Run your factory from the chat
Ask Claude or ChatGPT for work in plain words. The manager researches, asks questions, writes the stories and asks "proceed?". Then every station runs and you approve your own stations in the chat.
Ask Claude or ChatGPT to list the factories you can see, read a board or check which station needs your decision. Cloud work runs in CodeBook. To execute local stations in your checkout, use the Claude Code plugin. Open the same story on the CodeBook website to see its current state.
All of it goes through one connector, https://app.codebook.work/mcp. You sign in once with your Bookbag account, and the connector can do exactly what you can do on app.codebook.work, never more. A factory you cannot see does not exist for it. A station you may not approve stays refused.
Where you can do it
Claude chat means claude.ai and the Claude desktop and mobile apps.
| Claude Code (plugin) | Claude chat | ChatGPT | |
|---|---|---|---|
| Set it up | Connect Claude | Connect Claude | Connect ChatGPT |
| Start work in plain words | Yes | Yes | Yes |
| Approve or send back your stations | Yes | Yes | Yes |
| Run here: the stations run in your checkout, on your Claude plan | Yes | No | No |
| Run in CodeBook: CodeBook's agents run the stations | Yes | Yes | Yes |
| Read a cloud work-order report | /factory report <factory> <WO-7> |
Ask for the order by reference | Name the factory and work order, for example “Show WO-7 in Payments.” |
| Local proof, automatic transcript capture and checkout handling | Plugin features | Cloud connector only | Cloud connector only |
The Factory pane and /factory |
Yes | No | No |
| Asks you when a story reaches your station | Yes, by itself | When you ask how the story is going | When you ask how the story is going |
| The factory board and a story drawn as cards in the chat | No (the pane shows them) | Yes | Yes |
Claude Code runs local stations and benchmark workers on your computer. A cloud chat can follow a story assigned to a local worker and request proof for that story. Before approving a waiting station, ask “Where will this station run?” The story and board show its effective destination, including a personal setting that keeps a particular station in the cloud. Request and read repository proof in ChatGPT
Read a work order from chat
Try “Read work order WO-7 in Payments. Which milestone is blocked, what evidence was recorded, how much budget remains, and does someone need to decide a production hold?” A read can supply only what that factory has configured and what you may see. In Claude Code, /factory report Payments WO-7 performs the same lookup without starting a model. Work orders and overnight reporting
From a request to finished work
Say what you want: "Let customers export invoices as CSV", or "Tidy up the stale branches in the payments repository". In Claude Code you can also type it with /factory run.
In Claude chat and ChatGPT, and with Run in CodeBook, the chat starts the story you describe, or proposes several for you to confirm. From there, CodeBook's own manager and agents take it on, exactly as they do for a story started on the website.
With Run here, Claude becomes the factory's manager in your Claude Code session. It follows these steps, in this order, as fixed rules:
- Research, read only. The manager reads the factory's repositories, branches, pull requests and board to see what the request touches. Nothing is changed.
- Questions. It tells you what it found and asks what the stories need, one question at a time: the scope, what "done" looks like, and what must not change.
- The stories. The factory's story-writing agent, for example its business analyst, writes them. Each story has acceptance criteria and a risk. Stories that are already on the factory are left out, and the manager tells you which ones.
- "Proceed?" You see each story and a link that opens the proposal on app.codebook.work. You also see who merges the pull requests at Done, in CodeBook's own words. Nothing is created until you say yes.
- The stations. Every story moves station by station, and the chat tells you who is working, what moved and when a pull request opens.
The rule underneath is CodeBook's own: no story, no work. A request for housekeeping goes through the same steps. The manager offers you a story to approve, not a command to run.
Run here, or run in CodeBook
Run here is the default in Claude Code. For each agent station, Claude starts a sub-agent that becomes that station's agent and uses that agent's own skill from CodeBook. The sub-agent claims the run, works in your checkout, runs the tests and opens a draft pull request with your own git and gh. It writes its notes and artifacts on the card. A QA station runs the repository's tests on your computer. A crew station is led by the session, and each crew member gets its own sub-agent.
- It runs on your Claude plan. The card records each run as "own plan, not billed": CodeBook does not bill it.
- CodeBook picks each station's model. A station set to a Claude model runs on that model. A station that writes code runs on your session's model. Every other station (the story writer, QA, review, planning) runs on Sonnet, because those stations read, test and write text. Name a model for a station if you want a different one.
- A station that asks for another provider's model (GPT-5, say) runs here on the Claude model CodeBook picks instead, and the card notes which model it ran on. To run it on its own model, keep that station in CodeBook's cloud, for your stories only: the manager offers it, or switch it on in Factory settings › Your computer. The rest of the story still runs here.
- Stories run side by side. When several stories each have a station waiting, up to three workers start together, each code worker in its own git worktree. One story's stations still go one after another.
- Only the factory's repositories change. While a run is held, the plugin refuses a
git pushorgh pr createto any repository the run's brief does not list. Open Claude Code in the repository the factory works on. - Claude never merges. It never pushes to a default branch and never force-pushes.
- A factory can say no. If a factory's policy keeps work inside CodeBook, it refuses Run here, and Claude offers to run the story in CodeBook instead.
Your code, on your computer
Run here works in a folder you choose for each of the factory's repositories: an existing checkout, or a folder to clone it into. Set it in Factory settings › Your computer, or tell the manager in the chat. Before any work, research included, the manager checks the folder is there, is a checkout of that repository and that git can reach it. If not, setting it up is the first thing it does, after your yes. It never deletes or overwrites a folder. Only you see these settings. The same page can make your stories run here by default.
