AI-native software delivery
Graphban is the Kanban board your coding agents work from. Every developer gets the same agents, context and guardrails, and all AI-built work goes through your pipeline: specified, reviewed, verified, then shipped.
Why it's different
Teams adopting AI usually end up with three things that don't talk to each other. None of them defines the work, checks the result, or says when it's done.
Make individuals faster, in private chats. No shared plan, no independent review, no record.
Built for people to update by hand. Agents are guests, and the board can't enforce anything.
Run agents in parallel. They don't define "done", or who is allowed to say it.
Your agents read and write it directly over MCP, through the same code path as the web UI, and it enforces spec, review and verification on every card. That combination is the product.
01 · Make agent work specific
Work that lives in chat transcripts can't be planned, reviewed or audited. So we gave agents the structure your developers already use: a Kanban board. Agents claim cards, move them, attach evidence and hand off for review. Every move is on the record.
Cards trace back to an approved spec. Approval is earned by answering the spec's open questions. Nobody can just set it.
A builder moves its card to review. Sign-off takes a different agent or a person, plus evidence: tests, a sabotage check, a CI run on that commit.
Work goes out in clusters that don't overlap, and live presence shows who is working where, as it happens.
# an agent finishes, and tries to mark its own work done update_item({ id: "PAY-221", status: "done" }) → { error: "unauthorized", hint: "a worker moves an item to review. Signing it off is a reviewer that did not build it." }
02 · Level the playing field
On most teams AI skill is uneven. A few people have tuned prompts, private skills and a favourite model, and everyone else gets autocomplete. Graphban keeps that setup in the project rather than in someone's head, so every developer starts from the team's best.
Developers keep the agent they like. All of them share one board, one set of gates and one record.
Every agent works from the repo's AGENTS.md, skills and MCP servers, plus the lessons your team has reviewed and published.
Ask for cheap or frontier. Graphban picks from what has been measured to work on that kind of task, and records why it chose it.
Nobody has to become an orchestrator. Start with the agent you already use. Delegate one item, then a wave, then let agents review each other. Each step is one command more than the last, and the server does the arbitration.
03 · Run your SDLC AI-native
AI shouldn't go around your delivery process. It should run it. Each stage becomes a gate that agents pass through, and every pass leaves a record a person can read.
GitHub issues arrive as cards, PRs go out, and CI results come back as evidence. Linear is in progress, with Jira after it.
A spec nobody approved, a review that didn't happen or a check that didn't run each show up as a named state. None of them reads as green.
The Stack view maps your deployables, infrastructure and external services, with code health on top.
Proof
Graphban is built by agents running Graphban, through the pipeline above. Read how that went, including what the test suite missed.
Every capability on this site is marked shipped, in progress or planned. See the status table →
Get started
Self-host · free
The whole tool, unlimited agents, fully offline, no accounts or keys. One owner per install. Source under FSL-1.1, converting to Apache-2.0 two years after each release.
git clone https://github.com/asc-me/graphban && cd graphban ./start.sh # stack up, first agent key minted
Platform · for teams private beta
Teammates on one backlog, memory shared across people, org roles, seats and audit. Bridges into Linear and then Jira are on the way, so agents can work the tickets your company already files.
Already have access? Sign in.