nightly if you want. It runs whether you're there or not.

Full guide: MCP inside coding CLIs →
agentproto sessions tree
● s_2del supervisor · claude-code · "ship the release"
├─ ● s_9k2f executor · codex · "fix flaky test" · gate ✓
├─ ● s_x1a4 executor · hermes/glm · "lint sweep" · gate ✓
└─ ● s_p8c3 executor · hermes/glm · "changelog" · running
fan-in: waiting on 1 of 3
cron: nightly-audit in 6h 12m
# Claude Code (native HTTP transport)
claude mcp add --transport http agentproto \
  http://127.0.0.1:18790/mcp

# Codex, Cursor, Claude Desktop (stdio)
agentproto mcp-bridge

the part nobody else does

A gate between your agents and your main branch

Every time an agent finishes a step, a check you choose runs: a shell command (lint, tests) or a stronger model acting as reviewer. The commit waits behind that check for your explicit OK — from any device, or none. Close your laptop; the checks keep running.

Full features breakdown →
policy · session review · turn-end
turn-end review · gate lint+test attached
✓ pnpm lint — clean
✓ pnpm test — 214 passed
✓ judge: “diff matches the brief”
commit a41f9c2 staged behind gate
▶ awaiting human ack — agentproto ack s_7f2k

Deep dives — read the code

all deep dives →

Code-first, reproducible walkthroughs — each takes one real mechanism, walks the source, and ends with commands you can run on a cold clone. The three below turn a hostile agent (no ACP, no MCP) into a supervised orchestrator, in order.