Run in CodeBook works like starting a story on the website: CodeBook's own agents run the stations on its servers. You still follow the story and decide your stations from the chat. In Claude Code, switch with /factory here and /factory cloud, or with the two buttons in the Factory pane.
Your stations, in the chat
When a story reaches a person's station that is yours, the chat shows what was done, QA's verdict and the pull request. Then it tells you what approving will do and asks: approve, or send back and why? Your answer does exactly what the buttons on app.codebook.work do.
- Approving can merge. A factory may merge a story's pull request when its station is approved, even if the pull request is a draft. CodeBook gives the chat one sentence that says what approving will do, and in Claude Code the manager always shows you that sentence before it asks. In any chat, you can ask "what happens if I approve?". On app.codebook.work the same sentence is on the open card, next to Approve and Send back, and a card waiting for you shows a Review button on the board.
- Reviews and questions wait for you the same way. That covers a pass/fail review, a question from the manager, or an agent that asked for a person in the middle of a run.
- In Claude Code you don't have to watch. When a followed story reaches you, the plugin wakes the session, and Claude asks you there and then. In Claude chat and ChatGPT, ask how the story is going. The story card shows the station that waits for you.
- At Done, the manager sums up each story: its pull requests, its runs, its test reports and its link. If nobody at CodeBook will merge a pull request, it tells you plainly that the merge is yours to do on GitHub.
The board and the rest of the factory
The manager can do what you do on the factory's screens, with your own access:
- Cards. Move a card to another station, edit it, or delete it. A delete only happens after you say "proceed?" to that card, and CodeBook refuses it unless the card's exact title is given.
- Questions. A question for you goes on the card, where everyone sees it, and the chat asks you too, one at a time. Your answer goes on the card and the station carries on.
- The other screens. Work orders (plan, replan, edit), the optimizer's recommendations (see what one would change, then apply it), test runs, operations (pause and resume the factory, alerts), outcomes (report that shipped work broke production, or was deployed) and the dashboard. The manager reads them freely. Anything that spends, reaches outside CodeBook, changes the factory or pauses it waits for your yes. The kill switch, autonomy and emergency revoke stay on the website.
The manager in CodeBook
CodeBook's own manager, behind the board's Talk to … button (for example Talk to Harper), follows the same rules and reads the same screens. When it wants to change something there, such as pausing the factory or starting a test run, it changes nothing itself: it puts a Confirm button in the conversation, and the change happens when you click it, as you.
Switch factories
Name the factory in your request ("in the Payments factory, …") and the chat uses that one. Ask "which factories can I see?" to get the list. In Claude Code, /factory list shows them, ← Other factories in the pane picks another one, and /factory follow follows a story that is already on any board you can read.
Add agents from the marketplace
If a factory has no agent to write its stories and you may change the factory, the manager offers to add one. It looks in this order:
- an agent your workspace already has;
- one from the marketplace, your organization's own section first;
- a new agent whose skill it writes with you. It can be published to your organization's section only, never to the public marketplace.
The manager asks you before each change: installing an agent, creating one, publishing one, and putting an agent on a station. The first change can open a Bookbag sign-in that asks for one more permission, "Change your factories' agents and stations". You can also ask for these steps directly, for example "find a QA agent in the marketplace for this factory". A person who may only read the factory is refused, as on the website.
Scorers and benchmarks, when they help
If you want to measure the factory's work, ask for it. In Claude Code, the manager may also suggest it when it would help, with the reason. For example: "QA has 20 finished runs: shall I check whether Sonnet does as well as Opus?". These steps are never required for a story. Nothing here runs by itself, and the manager asks before every change.
-
Scorers. Scorers judge how good the factory's finished work is: "are the tests any good?", "is this over-engineered?". List the factory's scorers, and add one from CodeBook's ready-made questions or write your own. A new scorer only judges runs that finish after it is made.
-
Benchmarks. A benchmark replays a station's past work on other models, so you can see which model is good enough for the least. To create one, pick the factory's finished runs and freeze them; 10 to 20 runs is plenty for a first look. Then run it in either of two ways, or both:
- In CodeBook, on the models your organization runs. The replays run on the factory's budget.
- Here, in Claude Code, on Haiku, Sonnet or Opus. A sub-agent of each model answers the frozen tasks on your own Claude plan. These answers are text only: no repository, no tests and no pull request.
Either way, CodeBook's judge grades every answer in the same way, with the factory's scorers. Then ask for the report: quality, coverage, pass rate, cost and speed for each model.
-
Answered here costs "own plan". In the report, a run answered here shows its cost as own plan, not a price, and it is kept off the cost-and-quality chart. Its result tells you which Claude model to use for that station when you Run here. It can't be proposed as the model CodeBook runs for a station. To change that model, replay the model in CodeBook and propose the change; a second person approves it on app.codebook.work.
Adding a scorer and creating or running a benchmark change the factory, so they need the same permission as adding agents. A factory whose policy keeps work inside CodeBook refuses benchmark runs answered here, as it refuses Run here.
What is recorded
Every comment, run, artifact and pull request lands on the story's card in CodeBook. Comments are signed with the agent's name and where they came from ("Agent (role) via your Claude Code"). A move made from the chat names you and the app you used. Your code stays in your checkout, except what you push. The plugin has no servers of its own and no analytics: see Privacy.