PeckBoard
Self-hosted multi-agent orchestration

The kanban board that works itself

PeckBoard spawns Claude Code agents to work your project board. Add a card describing a task; a worker session — one of your flock — picks it up, edits real files in your repository, and moves the card across the board until it reaches Done.

One binary · your machine · your code never leaves the nest

A project board with cards spread across Backlog, In Progress, Review, and Done columns
How it works

You write the cards. The flock ships them.

01

Lay out the cards

Write cards the way you'd brief a colleague. Set priorities and dependencies; blocked cards wait until the cards they depend on are done.

02

A worker swoops in

The orchestrator spawns a worker session — a Claude Code agent — in the project's folder. It consults the project's experts, edits files, and runs your tests.

03

Review, then Done

Watch every step live, chat with the worker mid-run, inspect a side-by-side diff of what changed, and let the card land in Done.

The toolkit

Everything your flock needs

Sessions

Chat with any session

Every agent run is a session you can open: the full transcript, each tool call, and the model's replies stream in live. Interrupt, redirect, or just ask what it's doing — workers, experts, and plain chats all speak the same way.

Core concepts →
Chat view of a session showing user messages, agent replies, and a collapsed tool call
Review

Review documents with an AI pass

Annotate any markdown document — suggest, mark wrong, ask to expand — and run a pass: a reviewer session revises the document, resolves each annotation with a note, and asks an inline question when it needs your call. Every revision is a version with a diff, and nothing touches the source until you apply. Link a GitHub pull request and its review comments come along.

The review flow →
The review screen with a markdown document on the left and annotations with resolution notes in a rail on the right
Experts

Experts that know every branch

Long-lived expert sessions each read one area of the repository and answer workers' questions; a question expert remembers your answers, and a PM expert records decisions. Workers ask before they guess.

How experts work →
Experts view listing expert sessions grouped by scope, with Question, Knowledge, and PM badges
Providers

Every provider, with budgets

Run agents on Claude, Grok, Kimi, Cursor, or your own Ollama servers. Sign in multiple accounts per provider, give each a spending budget, and watch real subscription plan usage — every model shows up in one picker.

Models & providers →
Settings page with provider on/off toggles, Claude plan usage, and two signed-in accounts with spend and budget badges
Plugins

Plugins in two clicks

A built-in registry installs WebAssembly plugins — diff viewer, SSH fleet tools, a public REST API — and MCP server templates. Nothing runs until you approve the exact hooks a plugin asks for; deny and it stays dormant.

Browse the plugins →
The plugin registry's Browse tab listing installable plugins and MCP server templates, with search and filters
Replay

Replay every test flight

Recorded test runs replay like a session-replay product: the cursor moves, clicks ripple, the network waterfall and console scroll alongside, and rage-clicks are called out. Playback skips inactivity so a two-minute run reviews in seconds.

The Playwright Tests plugin →
A recorded browser test replayed with a frame stage, cursor, event timeline with a rage click, and a network waterfall
Also in the box

The rest of the flock's kit

Usage dashboard

Token and cost rollups per model, project, and session — with trends over time — so you know where the spend goes before the bill says so.

Repeating tasks

A saved prompt on a schedule: every N minutes, daily, or weekly. Each firing starts a fresh session in the task's folder — nightly dependency checks, morning summaries.

Multi-user access

Accounts with per-user sessions: each person sees and streams their own conversations, and admin-only surfaces stay locked to admins.

Plans before code

An agent proposes a plan you can read, comment on, and revise; a wizard then turns the approved plan into cards the flock works through.

Reports that outlive sessions

Agents write markdown reports — research, audits, findings — that collect in a browsable library long after the session that wrote them is gone.

Every tool call, readable

Whichever provider runs the agent, the transcript reads the same: the real command line, the model's one-sentence reason, and a diff card for every file edit.

Zero infrastructure

One binary. Your machine. Your code.

The web interface, database, and TLS certificates all live inside a single binary. State is a SQLite file in one directory — a backup is a folder copy. No external services, no cloud account, and your source never leaves the machine the agents run on.

Configuration covers the flags, accounts, and where data lives.

# grab a release, then hatch it:
./peckboard
# prints the admin login and listens on :3344