Answers from our docs and articlesesc to close
Try asking
Create an accountSign up

Build an app with Claude Code: ib-apps dev in one tab, Claude in the other

Run ib-apps dev in one terminal tab and Claude Code in another, with the app's MCP server connected: prompts for building, debugging and data.

Key takeaways

  • ib-apps dev keeps a folder in sync with the app's development environment and runs its dev server, so every file Claude Code saves is live at the app's address a few seconds later.
  • Each environment's MCP server gives Claude the app's real models, records, actions and event log as tools, so it works from the data instead of guessing at it.
  • Schema tools such as create_model write on the server, and ib-apps dev pulls the result into your folder, so the checkout and the server never drift apart.
  • Connect development to build, and production to ask questions about live data. A member's session gets only the reading tools on production.

Claude Code is good at reading and changing code. On its own it knows nothing about your app's data: what a record holds, which action failed this morning, how many shops have installed it. An Inference Bureau app gives it both. The folder on your computer holds the code, and the environment's MCP server holds everything else.

The setup is two terminal tabs. One runs ib-apps dev, the other runs claude.

Tab one: ib-apps dev

Install the CLI once, sign in, and create an app or pull one you already have:

npm install -g @inferencebureau/apps
ib-apps login
ib-apps create "Restock" --platform shopify
cd restock
ib-apps dev

ib-apps dev syncs the folder with the app's development environment, then keeps syncing both ways: a save on your computer is pushed, and a change from the dashboard, the assistant or a teammate is pulled every two seconds. For team owners and admins it also runs the dev server on your computer, reached through a tunnel, so the app's development address, its actions, Shopify webhooks and jobs all reach the code in your folder. It prints the address to open.

Leave it running. From here on, anything Claude saves in this folder is live at that address a few seconds later. The dashboard's Terminal panel shows each file it sends and receives, which is a useful place to watch what an agent is doing.

Tab two: Claude Code with the app's MCP server

Open a second tab in the same folder. Add the development environment's MCP server once per app:

claude mcp add --transport http restock-dev https://api.inferencebureau.com/v1/apps/restock/environments/development/mcp
claude

In Claude Code, run /mcp and choose Authenticate. Your browser opens the dashboard, you approve and pick the team, and Claude gets a session that acts as you. No key goes into the config. The app's Settings → API keys page shows the same command with the app filled in.

Claude now has two ways into the app:

  • The folder. It reads and edits files with its own tools, and runs ib-apps test in its shell. ib-apps dev carries every save to the server.
  • The MCP server. It reads the API reference, finds, counts and aggregates records, runs actions, reads the event log and the type errors from the dev server's last check, and creates models and fields with the right storage keys.

The two meet in the middle. When Claude calls create_model, the schema file is written on the server, and ib-apps dev pulls it into your folder within two seconds, so the checkout and the environment stay the same.

The starter also ships an AGENTS.md, a CLAUDE.md that imports it, and skills under .agents/skills/ for models, actions, records, admin pages, storefront blocks, testing and security. Claude Code reads CLAUDE.md when it starts, so it already knows the layout and the rules: records change through actions, a field's storage key never changes, every members-only route checks who is asking.

Prompts to start with

The app below is a back-in-stock list for a Shopify store: shoppers leave an email on a sold-out variant, and the merchant sees who's waiting. These prompts work in that order, but each one stands on its own.

Get oriented

Read AGENTS.md, then call get_reference on restock-dev and tell me which models
and actions this app has, which are synced from Shopify, and who may call each action.

Claude reads the guide from disk and the reference from the server. The reference is written from the environment's schemas each time it connects, so it names your real models and fields, including the synced ones such as shopifyProduct and shopifyProductVariant.

Add a model

Add a restockRequest model: email (required), variant (belongs to shopifyProductVariant),
and status (pending or notified, default pending). Use create_model so the storage
keys are set, then show me the schema file it wrote.

The file appears in your folder through ib-apps dev, and the answer includes what the change means for records that already exist. Ask for a rename later with change_field and the storage key stays, so existing values follow.

