wicked-crew — the harness for the coding agents you already use

Your agents, in sandboxes.

crew drives each agent’s own headless mode in a hardened container, on your accounts and plans. No billing, no reselling.

Only the job’s worktree is mounted.
The model gateway is the only way out.
Keys stay in a broker. The container sees placeholders.
Sign each agent in once, on studio’s Agents tab.

Agent	Sandbox image
Claude Code	ships with crew
Codex	ships with crew
GitHub Copilot	ships with crew
OpenCode	ships with crew
Pi	ships with crew
Antigravity	no image yet
A local model, or your own binary	your own image, offline

Every gate says what it checked.

A gate decides on the layers that actually ran. One failed check stops the job, and no model can approve on its own.

Layer	What it checks	Its power
Evidence floor	The job changed its worktree.	a pass with no judge reads “floor only”
Your checks and policies	Deterministic rules you approved, re-run against the worktree.	one failure stops the job
Evaluator	A different agent reviews the work.	can send it back
Judge	An AI judge reads the evidence cold.	can fail a gate, never approve alone
QE acceptance	Garden’s QE pipeline, read from the QE ledger.	only an attributed PASS delivers

When a gate is weaker than that, its receipt says so:

approved by default, not verifiedthe phase had no gate
floor onlyno distinct judge could review it
reduced assuranceyou let the builder review its own work
QE acceptance skipped by operatoryou skipped it, with your reason

A council settles disputes. It doesn’t pick agents.

Agents are routed by their skills. A council is convened only when a reviewer and a builder disagree about a finding.

FindingThe reviewer flags a problem, with the file and line.
DeclinedThe builder disagrees and says why.
HeldThe reviewer stands by it. Now it is a dispute.
CouncilThree agents, drawn at random with a recorded seed, rule on it.
RulingYes, no, or no verdict, with the agreement and the dissent.

The ruling is on the gate’s receipt.
Council size is a setting: 1 to 9, default 3.
After two send-backs, the job asks you instead.

Workflows are data. Jobs survive a crash.

Pick a built-in, or save your own steps as a preset (PUT /api/v1/presets/:name).

`feature`	clarify → design → build → adversarial-review → test → review
`bug`	triage → reproduce → fix → verify
`migration`	plan → execute → cutover → verify → cleanup
`mcp-server`	build an MCP server, then install it after your consent
`editor-plugin`	scaffold a studio editor, test it, install it after your consent

Crash-resumable. A daemon that dies mid-job resumes it at boot, at the step it was on.
One open API. Studio ships inside crew and uses the same /api/v1 and /ws your tools can.
Headless too. wicked-crew start, the REST API, or wicked-crew mcp for agents.
One machine today. A remote runner is the direction, not a shipped feature.

The wicked platform

One surface. One control plane.
One catalog. One record.

Your coding agents do the work. A different agent checks it against your rules. You get the evidence. Four layers make that work. Hover or tab through one to see what it does and what passes between them.

Experience Where you direct your agents The screens you use: start a job, watch it, answer the decisions that need you.

wicked-studiothe surface

Start a job with a chat, confirm the plan, then watch your agents work in sandboxes and answer the decisions that need you.

ws.wickedagile.com↗ repo

start a job · watch it · answer its questions — through one API

Control Intent in, gated work out Runs the jobs, checks every step, keeps the evidence: your coding agents as sandboxed workers.

wicked-crewthe control plane

Runs your agents in sandboxes. A different agent reviews the work, and app changes pass acceptance testing before delivery.

you are here

calls skills from sandboxed agents — one failed check stops the job; every verdict lands in the record

Capability What agents can do Skills, tools, playbooks and the QE specialists: how agents touch the record.

wicked-gardenthe catalog

The skills your coding agent can’t build alone: proof from evidence, graph-aware refactors, 40 QE specialists.

wg.wickedagile.com↗ repo

reads & writes the record through its contract — never around it

Foundation The system of record Code graph · memory · knowledge · evidence · events. Local, with no servers to run.

wicked-estatethe record

What your agents know about your system: a 114-language code graph, memory and knowledge in one local binary.

we.wickedagile.com↗ repo

wicked-interactivethe document engine

The engine behind the docs, decks and captioned demo recordings studio produces. You use it through studio.

repo ↗

Install it with the family. Studio comes with it.

Node 22 or later
A container runtime: Docker, Podman, Rancher Desktop, Colima or OrbStack
wicked-garden 12.40.0 or later, for the skills every agent follows

Terms of service: crew drives CLIs you installed under your own accounts. Check that each one’s terms allow programmatic use. · Why this exists →

Recommended · family installer

npx wicked-installer   # picks crew, garden and the rest

Installs crew and garden together. wicked-crew npm v0.15.7

Direct · wicked-crew only

npm install -g wicked-crew
wicked-crew serve   # opens studio on 127.0.0.1:7701

View on GitHub