New in Commission

Hand Claude Code the whole job.

Give a coordinator a goal, like a list of tickets or a feature. It breaks the goal into pieces, runs each one in its own Claude Code session, worktree and branch across your machines, and works out what can run side by side. You answer its questions and approve the merges, from your desk or your phone. Commission keeps the record of what shipped.

A coordinator in Commission: the batch bar reads Checkout fixes, SHOP-114 to 125, 12 tickets, 10 released in v4.2.0, 1 to merge, 1 working, beside the Work tab grouped into Needs you, Working and Released
The same batch on iPhone: the Work tab, with the merge that waits for you first
One real day

It opened as “fix three tickets, a PR each.”

Related bugs turned up as the sessions worked, so the coordinator added them and renamed the batch. By the evening, ten tickets had shipped in a release and an iPhone build. The eleventh was merged and waiting for the next release, and the twelfth, a redesign, carried on as its own piece of work. The report said so, under Not shipped yet.

  • 1goal
  • 12tickets
  • 11pull requests
  • 1 + 1release and iPhone build
  • 1day

The numbers are from that batch. The screenshots on this page come from a demo hub with made-up names.

How it works.

Start with a goal

New session › Coordinate. Say what you want done, the way you'd put it to a colleague: three tickets, a PR each, or a whole feature. Pick the machine and the folder, and Claude suggests a title. The coordinator writes each piece a brief, with the goal, the ticket, when it's done and what it may not touch, and starts a session for it in its own worktree and branch.

  • On any of your machines Settings › Coordinators allows, never the others
  • Up to six at once by default, twelve at most
  • Want to OK each start first? One setting puts them in Needs you
New session in Commission with Coordinate chosen: the folder ~/code/storefront, the title Checkout fixes and the goal Fix SHOP-114, SHOP-115 and SHOP-116, a PR each
The coordinator's reply on iPhone: where the batch stands, which ticket waits for a merge and which two sessions work on the redesign

It works out the order

The coordinator is Claude, reading the tickets and the code. It runs side by side what can, stacks what has to wait, and says why: “The build waits on the design's tokens, so I stacked it.” When related work turns up, it adds the ticket to the batch and renames it, so the list always says what the batch is really doing.

  • Scope changes are on the record: who added what, and the old title
  • Sessions on one piece of work, like a design and a build, share its ticket
  • Steer it any time, in plain words, like any Claude Code session
The batch's Timeline tab: started at 8:01, three sessions started, more tickets added at 9:38, the batch renamed at 9:40, a rule allowed for the batch, a question answered
The same timeline on iPhone, with filters for outward changes, decisions and sessions

Watch it from anywhere

The batch bar sits under the coordinator's header: its title, one segment per ticket coloured by where it stands, and what waits for you. The Work tab groups the tickets by what each needs next, with its pull request, CI and the sessions on it. The same bar is a header chip on your phone.

  • “12 tickets · 10 released · 1 to merge · 1 working”, the same on every device
  • Open a ticket for its branch, pull request, release and its own history
  • Each session stays a normal session you can open, read and steer
A ticket opened from the Work tab: SHOP-125 Checkout redesign, In Progress on branch shop-125, no PR yet, two sessions on it and its latest changes
The same batch on iPhone: the Work tab, with the merge that waits for you first

You stay in charge

Anything that needs a person comes to Needs you, as one kind of card: where it came from, what it is, the coordinator's pick and why, then your buttons. Its questions, permissions it won't decide alone, and ship plans as checklists of tickets. Commission runs the steps you approve; a merge runs only when you tap Merge. Never a terminal.

