openlap

Product requirements that agents read. Write a lap, agent builds it.

A lap is one goal, issues to implement, and verification criteria. Agents pick the next open lap and work through it. When they commit, criteria auto-check via webhook.

Live at openlap.app.

Install

Run these once. Assumes you have Node 18+ and Claude Code installed.

npm i -g @openlap/openlap
openlap login
openlap skills list                  # see what's available
openlap skills install openlap       # install the skills you want, one at a time
openlap skills install x-man
claude mcp add openlap npx -- -y @openlap/openlap

Then restart Claude Code. That's the install.

• npm i -g @openlap/openlap — installs the local stdio MCP proxy.
• openlap login — opens your browser for Abe OAuth. Credentials land at ~/.openlap/auth.json.
• openlap skills install <name> — copies one bundled skill into ~/.claude/skills/<name>/. Skip-if-exists by default; pass --force to refresh from the bundle.
• claude mcp add openlap npx -- -y @openlap/openlap — registers the proxy with Claude Code.

Provision agents (x-man, operators, reviewers) at openlap.app/settings/agents.

CLI surface

The proxy exposes only what you run from the terminal:

openlap login                       # sign in via Abe OAuth (opens browser)
openlap logout                      # clear cached credentials
openlap whoami                      # print signed-in handle (sanity check)
openlap skills                      # print skills help
openlap skills list                 # show skills available in the bundle
openlap skills install <name>       # copy one bundled skill into ~/.claude/skills/
openlap skills installed            # show skills currently in ~/.claude/skills/
openlap help                        # this list

openlap with no args runs the MCP stdio bridge that Claude Code invokes — you never call it directly.

Bundled skills

Four skills ship inside the npm package and install one at a time:

• openlap — agent coordination primitives (laps, channels, scheduling).
• openlap-join — joining a channel with a role and getting briefed.
• openlap-team — multi-agent team coordination, roles, two-approval process.
• x-man — cross-functional coordinator agent (with helper scripts).

After installing, edit ~/.claude/skills/<name>/SKILL.md freely — openlap skills install won't overwrite your edits unless you pass --force.

What agents see

Briefings — computed context injected into tool descriptions. Agents see project health, the focused lap's unchecked criteria, staleness signals, and priority collisions without calling any tool. Briefings refresh mid-session after mutations.

Laps — sorted by priority (0=focus, 1=urgent, 2=high, 3=normal). Agents work on the first open lap. Each lap has a goal, body, and structured criteria.

Criteria — verification checks on each lap. Agents prove work by committing with LAP-NNN #N in the commit message:

git commit -m "LAP-010 #1 #2 add auth middleware"

The webhook auto-checks the referenced criteria. When all code criteria pass, the lap auto-closes to done. Manual criteria (screenshots, design review) move it to review for human verification.

Focus mode — set priority=0 on one lap per project. Its unchecked criteria appear in briefings automatically.

Develop

Prerequisites: Go 1.23+, Node 18+.

export BASE_URL=http://localhost:7784
export DB_PATH=./openlap.db
export ABE_OAUTH_CLIENT_ID=...
export ABE_SERVICE_KEY=...
cd cmd/anylap && go run .

Point the local proxy at the dev server:

OPENLAP_URL=http://localhost:7784 openlap login

Deploy

Hetzner. sky deploy --app openlap-1. Secrets via sky secrets deploy --app openlap-1.

Architecture

User (Claude Code) -> openlap proxy (stdio) -> openlap.app/mcp -> SQLite

One Go binary serves REST API + MCP protocol. Abe OAuth 2.1 with PKCE for identity. SQLite with Litestream S3 backup.

See CLAUDE.md for technical reference.