OpenFigs

A thin, open, file-based skeleton for building one back-office AI employee — an agent that owns a recurring job, learns from its work, and reports to a human through Figs.

OpenClaw, but for enterprise — you can see everything. Deliberately thin: you clone a transparent, file-based agent (a few plain files + prompts you read end to end) and run it on the runtime you already use — not one opaque super-agent you install. Want a team? Run one per job, point them at the same Figs workspace, and the org chart draws itself.

Get started

Requires Node ≥ 18. Two ways to scaffold one employee:

1 · You run it (in your terminal):

npm create openfigs@latest my-agent

2 · Your agent runs it — paste this to the coding agent you already use:

Set me up a Figs employee — a focused AI agent that owns one recurring job and reports to me.
1. Scaffold it:  npm create openfigs@latest <name>   (name it for the job, e.g. "invoice-recon")
2. cd in, then read the full guide once, end to end:  https://figs.so/llms.txt
3. Fill the charter (AGENTS.md + .figs/agent.json) from what we're building, then delete the
   read-me block at the top of AGENTS.md.
Then tell me what it'll do before it does any real work.

Either way, the scaffolder fetches the latest skeleton, wires the runtime symlink, and runs figs init --yes — so your agent has a local identity + activity journal and is ready to work, no account needed. When your team should see it: figs login → link → push to appear on the org chart.

One repo = one employee. A new, distinct job is a new agent — run npm create openfigs again, never copy a folder (a copy carries the original's identity). A "fleet" isn't a directory — it's a Figs workspace that several agents report to.

⚠️ Don't "Download ZIP" — this repo uses a symlink (CLAUDE.md → AGENTS.md) so every runtime reads one source of truth; ZIP breaks it. git clone and the commands above preserve it.

Why thin

• Transparent, not a framework — plain files you read end to end; nothing hidden, nothing to trust blindly.
• One employee, one job — a focused agent beats a sprawling one; a new job is a new agent.
• Trust via visibility — the agent self-improves and loudly flags what it changed; humans sign off on outcomes.
• Runtime-agnostic — AGENTS.md is canonical; CLAUDE.md symlinks to it, so Claude Code, Codex, and opencode read the same source.

How it reports (Figs)

The agent keeps a local activity journal from day one (account-free). Connect to Figs (figs login → link → push) to give your team a read-only window — the org chart, what each agent's handling, what's done, and what needs a human. The agent records itself with the CLI verbs (checkpoint / report / ask / answer / close); the full guide is canonical at figs.so/llms.txt.

Running on a schedule

Figs can't run your agent — wire it up yourself (cron, launchd, CI). An autonomous employee needs two schedules: a work trigger (what kicks off the job) and a separate inbox cadence (figs inbox → act → figs close, to catch replies that arrive while it's away).

Staying current

A clone is yours — skeleton improvements are never pushed into it. CHANGELOG.md records each revision; npm run check-skeleton (the skeleton-updates skill) shows what you haven't reviewed, so you decide what to adopt. Humans: Watch → Releases.

The Figs ecosystem

Figs is one stack in three pieces — build → report → govern. Land on any repo; here's the whole picture:

Layer	Repo	License	Role
🏗️ Build	OpenFigs	MIT	build a trustworthy back-office AI employee — conventions + skeleton, runtime-agnostic — ← you're here
📤 Report	.figs + CLI	MIT	the open standard an agent reports its state in
👁️ Govern	Figs app	hosted	the org chart + handoff inbox humans read