  • The coordinator answers routine asks itself, and passes on what it shouldn't decide
  • A rule you allow for the batch can be taken back from its timeline row
  • Releasing stays yours. When you tag one, Commission finds it
The Needs you tab: a ship plan for SHOP-124 with the PR opened, the ticket moved to In Review and a Merge #222 button, then the coordinator's question about releasing v4.2.1, with its pick and why
Needs you on iPhone: the same ship plan, with Merge #222 one tap away

A record of what shipped

Every change lands on the batch's timeline, with who did it: you, the coordinator, a session or Commission. A pull request opened in a terminal, a merge on GitHub or a release you tagged by hand still lands on its ticket, marked as found. When the batch finishes, Commission writes the report from those facts, not from Claude's memory: what shipped, in which release and build, and what hasn't shipped yet, with its next step.

  • Totals: tickets done, pull requests merged, releases, lines changed
  • Not shipped yet, with a button for each next step
  • Copy the report as Markdown for Linear or Slack
The batch report: finished Monday 9 November at 18:40, 11 tickets done, 11 PRs merged, 1 release v4.2.0, then Not shipped yet with SHOP-124 merged and SHOP-125 still to do, and a table of every ticket with its PR and release
The same report on iPhone

Every batch in one place

The Batches page lists every batch, newest first: what needs you, what's running and what finished, each with its progress and counts. A running batch opens on its Work tab; a finished one opens its report. Search by title, goal or ticket to find the batch that shipped it.

  • In the sidebar beside Activity, and in ⌘P
  • On the phone, a Batches segment at the top of Sessions
  • Filter by space, to keep each job and client apart
The Batches page: Checkout fixes needs you, Shadow artifacts is running, Search filters and sort finished, each with its progress strip and counts
Batches on iPhone, grouped into Needs you, Running and Finished

How it works underneath.

There's no new kind of agent and no hidden scheduler. The coordinator is a Claude Code session, and so is every piece of work. Commission's hub, on one of your machines, carries what they say to each other and keeps the record.

You talk to the coordinator, a Claude Code session: you send your goal, steers and answers, and it replies with its picks. The coordinator makes coord calls, by CLI or MCP, to Commission's hub, and the hub brings it events, grouped. The hub carries briefs, steers, answers and decisions to worker sessions on Studio, Laptop and Server, each in its own worktree, and brings back their asks, permission requests and reports. What only you decide, such as risky permissions, ship plans and merges, goes from the hub straight to your Needs you. Workers never push, merge or move tickets themselves.
Every message goes through the hub, which delivers it and records it on the batch's timeline.

The coordinator is a Claude Code session

You talk to it the way you talk to any session: type, attach, steer, stop. What makes it a coordinator is a set of tools Commission gives it, to start sessions, answer them and propose what to ship. Each call shows in its conversation as one line, like “Answered SHOP-125 build”, and the batch bar above keeps count.

  • It wakes when a worker asks, reports or fails, with the news grouped into one turn
  • Any session can become one: it asks, and you approve it once
  • Its workers can't start sessions of their own, so a batch never nests
The coordinator's conversation under its batch bar: two rows from the workers, SHOP-125's build asking whether to wait for the design's spacing tokens and SHOP-124 reporting done, then the coordinator's actions, Answered SHOP-125 build: wait for the design's tokens and Proposed a ship plan for SHOP-124, and its reply. Beside it, Needs you holds the ship plan with Merge #222 and the coordinator's own question, with its pick
The same coordinator on iPhone, with the batch chip in its header

Sessions talk through the coordinator

Workers don't chat with each other or with you. Each sends its coordinator short typed messages, and the coordinator answers them, or passes on what it shouldn't decide. The hub delivers every message, wakes whoever it's for, and records it.

  • Everyday permissions, like edits in its own worktree or running the tests, the coordinator can allow
  • Anything outward or risky, like a push, comes straight to you
  • A request the coordinator leaves for two minutes comes to you too
  • When Claude asks a question the usual way, it still becomes an ask
From a worker
Ask
A question, with options
Permission
A command or an edit it wants to run
Report
Done, blocked or failed, with its commits and tests
From the coordinator
Brief
The goal, the ticket, when it's done, what it may not touch
Steer
Change course, carry on, or stop
Answer
To an ask, from the coordinator or from you
Decision
Allow or deny, once or for the batch

Information, not orders

A Claude Code session treats what another session sends as information, not as orders. A plain message from one session to another arrives marked “Another session sent this, not the user”, and Claude reads it that way. That's a safety feature: a session that has gone off course can't talk another one into something you never asked for.

