Confab is a desktop app that connects to your board and builds the stories on it. Dispatch a card and your own Claude builds it on your machine, pushes a branch to your own git host, and opens a pull request with the cost, tokens, and time reported back to the card. Nothing merges without you.
// bring your own Claude · runs on your machine
#218 moved to Queued: “Return a Retry-After header with every 429.”#218. Building on branch confab/218-retry-after, right here on your machine.$0.84 · 41.2k tokens · 7m 05s, reported to the card.Confab is the execution arm of your planning board. You decide what to build, it does the building, and you review everything that ships.
Plan the way you already do: cards on a board. When a card is ready, drop it in Queued. Confab picks it up with the full context behind it: the story, the acceptance criteria, and the architecture decisions your team already made.
It builds against your real codebase, in a workspace bound to your repo folder. No copy-pasting tickets into a prompt.
#218 “Return Retry-After with every 429” moved to Queued#218 with 4 acceptance criteria and 2 linked architecture decisions. Building.confab/218-retry-after pushed. Pull request open. Review →Confab builds work units sized for a human to read: one card, one branch, one pull request. It self-reviews the diff against the card's acceptance criteria before handing it over, and then it stops. Nothing merges itself.
If a card is too big to review comfortably, it's too big to dispatch. Small cards are how you stay in charge of an AI that codes.
+64 −12 across 3 filesWhen the pull request opens, the card gets the receipts: what the build cost in dollars, how many tokens it used, and how long it took. No credit meters, no mystery math.
Over time your board becomes its own cost model. You can see what a story of a given size runs, and plan with real numbers instead of vibes.
$0.84 · 41.2k tokens · 7m 05s$1.12 · 58.9k tokens · 9m 41s$0.31 · 17.4k tokens · 3m 58sConfab runs your own Claude under your subscription or API key, on your machine, and pushes to the git host you already run: self-hosted Forgejo, GitHub, and friends. Nothing about your setup has to change.
And when you want to steer, each workspace has its own chat with the agent. Course-correcting a build is a message, not a meeting.
#218, reuse the middleware pattern from #204 instead of a new helper.Not another chat bot. A desktop app that ships the cards on your board, on your own machine, one reviewable pull request at a time.
Mark a card ready and Confab picks it up with its full context: the story, the acceptance criteria, and your architecture decisions. Your backlog becomes a build queue.
One card, one branch, one pull request, sized so a person actually reads the diff. By default nothing touches main and nothing merges itself; the PR waits for you.
A per-workspace bridge runs your own Claude locally against one repo folder. One window per project, nothing bleeds between them, and your code never leaves your computer.
Use your existing Claude subscription or API key. Your quota, your bill, no markup, no proxy, no reselling your tokens. Compliant with provider terms by design.
Every card reports what it cost: dollars, tokens, and wall-clock time, posted back to the card when the pull request opens. You always know what your board is spending.
Branches and pull requests land on the git host you already run: self-hosted Forgejo, GitHub, and friends. Your code never has to move anywhere new.
Real cards, real pull requests. Not one person babysitting a terminal: a plan turning into reviewed changes.
Queue three small cards before lunch, review three pull requests after. The backlog burns down while you stay on the work only you can do. Every change waits for your review.
Write the bug up as a card the way you already do. Confab reproduces it against the real code, fixes it on a branch, adds a test, and opens a pull request with the cost on the card.
Telemetry lands on each card: dollars, tokens, wall-clock. Your board becomes a real cost model per story size, and the branch-and-PR default means the audit trail is your git history.
The card's acceptance criteria are the spec. Confab builds against them and self-reviews the diff criterion by criterion before the pull request opens. Good cards in, checkable PRs out.
Make test debt visible on the board instead of a wiki page. A coverage card gets built like any other story: Confab reads the source, writes the missing tests, and opens a PR you can inspect.
Confab pushes branches and opens pull requests on your own git host: self-hosted Forgejo, GitHub, and friends. The build runs on your machine, so your code never transits our servers.
Cloud agents run your work on someone else's machine. A terminal session needs you watching it. Confab connects your plan to your own AI and hands you pull requests.
| Category | Confab | Cloud coding agents | Claude Code in a terminal | Chat coding bots |
|---|---|---|---|---|
| Where the work runs | Your machine | Vendor cloud, code uploaded | Your machine | Vendor cloud |
| Driven by your plan | Board cards in, PRs out | One prompt or ticket at a time | You type each task | One chat thread at a time |
| Work unit | Review-sized story card | Often whole-feature, hard to review | Whatever you ask for | Whatever you ask for |
| Review gate | Branch + PR by default, nothing merges itself | Varies by vendor and config | You are the gate, live | Varies |
| Works while you do something else | Yes, until the review | ✓ | No, it's interactive | Per message |
| Who owns the AI bill | Your subscription or keyConfab runs your own Claude under your subscription or API key. Your provider bills you directly at standard rates. Confab never marks up your AI costs. | Credits, usually with markup | Your subscription | Varies, often credits |
| Cost visibility | Dollars, tokens, time on every card | A credit meter | Per session, if you look | ✗ |
| Your own git host | Forgejo, GitHub, and friends | Often GitHub-only | ✓ | Varies |
Cloud agents typically buy model capacity wholesale and sell it back as opaque credits. Real-world reports show $200-400/month in sustained use, even when the headline price says $50. Confab runs your own Claude, so there is nothing to resell.
Some agent platforms charge per seat, per parallel run, or both. Five people running builds means five meters. Confab charges a flat team price, and the AI bill stays whatever your provider already charges you.
Providers keep changing how they price and enforce third-party tool access, and some have suspended accounts without warning. Confab runs your Claude under your own login, the pattern providers actually sanction.
An agent that loops overnight on real money is a budgeting problem you find out about later. Confab reports cost, tokens, and wall-clock time back to every card, so the spend is visible per story, as it happens.
Confab keeps it simple. Bring your own Claude subscription or API key and pay Anthropic directly at standard rates. We charge a flat team price, and every card shows what it cost to build. No markup, no credits, no surprises.
Solo? $9.99/mo. Team of 5? $49/mo.
The same team on typical per-seat pricing: $75–400/mo
Pay annually and save 20%. You bring the AI. We don't mark up your provider costs.
No servers. No terminal. No IT ticket.
Download Confab for Mac or Windows and open it. Everything runs locally; there is nothing to deploy.
Point a workspace at a repo folder on your machine and connect your board. Your own Claude runs inside it, against your code.
Drop a card into Queued. Confab builds it on a branch, self-reviews it against the acceptance criteria, opens a pull request, and posts cost, tokens, and time back to the card.
Confab runs on your own machine and touches your code. We take that seriously. Every claim here is backed by an architectural decision, not a marketing checkbox.
Start free. Bring your own AI. We don't mark up your provider costs.
See what Confab can do
Solo power user
5 users included
20 users included
Confab turns the cards on your board into reviewed pull requests, built on your own machine with your own Claude, with the cost reported back to every card. You stay in control: every change is a branch, every branch is a pull request, and nothing lands until you say so. That's Confab.
// bring your own Claude · runs locally · your code stays yours