Quickstart
From nothing to a staffed team in four steps. This page assumes a terminal, Node 24, and a model endpoint you can reach.
Before you start#
- Node.js 24+ — check with
node --version. - A model backend — a local OpenAI-compatible server needs no key. For hosted models, keep the key and base URL ready.
- An MCP client — install and open it once before running setup, so its config exists and can be detected.
1. Install#
Clone the repository, install dependencies, and link the two binaries. On macOS and Linux, install.sh does all three.
git clone https://github.com/lab1207/Wednesday.git
cd Wednesday
./install.sh
git clone https://github.com/lab1207/Wednesday.git
cd Wednesday
npm install
npm link
npm link puts wednesday (the operator CLI) and wednesday-mcp (the server) on your PATH. The full guide, including the manual route, is Install & connect.
2. Connect your client#
One command registers the server with every client it can find on this machine.
wednesday setup
# plan only, write nothing
wednesday setup --dry-run
# one client
wednesday setup --client cursor
Each file is backed up as <file>.wednesday-backup-<timestamp> before it is written. A config that is not strict JSON is reported and left alone. The registration pins this checkout with an absolute command, so it does not depend on a shell shim.
3. Staff the crew#
The team is the operator’s decision, not the calling model’s. There is no MCP tool to change it. Create a crew and make it active with the CLI.
# create a crew and make it active
wednesday team add demo coder=<model> reviewer=<model> --shape auto
wednesday team use demo
# inspect it
wednesday team show
A crew carries seats, per-seat models, and a policy: shape, repair cap, token budget, dollar budget, test command, and required files. Every MCP connection adopts the active crew at startup. If a crew declares no shape, auto is used and the task difficulty decides.
See Crews & shapes for all nine seats and the five shapes.
4. Hand over a request#
In your client, ask for the work in plain language. Many clients also surface the Wednesday prompt as a slash command — /wednesday:run in OpenCode, /mcp__wednesday__run in Claude Code.
# the prompt is the reliable path: it tells the host model to hand over
/mcp__wednesday__run add a quantity input to the cart page
# or call the tool directly from the client
wednesday_run { "request": "add a quantity input to the cart page" }
The tool returns a job id immediately. Follow it with wednesday_wait, then read wednesday_result. wait returns within 55 seconds with done: false if the run is still going; call it again rather than treating the gap as a failure.
What to expect from the first run#
- It runs unattended. Over MCP the team works inside its sandbox. A client-side approval gate wired to MCP elicitation is not built yet.
- Tests decide. A declared verification command that fails blocks the run and feeds the repair loop. If no test command is declared, the tester passes with a note; the run is not treated as failed for the absence of a plan.
- Nothing is committed. The run changes files in the workspace and leaves a receipt. Committing is your action.
- One job at a time. The queue is deliberate. A second
wednesday_runwaits behind the first.
Next#
Read the MCP tool reference before wiring a client integration. Read Verification before trusting a green result. Read Security & data before putting a real key on the machine.