Capabilities
Everything the board can actually do.
There are six groups. Orchestration and isolation are the parts other tools also have. Memory, cost accounting, cowork and customization are the parts they do not.
01 · Orchestration
Cards spawn sessions
Every card is a real Claude Code session, not a queued job or a templated prompt.
Lanes carry state
A card sits in backlog, in flight, review, or needs you, and that lane reflects the session's actual state rather than a label somebody dragged across the board.
Cross-project inbox
One view shows every card across every project, and you can filter it by what needs pickup, what is in flight, and what is waiting on review.
Projects and groups
You organise repositories into groups, and each one keeps its own checkout path and its own lanes, so a board covering eight client codebases stays legible.
02 · Isolation
Git worktrees, not sandboxes.
Every card gets a real checkout on a real branch. There is no container to configure, no image to keep current, and nothing to export when you want to finish something by hand.
| Property | Behaviour |
|---|---|
| checkout | One git worktree per card, created on demand. |
| branch | One branch per card. Nothing is committed to the default branch by a session. |
| object database | Shared with the origin repository, so worktrees are cheap to create and destroy. |
| stash | Shared across worktrees, which is why sessions are instructed to use WIP commits rather than bare stash. |
| cleanup | A worktree that produced nothing is removed. One that produced work stays until you merge or discard it. |
| takeover | Open the worktree in your own editor at any point. It is a normal directory on a normal branch. |
03 · Memory
Memory is the part nobody else ships.
Sessions end, but what they learned should not end with them. Memory is routed onto the prompt that needs it, indexed as it is written, and published to the team, so one person's hard-won knowledge is never trapped inside a single transcript.
Of the four parallel-agent tools we compared in August 2026, not one names persistent memory or team memory anywhere on its public pages.
The biggest win
The only agent board that tells you what the work cost.
Authoritative accounting
Usage is read from the per-turn transcript Claude Code actually sends, then attributed to individual tool calls. It is not an estimate.
Re-read detection
A session that already read a file is told so, with a token estimate, before it reads it again.
Cost watchdog
You see spend building while the session runs, instead of meeting it on the invoice at the end of the month.
Figure measured on one machine across 1,618 Claude Code sessions and 105,724 assistant turns, priced at published cache multipliers. It is a measurement, not a product benchmark.
05 · Cowork
Not all of the work is code.
A second surface for everything that surrounds shipping: drafting a reply, writing the proposal, researching a competitor, running a security pass. It uses the same board, the same memory and the same sessions, without pretending every task has to end in a diff.
Sources and drafts stay attached to the session, so the material a piece of work was based on is still there when somebody asks where a claim came from.
Each one starts a session with the right context already attached, so the work begins from something rather than from an empty prompt.
Write your own
A skill is just a directory with a markdown file in it. Add one, commit it, and every teammate has it on their next pull.
Your own workflows
Write slash commands that encode how your team actually works, rather than how a vendor assumed you would.
Rules that always fire
The harness enforces these rather than trusting the model to remember them. They cover guards, change logging, formatting and test runs.
Reviewers you choose
The diff decides who reads it, so a security, TypeScript, Laravel or tech-debt reviewer is dispatched based on what actually changed.
Per-repository knowledge
Project-level skills live inside the repository, right next to the code they describe.
Team and personal layers
Everyone runs the same shared team configuration, with a personal overlay on top. Nobody has to maintain their own fork.