MCP server · beta · Apache 2.0

Small open models,
staffed as a
coding team

Install wednesday. Bring the models you already have.
Connect it to the harness you already use.
One tool in. The whole team inside.

Node 24+ · your key or your local model · runs on your machine

Cheap models can write code.
They cannot ship checked code.

Training a frontier model is a lab’s job. Shipping a verified coding seat is not.

One model, alone
small model → unchecked output

Guesses. Hallucinates. Ships broken.

→
wednesday crew
planner → coder → real tests ⇄ repair → checked + receipt

Same models. The structure around them is the product.

Verification runs real commands Repair feeds failures back Roles: planner · coder · vision · worker No token meter on your side

One request.
A verified result.

01

Request

You say what you want in the harness you already use.

02

Plan

wednesday plans the work, the files it will touch, and the verification commands.

03

Staff

Your crew runs it: the seats and models you chose, the shape you chose.

04

Build

Controlled sandbox. Explicit tools. Real file changes.

05

Verify

Real tests + security scan. The tests decide, not the model.

06

Repair

Failures feed back into a capped repair loop until green.

07

Review

You see the diff. Then ship with a full receipt.

Two buyers.
One install.

01 — Host + Resell

Sell AI coding without training a frontier model

Install wednesday, download open-weight models, host them on your own machines, and offer a coding service as an MCP server your clients connect to. Your margin is cheap hosting sold as a seat.

  • Software agencies of 10–80 people
  • GPU hosts & IT firms
  • Telcos entering AI
  • Regulated buyers needing data residency
02 — Switch inside product

Already ship a coding product

Install wednesday on your own servers and offer it as a switch users can pick. Everyday coding sits on a measured small-model team you host. Your GPU budget stays on the flagship models you still want to build.

  • Labs and coding-product vendors
  • Agencies with capped coding plans today
  • A low monthly plan that stays on all day
  • Meter is your GPU, not a token cap

One tool in.
The whole team inside.

wednesday speaks MCP over stdio. Your harness gets a small, coarse surface; the seats stay on this side of the wall.

wednesday_run

Start a run. It validates the workspace and hands back a job id immediately, so the client never blocks on a long call.

wednesday_wait / status

Clients kill a tool call on their own timeout. wait returns within 55 seconds and reports done: false rather than hanging — and streams progress notifications when the client supports them.

wednesday_result

The verdict: status, summary, files changed, real test results, acceptance, diagnostics. A receipt, not an opinion.

wednesday_cancel / jobs

Stop a queued or running job, or list this session's runs, newest first. Nothing is hidden from the operator.

wednesday_team

Read-only. The seats, models and policy you staffed — and whether this connection adopted the active crew.

No seats. Ever.

The outside model cannot call architect or coder. If it could, it would become the team lead and the small-model crew would become day labour. The surface is coarse on purpose.

Claude Code

claude mcp add wednesday -- wednesday-mcp

OpenCode

opencode mcp add wednesday -- wednesday-mcp

Cursor / Claude Desktop

{"mcpServers":{"wednesday":{"command":"wednesday-mcp"}}}

Codex CLI

[mcp_servers.wednesday] with command = "wednesday-mcp"

Any OpenAI-compatible endpoint.
Any mix of sizes.

Local servers need no key. Hosted models use your own key. Teams let you assign a different model to each role.

Local first

Auto-detects Ollama (:11434), LM Studio (:1234), llama.cpp (:8080), vLLM (:8000) and any local server (:1337). Zero configuration when the model is already running.

Hosted

OpenRouter, NVIDIA NIM, DeepSeek, Groq, Together, Mistral, or any OpenAI-compatible base URL. Keys stay in ~/.wednesday/.env.

Named teams

Six real seats — architect, planner, coder, tester, reviewer, vision — each with its own model, and a team policy that says how the team works: solo, duel, vote or relay, the repair cap, the token and dollar budgets, and the command every task is verified with. The receipt shows what each seat spent. Teams are plain JSON in your local state, staffed by the operator and adopted by every client connection.

Small models preferred

Built for models under ~7B (down to 1B and below). You choose: pure small, frontier, or any mix. Turn on team-up and the factory probes your model pool, scores every model per capability, and hands each seat to the model that scores best for it — one strong model can hold several seats. Context windows are read, not guessed: Qwen2.5-Coder-3B is a 32K model, but the Colab T4 deployment serving it runs a 13,312-token window — the smaller one wins.

Bring your own knowledge.
And your own tools.

The factory's structure is the point — so the operator's own playbooks and the operator's own tools belong in it.

The team asks

When a request is ambiguous, the architect and the planner can stop and put one question to you at a time instead of guessing at what you meant. The question flow rides the chat route today; MCP elicitation is on the list.

Images reach the model

Images reach the model: attach the picture to your request and the prompt carries the image itself, not a path the model cannot open.

Your own skills

A skill is a folder with a SKILL.md in it. Drop it into ~/.wednesday/skills/ (or your project's .wednesday/skills/) and Wednesday finds it on the next request. A .skill package installs the same way — and the same audit that vets the factory's own skills refuses one that carries an instruction-override directive.

Your own MCP servers

Attach a server over stdio or Streamable HTTP and its tools become callable by the model as mcp____. Wednesday speaks MCP itself, over stdio and Streamable HTTP, with no extra dependency.

The tests decide.
Not the model.

Verification runs real commands. Repair loops feed failures back until green. Every run leaves a receipt.

01

Real tests

Not simulated. Actual test runners and security scans gate the result.

02

Capped repair

Failures are fed back into the crew with a budget. It stops when green or the budget is exhausted.

03

Receipt

Every run is persisted with its plan, events, and evidence. wednesday_result · wednesday_status · wednesday_jobs

One command.
Your machine.

Recommended

From source

git clone https://github.com/lab1207/Wednesday.git
cd Wednesday && npm install
npm link
Register

One command

wednesday setup
# every harness it finds
# backups before every write
Staff

Pick the seats

wednesday team add demo \
  coder=<model> reviewer=<model>

Requires Node.js 24 or newer. wednesday setup registers Claude Code, Claude Desktop, Codex, Cursor and OpenCode — through their own mcp add command where one exists, and with a backup of every file it touches.

The free core is the whole product.
Nothing held back.

Apache 2.0. Full source on GitHub. PRs welcome.

Install once.
Own the factory.