Local-first · open source · Apache-2.0

Your agents.
Your machine.
Your data.

chief turns your own agent CLI into a team of workers. Describe the workflow to chief bot and watch it draw the steps, press Run, and answer the questions it stops to ask. It runs on this machine, writes everything to one folder, and talks to nobody unless your agent does.

Prefer the browser? Run it in a browser instead — nothing to install if you have Node. Either way it is free and unsigned; macOS asks once, and that is what the one command clears.

  • Your agent CLI — not a model chief resells
  • Everything on disk — one folder, one database
  • No telemetry — no account, no analytics, no phone-home
The chief designer: building blocks on the left, the workflow as a graph in the middle, the selected step's settings on the right.
The designer. Arrows are the order. It saves itself as you draw.
Works with Claude Code Codex Cursor Antigravity OpenCode Hermes

Built the opposite way round

Most “AI workflow” tools are a cloud account with your data inside it. chief is a program on your laptop that drives the agent you already pay for.

Bring your own agent

Claude Code, Codex, Cursor, Antigravity, OpenCode or Hermes — whichever you are signed in to. A step can even use a different one from the step after it, so a personal subscription and a work one can share a workflow.

Nothing leaves the machine

One server on 127.0.0.1, no login, one SQLite file in ~/.chief. Runs, prompts, results, costs and gate answers stay there. Delete the folder and chief has never heard of you.

No telemetry, ever

No analytics, no crash reporting, no account. The only traffic chief makes on its own is one request per launch to GitHub, to ask whether a newer version exists — and CHIEF_NO_UPDATE_CHECK=1 turns even that off. Everything else is what your agent CLI makes to the service you already use, and what an MCP server does for it.

Three moves, then a coffee

  1. 1

    Draw it

    Drag an Agent, a Gate, a Script or a Map onto the canvas and connect them with arrows. Write the prompt in plain words.

  2. 2

    Run it

    chief starts each step as its own agent session, hands it the results of the steps before it, and records every message as it happens.

  3. 3

    Answer it

    Where you asked for a say, the run waits. A gate shows a form worked out from what the step before found — pick, approve, or type. Then it carries on.

chief bot

Talk it into being

Say what the workflow should do. chief bot answers and builds it: the steps, their prompts and the arrows appear on the canvas as you talk, and when a choice is open it offers a few options you can pick or overrule. No dragging blocks around.

When a run goes wrong, open the run and ask the same bot what happened. It reads that run's own record — every prompt, reply, tool call and error — and tells you the cause and the next move, and can suggest the fix for the step that failed.

It runs nothing and changes nothing during a run: it writes workflows and their schedules before a run and reads records after one. Every change it makes to a workflow is listed under its answer with Undo.

chief bot in the designer: a conversation on the right, and the workflow it built on the canvas.
chief bot building a workflow from a description.

What is in the box

Everything below is in the app and over HTTP; the CLI is the same features for scripts and for when a page is not enough.

Visual designer

Blocks, arrows, a side panel per step. Autosave: what is on the canvas is the workflow the moment it is valid. There is no Save button anywhere.

Live runs

Watch a run on its own diagram, each step coloured by status. Open a step to read the prompt it was given, what it returned.

Gates, in plain words

Describe what you want to be asked; an agent builds the form from the earlier results. Pick several, pick one, approve, or type with options.

Replay every run

Every prompt as sent, every reply, tool call and result, stdin, stdout, exit code, gate answer and error — recorded as it happens, never re-executed.

Schedules

Cron, with the expression spelled out in words and the next firing shown. Add, disable and remove from a page; they fire while chief runs.

A CLI per step

The step's own agent CLI beats the workflow's, which beats the default. A step that chose its own says so on the canvas, so a mixed workflow reads at a glance.

Tokens per step

How many tokens each step used, per run totals, and a limit on how much of an earlier result goes into a later prompt — so a prompt cannot grow without end.

Undo a run

Each step reports the side effects it caused, so a run can be reverted in reverse order. Anything irreversible was gated before it happened and is listed for you.

Sandboxed scripts

A script step is Python, shell or Node with no network and no writes outside its folder unless you approve it. Protected paths cannot be granted.

Keeps going when you close it

Shut the window and the app stays in the Dock: runs finish, schedules fire, gates still notify. Quit asks first if a run is in flight.

Light, dark, text size

Light, Dark or System, and one text scale for the whole app from Small to Extra large. Both are chief's settings, so the window and the browser agree.

Have a look

Made-up data, a stand-in agent, the real app.

A run waiting at a gate: the form to pick tasks is inside the step that waits.
A gate mid-run. The run is stopped exactly here until you answer; closing the window does not lose it.
The run list: every run with its status, a new-run form above it.
Every run, with its status. Start one without leaving the page.
The schedules page: a workflow, its cron expression in words, the last and next run, and a switch.
Schedules. The expression is spelled out, and shows when it fires next.
Settings: the agent CLIs as cards with a switch each, their paths found automatically, and appearance above.
Your agent CLIs, found for you. Switch one off and it is offered nowhere.
A run open in a panel over the run list, in the dark theme.
The dark theme, with a run open in the panel over the list.

Two ways to run chief

The same engine, the same database in ~/.chief, the same screens. Pick the one that fits the desk you are at.

The desktop app

macOS 13+, Apple silicon

A window of its own. Close the window and it keeps running in the Dock — runs finish, schedules fire, and the bell tells you when a gate is waiting — until you Quit.

  1. Download it chief-0.2.7-apple-silicon.dmg Apple silicon · 128 MB · v0.2.7 · unsigned
  2. Drag it into Applications and open it macOS says it cannot be verified — that is the missing notarization, not a broken download.
  3. Clear the warning once The first time you open it, macOS says “chief” cannot be opened because the developer cannot be verified, with only Move to Bin offered. That is macOS checking an app nobody has notarized — not a damaged download, and the app is not going anywhere. Click Done, then run this once in Terminal: it clears the “downloaded from the internet” mark from this copy, and chief opens normally from then on.
    xattr -dr com.apple.quarantine /Applications/chief.app

The web UI

Any machine with Node 22.18+

Nothing to install but Node: chief serves the same app on this machine and you open it in a browser. Any computer that can run the agent CLI can run this, and the tab is only a viewer — closing it does not stop a run.

  1. Get chief
    npm ci
  2. Start the server
    npm run chief serve
    It starts the web UI, the scheduler and the API on 127.0.0.1:8080, and opens your browser.
  3. Open 127.0.0.1:8080 Runs keep going while chief serve runs, whatever the browser does.

Both need

  • One agent CLI, already signed in: Claude Code, Codex, Cursor, Antigravity, OpenCode or Hermes. chief finds its path for you in Settings → Agent CLIs
  • Node 22.18+ for the web UI, and for the desktop app only if a workflow uses a script step
  • Nothing else: no account, no cloud, no key of its own

From source, into an app

npm ci && npm run build:desktop

What it is careful about

Your keys are not its keys

chief has no API key of its own and no MCP configuration. Your agent CLI is already signed in; chief just runs it. It never speaks MCP itself and never edits a server list.

Local means local

The server binds 127.0.0.1 and rejects requests whose Host or Origin is not its own address. No login is needed because nothing is reachable from anywhere else.

Scripts run in a sandbox

No network, writes only in the step's folder and a scratch directory, an allowlist for environment variables, and protected paths (credentials, start-up files, git hooks) that can never be granted.

Irreversible steps ask first

A step that sends a message or posts a comment is gated before it runs, and a step that reported side effects is never retried automatically.