Build against real data

Find variants that are out of stock right now with find_records on
shopifyProductVariant (inventoryQuantity at most 0). Create 15 restockRequest
records spread across them so I have something to look at.
Add an admin page that lists pending requests grouped by product, with the count
for each, newest first. Read the shopify-admin-pages skill before you start.

Claude writes the page in web/, ib-apps dev pushes it, and the dev server hot-reloads it at the app's address. It can check the page's numbers against aggregate_records rather than reading every record.

Check its own work

Call get_problems and fix any type errors. Then run ib-apps test and fix what fails.

get_problems returns the type errors from the dev server's check a few seconds after each save, the same list as the dashboard's Problems panel. A TypeScript error there would fail production's build, so this is worth asking for before every deploy. Tests run in Claude's own shell from the folder; they use in-memory records and touch nothing on the server.

Debug from the event log

Something failed overnight. Call list_events with failed: true, read the logs on
the failed runs, find the cause in the code, fix it, and rerun the action on one
of the affected records to confirm.

A failed action's event carries its error and its logs. When Claude reruns it with run_action, the answer has the record as the action left it and the new run's event id, so it can look that run up too. If the development dev server is asleep, the run wakes it and waits instead of failing.

Ask about the data

How many pending requests does each product have, and what's the average wait
since createdAt? Use aggregate_records, not find_records.

Ship it

Problems is clean and the tests pass. Run ib-apps deploy and tell me when
production is live.

ib-apps deploy syncs, ships development to production, and waits until it's live at the app's inferencebureau.app address, or says why the build failed. The MCP server has a deploy tool too, but it's only for when you've asked for one, and it takes a team owner or admin.

Writes go through your actions

When Claude creates, updates or deletes a record through the MCP server, it's the same write the REST API makes. A model's own create, update and delete actions run when it has them, with their validations and onSuccess hooks, so Claude can't save a record your code would refuse. A refused write comes back with each field's problem, and a thrown action with its message and logs, which is usually enough for Claude to fix the cause and try again.

Every call is logged as an mcp event in the environment's Logs, with its arguments (secret fields redacted), how long it took and whether it failed. After a long session you can read back exactly what Claude did.

Add production for questions, not edits

Production is a separate server with its own records. Add it under its own name:

claude mcp add --transport http restock-prod https://api.inferencebureau.com/v1/apps/restock/environments/production/mcp

Then ask about live data by naming the server:

Using restock-prod, which ten products have had the most restock requests in the
last 7 days, and how many of those variants are back in stock now?

Production's server can change live data for a team owner, an admin or a key. A member with Edit gets every tool on development and only the reading tools on production; the writing tools aren't listed for them at all. Production's files change only by a deploy, so its server has the file-reading tools and none of the writing ones.

Each developer gets their own environment

ib-apps dev runs the dev server on one computer at a time. If a second computer starts serving the same environment, the newer one takes over. When two people build at once, give each their own development environment under Settings → Environments and point the CLI at it with --env. Their MCP servers are separate too, at that environment's name, so each person's Claude reads and writes only their own records.

Frequently asked questions

Do I need the MCP server if Claude already has the folder?

For code alone, no: ib-apps dev syncs whatever Claude saves. The MCP server is how it reaches the app's records, actions, event log and type errors, and how it creates models and fields with the storage keys set.

Does this work with other coding agents?

Yes. Codex, Cursor and other MCP clients that speak Streamable HTTP take the same server, as a command or as JSON in their mcp.json, and read the same AGENTS.md and skills.

What does it cost?

Each MCP tool call counts as one request toward the app's plan, like an API call. Running the dev server on your own computer uses no dev server hours. Claude Code's own model usage is billed by Anthropic as usual.

Can Claude work on the app without a checkout?

Yes. Development's MCP server has tools to list, read, search, write and edit the app's files, and the server tells the agent to read AGENTS.md first. The one thing it can't do there is run the app's tests, which need a checkout and ib-apps test.

One bill for your apps, your models and your reviews.

Host apps and functions next to your model traffic. Routing picks each turn's model by your quality floor, and Auditor reviews every pull request against your own rules.