OpenFigs

A small, open, file-based skeleton for building back-office AI-employee agents — that report to a human through Figs.

The agent runtime (Claude Code, Codex, opencode) is commoditizing. What doesn't: a good way to build an agent you can trust with a recurring, consequential job, and the trust layer that lets you actually delegate it. OpenFigs is the build side; Figs is the trust layer.

OpenClaw, but for enterprise — you can see everything. An army of focused, single-purpose agents you can trust, not one opaque super-agent you install. You clone a transparent, file-based fleet (plain files + prompts you read end to end) and run one agent per job. Nothing hidden, no risky third-party sprawl; the Figs app is the live window onto what they're doing.

Get started

Requires Node ≥ 18. Scaffold a fleet — and your first agent — in one line:

npm create openfigs@latest my-fleet

That fetches the latest skeleton, wires up the runtime symlinks, and (optionally) scaffolds your first agent. (npx create-openfigs my-fleet is equivalent — as are yarn create openfigs and pnpm create openfigs.) Then:

cd my-fleet
# fill in agents/<name>/AGENTS.md — its role, mandate, and the loop it runs
# add more agents any time:  npm run new-agent <name>
# then connect to Figs so your manager can see it (see "Connect to Figs" below):
npx @figs-so/cli@latest login && npx @figs-so/cli@latest init

⚠️ Don't "Download ZIP" from GitHub. This repo uses symlinks (CLAUDE.md → AGENTS.md, .claude/skills → .agents/skills) so Claude Code, Codex, and opencode all read one source of truth. GitHub's ZIP export turns them into broken text files. The commands above (and git clone) preserve them; if a symlink ever breaks, run npm run fix-symlinks to repair it.

Philosophy

• Infra, not dictation — conventions + guardrails in plain files; what the agent does is between the agent and its user.
• The agent is an employee — it does the job, learns from its work, keeps its own code clean, and reports for sign-off.
• Trust via visibility, not prohibition — agents self-improve and loudly flag what they changed; humans sign off on outcomes.
• Small & in control — plain files, auditable, no heavy framework. Runtime-agnostic via AGENTS.md.

How it works

• One folder = one agent — agents/<name>/ with its own AGENTS.md + MEMORY.md + reports/ + .figs/. The root AGENTS.md is shared by every agent and composes above each one.
• Small, focused agents — one agent, one job. Spin up a new dedicated agent for a new scope rather than growing a do-everything one.
• Runtime-agnostic — AGENTS.md is canonical; CLAUDE.md is a symlink to it, so Claude Code, Codex, and opencode all read the same source of truth.
• Scaffold a new agent, fill in its guide, run it on your runtime, connect it to Figs.

Start with AGENTS.md — the operating guide every agent inherits.

Connect to Figs

Every agent keeps a local activity log (.figs/runs.jsonl · asks.jsonl · artifacts/) — useful on its own. To give your team a shared, read-only window into your agents — the org chart, what each is handling, what's done, and what needs a human — connect them to Figs: install the figs CLI, run figs init in an agent folder, then figs push. The agent already produces the log; Figs just mirrors it. (CLI + UI are open source.)

The Figs ecosystem

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

Layer	Repo	License	Role
🏗️ Build	OpenFigs	MIT	build trustworthy back-office AI employees — conventions + skeleton, runtime-agnostic — ← you're here
📤 Report	.figs + CLI	MIT	the open standard an agent reports its state in
👁️ Govern	Figs app · app.figs.so	AGPL-3.0	the org chart + handoff inbox humans read

Status

Early — working name OpenFigs, MIT licensed. Distilled from two real, dogfooded agent fleets (a finance/compliance execution fleet and a product-analytics fleet).