Ask Repave from Claude
Repave can answer questions about your features and what they cost from inside Claude, so you do not have to open Repave and read a table to find out where something has got to.
Once connected, you can ask in plain language:
- "What's the status of the checkout feature?"
- "Which features in the CRM project are still unimplemented?"
- "Has the order export feature merged yet?"
- "What's left in the billing module?"
- "What has the CRM project cost so far?"
- "How much did analysing the legacy code cost?"
- "What did we spend on the checkout feature last week?"
- "Which runs were the most expensive this month?"
What it can and cannot do
It is read-only. It can list the projects you have access to, list their features, report one feature's progress, and report what the AI runs on a project cost. It cannot change anything, start a job, approve a prototype, or see any project you could not already open yourself.
You connect once, per person. What Claude can see is exactly what you can see when you sign in — connecting does not grant access to anything new, and it does not let anyone else read your projects.
Connect Claude Desktop
- In Repave, open the account menu in the top right and choose Connected apps.
- Copy the connector URL shown under Connect a new app.
- In Claude Desktop, go to Settings → Connectors → Add custom connector.
- Paste the URL and choose Add.
- A browser window opens on Repave. Sign in if you are asked to, then choose Allow.
That is the whole setup. There is no token to copy, no configuration file to edit, and you will not be asked to sign in again.
Connect Claude Code
Run this once, using the connector URL from Connected apps:
claude mcp add --transport http repave <connector-url>
Then run /mcp, choose repave, and choose Authenticate. A browser opens on the same
consent screen; choose Allow.
What you get back
Asking about one feature returns its position in the Repave pipeline, in order:
| Stage | What it means |
|---|---|
| As-Is | The legacy behaviour has been captured as scenarios. |
| To-Be | The specification for the new system has been written. |
| Prototype | A UI prototype exists, with its version and whether it is approved. |
| Implemented | How many scenarios an agent has written code for. |
| Verified | How many scenarios a test run has actually seen pass. |
| Pull request | Whether a pull request is open, merged, or closed. |
| Merged | Whether the work reached a branch, and which one. |
Implemented and verified are deliberately separate. "Implemented" is the agent's own report that it wrote the code. "Verified" is a test run's observation that the code passes. A feature can be fully implemented and not yet verified, and the two are never reported as one number.
Some stages report not applicable rather than outstanding. A batch job project has no user interface, so it has no prototype stage; a feature that did not go through user acceptance testing has nothing recorded for verification.
Costs
Cost answers use the same figures as the project's Costs page, so the two always agree. Anyone who can open a project can ask what it cost; connecting grants nothing extra.
You can ask for:
| Question | What comes back |
|---|---|
| A project's cost | The total, the split by pipeline stage and by model, what legacy analysis cost, the most expensive features, and spend that belongs to no single feature. |
| A feature's cost | Its total by stage and by model, next to its status, and the cost per To-Be scenario. |
| Spend over time | Spend per day or per week, for a project or one feature, over any date range. |
| Individual runs | Each AI run with its cost, model and tokens, filtered by feature, stage, legacy analysis or date, newest or most expensive first. |
Dates are in UTC, weeks start on Monday, and amounts are in US dollars.
Estimated and unknown costs are reported as such. Some providers do not report what a run cost, so Repave estimates it from the provider's published prices; those amounts are marked estimated. If a model has no published price, the run's cost is unknown rather than zero, and it is counted separately so it is never mistaken for a free run.
Legacy analysis is the work of reading the legacy code: discovering its features, capturing its current behaviour as scenarios, and extracting its architecture and data. It is reported as its own figure, and it is part of the Other stage rather than added on top of it, so the stages always add up to the total.
When a list of runs is cut short, the totals still cover every run that matched, so a figure is never a sum of only the runs that were shown.
Naming a project or feature
You do not need identifiers. "The checkout feature" is enough. If more than one feature matches what you asked for, Claude is told which ones matched and will ask you which you meant rather than guessing.
Disconnecting
Open Connected apps and choose Revoke next to the app. Access stops immediately — the next thing that app tries to read is refused. You can connect it again later at any time.
The same page lists every app you have connected, when you connected it, and when it was last used, so you can tell at a glance whether a connection is still being used before you remove it.
If the connector will not connect
Claude Desktop and Claude Code connect from your own computer, so they can reach a Repave installation on your company network or VPN.
Claude in a web browser is different: it connects from Anthropic's servers, which have to be able to reach your Repave installation over the public internet. If your installation is on an internal network, behind a VPN, or has no inbound internet access, use Claude Desktop instead.