Manage the build, not the chat.
Bennch is the workspace for software built with AI. Your coding agent connects to it over MCP, builds from a clear plan of cards and reports back, so you can see what’s planned, what’s being built, what needs your decision and what’s actually been checked.
Live now with Claude Code. Codex and more coming soon.
Ledgerly, an example project in Bennch: the next step, what needs you and where the build stands.
AI writes the code. Bennch keeps the project on track.
Building fast is the easy part now. Knowing where the project stands is what gets hard.
The whole project, every session.
The plan, your decisions and design links live with the project, so a new session never starts from zero.
Your AI builds what you approved.
Extra ideas come back as proposals. Choices that are yours come back as questions.
Done means you’ve checked it.
Not when the AI says so. Each card shows its commits and tests, and only you can accept it.
Press Build once. Claude works through the list.
Every piece of work is a card. Claude takes them in order, in the same Claude Code session, and hands each one back with proof. You don’t start them one by one.
Up next 4
Claude works down the list
Turn an accepted quote into an invoice
Attach photos of the finished job
Mark an invoice as paid in cash
Monthly income summary
Claude is building 1
One at a time, in order
Today screen: jobs and unpaid invoices
iOSAndroid
2 commits so far
Tests passing
Next up: LD-24, in the same Claude Code session. Nothing for you to start.
Your check 2
Built. Only you can finish these
Sign in with an email link
iOSAndroid
4 commits · tests passed
Remind clients when an invoice is overdue
2 commits · tests passed
1You approve the cards
What each is for and how you’ll know it works. Unapproved cards can’t be started.
2Claude builds them in a row
One card, or as many as you like. Claude finishes one, starts the next, and only stops when something needs you.
3You check the proof
Try it on your own devices. A failed step goes back to its card with your note.
More than a task board.
The plan, the decisions, the code and the tests stay connected from the first card to launch.
A plan that covers the whole product.
Take Ledgerly, an invoicing app for tradespeople. It has a phone app, a page where clients pay, and a backend. Each card belongs to one of those parts, sits on a board for its kind of work, such as design, backend or security, and says where it runs: iOS, Android or web.
See the whole product at once, or just the phone app’s next few cards, and have your AI work through them in order.
Decisions stay with the work.
When a choice is yours, such as how data is stored, payments or security, your AI asks with options and a suggestion instead of deciding. Answer in the chat or in Bennch.
Your answers stay on record for every later session, and cards link to where their designs live, like a Figma file.
Pick up where you left off.
The Overview shows the next step and what’s waiting on you. If a session stops halfway, the next one picks that card up first. Run several AI sessions at once and each card shows which one is building it.
Evidence attached to what was built.
Each card shows its commits, the tests the AI reported, your CI result and which platforms were covered. Work that grows far beyond its card is flagged.
For whole journeys, like sending an invoice and getting paid, the AI writes a step-by-step test guide linked to the cards it tests, so one run on your phone can accept several cards.
Start from an idea, or from code you already have.
You don’t pick a mode. Claude looks at your project and starts the right way.
Plan it together
Claude asks what you’re building, who it’s for and what the first version must do. It plays back a short brief, then proposes cards for you to approve.
Claude: “What are we building?”
Map what’s there
Claude reads the code and proposes cards for what’s built and what’s left. Built work comes to you to check in one go; new cards are only for what’s missing.
Claude: “This folder has code, so I’ll map what’s built and what’s left.”
Up and running in a few minutes.
Join the waitlist
We’ll email your invite. A short practice project, about 20 minutes, shows you how it all works.
Add Bennch to Claude
Add it as a custom connector in Claude (Customize → Connectors) with app.bennch.co.uk/mcp. It connects over MCP and covers all your projects. No keys to copy.
Ask Claude to start
Paste the start prompt from Bennch. Claude plans or maps the project with you; approve the cards, press Build, and it gets to work.
Works with the tools you already use
Claude Code
Works from your approved cards over MCP and reports what it did. Bennch runs no AI of its own.
Codex
Coming soonThe same cards and checks for OpenAI’s Codex: it builds what you approved and reports back to Bennch.
GitHub
Commits that name a card appear on it, with their size and your Actions result. If the AI says tests passed but CI failed, the card shows it.
GitLab
Coming soonCommits, pipelines and change size on each card, for projects on GitLab.
Sentry
Coming soonA new error after a release shows on the card that shipped it, and can become a proposed card for you to approve.
External integrations such as GitHub are read-only.
What Bennch enforces.
- Bennch never stores your codeYour code stays wherever you keep it: your computer, GitHub, GitLab or elsewhere. Bennch keeps your plan and what your AI reports: commits, tests and results.
- Only you approve and acceptYour AI’s connection has no tool to approve cards or mark work done. Those actions only exist in your signed-in account. If you answer a question in your AI chat, it’s saved in your words, labelled, and you can change it.
- Secrets are refusedValues that look like API keys or tokens are rejected on cards and in prompts.
- Approved work onlyBennch refuses to start a card you haven’t approved, or one that’s waiting on you or someone else. External integrations such as GitHub are read-only.
- Invite-only for nowEarly access is by invitation from the waitlist. You sign in with your email and password, plus an authenticator app if you turn it on.
Questions.
Which AI coding tools work with Bennch?
Claude Code works today, and Codex and other coding agents are coming. Your AI does the building; Bennch holds the plan, the decisions and the evidence.
Does Bennch run AI?
No. Everything the AI does happens in your own coding tool, under your own account.
Can Bennch stop the AI going off-plan?
Bennch blocks what goes through it: your AI can’t start an unapproved card, approve anything or mark work done. It can’t control what the AI writes in your files, so every commit is linked to a card, unusually large changes are flagged, and nothing counts as done until you’ve checked it.
Can my AI work through several cards?
Yes. Press Build, pick a board or part of the product, and choose one card or several in a row. Bennch shows exactly which cards it will take, in order. You can also run several AI sessions at once.