cheep — a multi-agent coding shell you can watch

Live anatomy of a task

One prompt in. A supervised fleet out.

You talk to one agent. It decomposes the task, fans subtasks out to cheap executors in parallel — each in an isolated git worktree — validates the diffs, and merges only what passes. Every agent gets a tab. Every token gets metered.

How it works

Plan · delegate · validate · land

Why cheep

Frugal by architecture,
safe by default.

/nomistakes on
┌ approve: merge → cheep/local-qwen-1 ─┐
│ + func LoadTOML(p string) …          │
│ - func LoadJSON(p string) …          │
│ y: allow · n: deny · ↑/↓: scroll     │
└──────────────────────────────────────┘

Showcase

Impressive things, and exactly
how to run them.

› audit this repo: dead code, missing error
  handling, flaky tests — one prioritized list
# three kind:"scout" subtasks, one delegate call,
# a tab per scout — checkout untouched

/budget 2
› fix ./internal/parser until
  `go test ./internal/parser` passes
# iterate_until{subtask, check} on the
# cheapest executor — rounds are ~free

/nomistakes on
› migrate the config package to TOML
# … executors work, checks pass, then the
# approval overlay shows the branch diff

/delivery pr
› split User into Identity + Profile,
  update all call sites
# result: "pull request opened" + URL each

/fork   # pick the turn to redo
› use a worker pool instead
/tree   # both attempts side by side

{ "orchestrator": { "model":
    "claude-sonnet-4-6:high" },
  "executors": [ { "model":
    "qwen/qwen3.6-35b-a3b" } ] }

cheep pi add pi-web-access
› research the top 3 Go HTML sanitizers,
  recommend one
# executors now have pi__web_search & co.

$ cat .cheep/prompts/review.md
Review the $1 package. Focus: $ARGUMENTS
/review internal/worktree locking

› run the test suite and fix failures
  every weekday at 9am
# scheduled 0 9 * * 1-5 · manage with /scheduled
cheep daemon  # fires jobs on cadence

› research the latest Postgres 17 features
# qwen has no web tool → auto-improve spots it:
🔧 fetch would help here — /autoimprove install
# added; agents can now read the web

"mcp": { "claude": {
    "command": "cheep-claude-mcp",
    "roles": ["orchestrator"] } }
# orchestrator gains claude__ask_claude —
# hard bits run on your plan, grunt work stays local

# in Claude Code, on your Pro/Max plan:
› build the appointments module
# Claude delegates the grunt work →
$ cheep run "write the routes + tests
  for appointments" --workdir . --json

Run it anywhere

Walk away mid-task.
Pick it up on your phone.

cheep is a plain terminal app, so the most reliable portable setup needs zero new software: run it in tmux on your machine and attach from anywhere over Tailscale SSH — the same live session, running task, agent tabs, and pending approvals. A long delegation keeps working while your phone is in your pocket, and crash markers mean interrupted work is surfaced, never stranded.

tmux new -As cheep   # attach or create
cheep
› refactor the storage layer;
  run the full suite when done

ssh my-mac -t 'tmux attach -t cheep'
# flaky signal? mosh roams + survives sleep:
mosh my-mac -- tmux attach -t cheep

Pair with /nomistakes on so nothing merges without your sign-off while you're away, and /stow before a break to write a handoff note. Full setup — Tailscale, phone keyboards, cheat sheet — in the portable guide.

Install

Thirty seconds to a fleet.

brew install TedHaley/homebrew-tap/cheep

go install github.com/TedHaley/cheep@latest

Then just run cheep — first launch discovers local model servers (Ollama, LM Studio, vLLM, llama.cpp) and API keys on your machine and walks you through picking an orchestrator and executors. Setting up a project? cheep init scaffolds AGENTS.md with detected validation checks.

Reference

Commands