A coordinator's briefs, steers and answers count for more only because you said so when you started the batch. Commission tells each worker that when it starts, and the hub, not the model, marks whether an answer came from the coordinator or from you. Even then it covers the batch's work only. It can't widen a worker's brief or loosen its limits, and no worker pushes, merges or moves a ticket. Commission runs those steps when you approve them.

The same calls from anywhere

Every coordinator action is a commission coord command with a matching MCP tool: coord answer is mcp__fleet__coord_answer. The hub checks each call, for the batch, its space, its cap and who may act, so both behave the same.

Every Claude Code session Commission runs can read the fleet the same way: list sessions, read what another one said, search them, and send one a message. So a session checks who else is in a checkout before it changes anything.

  • From inside a session, it sees only what that session's space may see
  • Approving another session's request is off unless you turn it on
# In the coordinator
commission coord status --work
commission coord answer a_19 --choice Wait
commission coord steer 3f2a91c0 stop

# In a worker
commission coord ask "Wait for the design's tokens?" \
  --option Wait --option "Build on the old ones"
commission coord report done "Estimate shows in the cart"

# In any session
commission session list --here
commission session tail 3f2a91c0 --turns 3
commission search "spacing tokens"
commission session send 3f2a91c0 "The tokens are merged."

Good to know.

  • It's Claude Code underneath. Every piece is an ordinary session with your own Claude plan, skills and MCP servers. A batch of twelve uses about what twelve sessions would.
  • Claude plans, Commission keeps count. The coordinator decides the order. Commission tracks the tickets, pull requests, merges and releases, and runs the steps you approve.
  • Facts need connections. Pull requests and CI need GitHub on the repository's machine; ticket states need Linear or Jira. What it can't read shows as unknown, with the reason.
  • Releases are found by tag. Tags like v4.2.0 by default, per repository if you like. A repository that releases without tags shows “merged” until the coordinator records the release.
  • Off until you start one. Coordinators are sessions you start on purpose. Plain sessions work exactly as before.
  • One person's hub. Batches run on your own machines and hub. A version for teams is in the works.

Questions.

Does a batch use more of my Claude plan?

Yes, in proportion. Each piece of work is an ordinary Claude Code session on your own plan, and the coordinator is one more. A batch of twelve tickets uses about what twelve sessions would. Settings › Coordinators caps how many run at once: six by default, at most twelve.

Who decides the order, Commission or Claude?

Claude. The coordinator is a Claude Code session that reads the tickets and the code, and decides what can run side by side and what has to wait. Commission keeps the list of work, starts the sessions it asks for, runs the ship steps you approve and records what happened. It isn't a scheduler with dates.

Can it merge or release on its own?

No. A coordinator proposes a ship plan; pushes, pull requests and ticket moves run when you approve it, and a merge runs only when you tap Merge on that card. Releasing stays yours: when you tag a release, Commission finds the tag and marks what went out in it.

What do I need connected?

Claude Code on each machine, as for any session. For pull requests, CI and merges, connect GitHub on the machine that holds the repository; for ticket states, Linear or Jira. Anything Commission can't read shows as unknown, with the reason.

Does it see work done outside the batch?

Yes. A pull request opened in a terminal, a merge on GitHub or a release you tagged by hand lands on the same ticket, marked as found by Commission.

Can one session tell another what to do?

Not on its own say. A session treats what another session sends as information, not orders. A coordinator's briefs and answers count only in a batch you started, only for that batch's work, and they never let a worker push, merge or move a ticket. How it works underneath.

Can I start one from my phone?

Yes. New session › Coordinate works the same on the Mac app, in a browser and on the iPhone app, and so do the batch bar, Needs you, the timeline and the report.

Set up your hub in about ten minutes.

Free while Commission is in beta. No account to create.