PGAI AGENT KANBAN

Ship your software while you sleep.

public release: v3.5.0 · MIT

A requirements document goes in. A tested, tagged release of your software comes out. Repeat.
Runs on your VPS, your repository, your language model.

The kanban's operator window in a terminal. On the left, the board: bugs, priorities, requirements and agent tasks for three projects, one colour per project, with a command line under it. On the right, the workbench: an assistant answers 'what shipped overnight, and does anything need me?' with three shipped releases and one question waiting.
The operator's window: the board for three projects, a command line, and Claude Code or Codex in the workbench beside it. Select the image to see it full size.

How It Works

You write a requirements document. The kanban does the rest, one release at a time per repository, on purpose.

PLANA planner breaks your document into small tasks, each with acceptance criteria.
BUILDCoders build each task in its own git worktree and merge it back.
TESTA tester checks the result against your criteria, with your own test commands.
SHIPSquash, tag, push, release notes, changelog. Bugs it found are fixed next round.

The Team You Get

PMturns your requirements into tasks
CODERbuilds each task in a worktree
WRITERdocs, guides, prose deliverables
TESTERverifies against your criteria and files every bug it finds
CMbranches, tags, releases; the only one allowed to push to your origin
POon request: turns a one-page brief into a reviewed requirements document
OVERWATCHthe watchdog: resolves mechanical blocks and stops runaways, with no model in the loop

The Road So Far

The releases that mattered most. Each one was built by the one before it.

v1.0.0Jul 2026

The core loop: a requirements document in, a tagged release out. Role agents, cron, a terminal dashboard. Builds itself from day one.

v1.23.0Jul 2026

Workflows become plugins (release, document, testing-only). Operator REST API. OVERWATCH and the PO intake agent arrive.

v2.4.0Aug 2026

Resilient autonomy: findings become bugs for the next loop, not reasons to stop. Blocks resolve themselves.

v2.41.0Oct 2026

Walk away from it: why.sh, an API twin for every operator verb, ship by default, small bugs bundled.

v3.0.0Oct 2026

It stops stopping: operator inbox, OVERWATCH decisions and runaway detection, provider-error retries, your own test runners and release hooks.

v3.5.0Nov 2026

The operator's workbench: Claude Code or Codex beside the board, with skills, so you instruct instead of typing commands. Time limits on hooks. Reset keeps work.

FUTURE PLANS

The code graph: agents look your code up in a map of the repository instead of searching it.

Six of eight public releases, and what is planned. Every release and its notes are in the repository.

What v3.5 Gives You

In production. Not planned.

  • Ships by default, with a release status you can read; never a silent stall
  • Questions go to an inbox; the chain keeps moving
  • OVERWATCH fixes mechanical blocks and stops runaways
  • Your test commands in every task's acceptance criteria
  • Release hooks for your own release steps, time-limited
  • A cockpit with Claude Code or Codex beside the board
  • Skills, so that assistant already knows the kanban
  • Reset keeps committed work
  • Containers that start with a proper init and stay healthy
  • One line answers "why isn't it moving?", with the fix attached

Built the Right Way

The size and discipline behind the thing you would be trusting, at this release.

18,288unit tests, green on every release
44checks on every merge
53REST routes, 35 operator verbs
400+releases of itself, built by itself
Lines of code at this release. Blank lines and comments not counted.
BASHPYTHONTOTAL
PRODUCTION60,000132,000192,000
UNIT TESTS—217,000217,000
INTEGRATION TESTS—49,00049,000
FIXTURES AND OTHER TESTS1,00027,00028,000
TOTAL61,000425,000486,000

More test code than product code: about 1.5 lines of test for every line that ships.

A long bug list is the standard being enforced, not the product breaking. A bug here is anything the tester can show is wrong, and that includes a misspelled message or a stale line of documentation. Most are fixed in a few lines.

What It Costs

~$200 / yearHardware: a VPS. The hardware is solved.
~$300 / projectOn an API key, an average working day at list prices. It varies with the model you choose and how much work you queue.
$200 / month planOne project shipping steadily. Two shipping hard hit the weekly cap; the extra is billed at API rates.

Measured from real token records, not estimated.

Run It

# your VPS, your repository, your choice of language model
$ git clone https://github.com/purdygood-ai/pgai-agent-kanban.git
$ cd pgai-agent-kanban
$ pip install -r requirements.txt
$ ./install.sh
# then walk the ten-minute demo: a requirements document in, a tagged release out
$ less team/demos/helloworld-demo/README.md

Four complete demos ship in the repository with every input file included: hello world, an arcade game built release by release, a children's story through the document workflow, and an audit of any repository. Everything is open: docs, role files, workflow definitions.

Future Plans

The next features planned. Nothing is claimed until it ships.

  1. A code graph for the agentsWhere your code is, what calls it, what it calls. Lookups instead of scans: fewer missed callers and fewer tokens per task, measured. Built and running in our own shop.
  2. The walk-away runFifty or more requirements on a new project with no human interaction. The bar this was built for. Not yet claimed.
  3. More providers, on demonstrated needOne full release under a provider before it is claimed.