Run a fleet, not a tab farm
Launch three, five, ten features. Cap how many run at once, triage only what needs you, batch-approve low-risk gates, and retry the failures. A circuit breaker flags runs that keep failing.
Open-source AI coding agent orchestrator
Shep runs parallel agents — Claude Code, Codex, Copilot, Cursor, Gemini, Kimi Code, Cline, or local models — each in its own git worktree, through implementation, CI, and review. Drive them from the web, the CLI, or your terminal. Open source, local-first, cross-platform.
Runs locally with Node 22+. Opens the dashboard at localhost:4050, or run shep tui in your terminal.
Free and MIT-licensed. Bring the agent subscription you already have.
describe → (spec) → agents → review → pull request
Every feature moves through the same pipeline — you decide how much of it runs on its own.
Type it in the dashboard, run shep feat new, or start an app with shep app new. Shep cuts a branch and a worktree.
Enable spec-driven mode and Shep writes requirements, research, and a plan as YAML you approve before any code.
# specs/003-add-pagination/spec.yaml name: add-pagination number: 003 oneLiner: Cursor pagination for GET /todos. requirements: - page + limit query params, default limit 20 - response includes total count tasks: - id: task-1 # cursor decode - id: task-2 # paginated query gate: awaiting your approval
Each feature gets its own git worktree and agent session — Claude Code, Codex, Copilot, Cursor, Gemini, Kimi Code, Cline, or a local model. Pick the agent per feature.
Shep commits, pushes, opens a draft PR, watches CI, and pushes fixes when it fails. When an agent needs you, it asks with a recommended answer.
Features land as standard pull requests on named branches. Your CI, your reviewers, your merge — unless you flag --allow-merge.
feat/add-paginationmain
all checks passed — awaiting your review
shep tui
shep tui is a live console of every feature. Whatever needs you is pinned to the top, and single keys approve, reject, stop, and resume. It reads the same local state as the dashboard and runs without it.
Inside tmux, Zellij, WezTerm, or kitty, t and f open a feature's shell or live agent log in a split pane.
features
Launch three, five, ten features. Cap how many run at once, triage only what needs you, batch-approve low-risk gates, and retry the failures. A circuit breaker flags runs that keep failing.
Each AI agent works in its own git worktree on its own branch. No clobbered files, no port collisions, no stash juggling. Your checkout stays untouched.
After push, Shep watches your pipeline. When it fails, the agent reads the logs and pushes a fix — up to three retries, configurable.
Unsure agents ask instead of guessing, with a recommended answer and a deadline. Answer from the web, the CLI, shep tui, or chat.
Shep snapshots the worktree after every agent step. Diff any two steps, or restore one: shep feat checkpoints, diff, restore.
A chat-first dashboard at localhost:4050, with specs, previews, and plans inline in the conversation. Every action is also a shep command, and shep tui is a live console of the fleet that approves, rejects, and opens split panes.
Claude Code, Codex, Copilot, Cursor, Gemini, Kimi Code, Cline — or models through OpenRouter, Together AI, and Ollama. Swap per feature, per repo, anytime.
MIT-licensed. State stays in ~/.shep/; code goes only to your agent. shep telemetry off stops usage metrics.
what is shep
Shep is an AI coding agent orchestrator that lets you run multiple Claude Code, Codex CLI, Copilot CLI, Cursor CLI, Gemini CLI, Kimi Code, and Cline sessions in parallel — each in its own git worktree, through CI and review to a pull request. Manage the whole fleet from the web dashboard, the CLI, or the shep tui terminal console.
faq
Run `shep feat new` once per feature. Shep gives each session its own git worktree and branch, so parallel Claude Code sessions never touch the same files. Shep manages the worktrees for you — creating them, committing in them, and opening PRs from them — while you follow everything from the web dashboard, the CLI, or the `shep tui` terminal console. The same works for Codex CLI, Copilot CLI, Cursor CLI, Gemini CLI, Kimi Code, and Cline — and you can mix them, one agent per feature.
A control plane that runs and coordinates multiple AI coding agents: it assigns each one an isolated workspace, drives the git workflow around it — branch, commit, push, PR — watches CI, and reports status in one place. Shep is an open-source AI coding orchestrator: MIT-licensed, local-first, and agent-agnostic, so one dashboard manages parallel AI coding agents from different vendors. It also lets you cap how many features run at once, triage only what needs you, and batch-approve gates across the fleet.
Every feature gets a dedicated git worktree on its own branch. AI agents working on different features never share a checkout, so there are no clobbered files, port collisions, or half-finished edits crossing between tasks — and your own working directory is never modified. When the agent finishes, Shep commits, pushes, and opens the pull request from that worktree, which is preserved until you delete it. Shep also takes a git checkpoint after every agent step, so you can diff or restore any step (`shep feat checkpoints`, `shep feat diff`, `shep feat restore`).
7 CLI agents are built in — Claude Code, Codex CLI, Copilot CLI, Cursor CLI, Gemini CLI, Kimi Code, and Cline — plus OpenRouter, Together AI, Ollama, and any OpenAI-compatible proxy. Pick the model and reasoning effort per feature (`--model`, `--effort`), and swap agents per feature, per repo, anytime: run Claude Code on one feature and Codex CLI or Gemini CLI on another at the same time.
Yes. Shep is MIT-licensed and runs on your machine. All state lives in `~/.shep/` as a SQLite database, and your code goes only to the agent you configure, under that agent's own terms. The one thing Shep itself sends is anonymous usage metrics — which commands and screens are used and how runs end, never prompts, code, file paths, or repository and branch names. They are on by default; `shep telemetry off` (or `DO_NOT_TRACK=1`) turns them off, and they are always off in CI. Start with `npx @shepai/cli` — no account required. Prefer zero install? app.shep.bot runs Shep in your browser.
Shep puts several layers between the agent and your main branch: the agent works in an isolated worktree, never your checkout, with a git checkpoint after every step; its output lands as a draft pull request that your CI, linters, and security scanners run against; and nothing merges without your approval unless you explicitly pass `--allow-merge`. Every action is logged, a circuit breaker flags runs that keep failing, and `shep agent stop` halts a run immediately — the worktree is preserved.
Yes. `shep tui` opens a live console of every feature. Features that need you are pinned to the top, and single keys approve, reject, stop, or resume a run. Inside tmux, Zellij, WezTerm, or kitty it opens a feature's shell or agent log in a split pane. It works without the web dashboard, and every action is also a plain `shep` command.
Yes. `shep app new "<idea>"` starts a new app on any stack; its first feature is planned first by default, so you can review the plan before code is written (pass `--fast` to skip that). Or pass `--starter vite-shadcn` for a Vite + React + Tailwind + shadcn app built straight from your prompt, with a live preview. Then add features to it with `shep feat new`.
Free and MIT-licensed
Shep adds the workflow around your agent — worktrees, commits, CI watching, pull requests. No Shep account, no usage meter, no cut of your tokens. Your code goes only to the agent you pick; anonymous usage metrics are on by default and shep telemetry off turns them off. The source is on GitHub; read it before you run it.
Live count, straight from the API. We don't cache our own traction.
Open source
local-first · your code stays yours
Works with Claude Code, Codex CLI, Copilot CLI, Cursor CLI, Gemini CLI, Kimi Code, and Cline — or models through OpenRouter, Together AI, Ollama, or any OpenAI-compatible proxy.
CLI agents
Model providers
Pick the agent and model per feature, per repo, anytime.