Claude Code on Flux
Fix a failing test and ship the PR
Claude Code finds why the checkout test broke, writes the fix, opens the PR and tells the team.
“The checkout test fails after the payments upgrade. Fix it, open a PR and post the link in #eng.”
- ModelLive
- Page readingLive
- AppsShipping
Murage- Read the failing test with Claude$0.021
- Check the library changelog$0.005
- Write the fix with DeepSeek$0.008
Open the pull request on GitHubShipping$0.010
Post the link in #engShipping$0.010
Examples. Names, prices and costs are illustrative.
What Flux did, step by step
Flux split the ask into jobs and sent each one to the model or tool that fits it. It picks among vetted models for each kind of task, cost-aware.
Read the failing test with ClaudeLive
Model
Working out why a test fails takes careful reading, so Flux sent it to a frontier model.
Check the library changelogLive
Page reading
The failure came after a library upgrade, so Flux read the changelog page instead of guessing what changed.
Write the fix with DeepSeekLive
Model
Once the cause was clear, the fix was routine code. Flux picked a lower-cost coding model for it.
Open the pull request on GitHubShipping
Apps
The pull request went through the GitHub connection.
Post the link in #engShipping
Apps
The link went to #eng through the Slack connection.
Result Fixed. PR #412 is open, the tests pass and the link is in #eng.
The receipt
Every call comes back with the model or tool that ran it and what it cost, all on one bill.
These costs are illustrative, not a quote. flux-auto bills each model step at the Standard rate, or the Fast rate when it sends the step to the Fast lane and a Fast model answers. Never more than Standard. Pin a model and you pay its own rate. See pricing
Example receiptIllustrative
| Step | Cost |
|---|---|
| Read the failing test with ClaudeModel | $0.021 |
| Check the library changelogPage reading | $0.005 |
| Write the fix with DeepSeekModel | $0.008 |
| Open the pull request on GitHubApps | $0.010 |
| Post the link in #engApps | $0.010 |
| One bill | $0.054 |
Run it with your own agent
- Get your key.
- Connect Claude Code to Flux with flux-auto as the model. Claude Code setup
- For page reading and images, add the Flux tools server to Claude Code’s MCP settings as a remote HTTP server at
https://api.fluxrouter.ai/mcp-tools/, with your Flux key as the bearer token (Authorization: Bearer sk-flux-…). It offersweb_fetchto read a page andimage_generateto make an image. - Ask it for the job in your own words, like the ask at the top.
What works today
- ModelLiveModel steps work today, as soon as your agent points at Flux with flux-auto.
- Page readingLivePage reading works today on a paid key. It covers public pages only.
- AppsShippingApps on your own key are shipping. Until then, your agent can use its own tools for these steps, and Murage runs them today with apps built in.
Try it with another agent
One key, any agent. If it can talk to OpenAI or Anthropic, it can connect to Flux.
- >_
Questions
Does Claude Code still use Claude?
Where Flux picks it, yes. With flux-auto, Flux picks among vetted models for each kind of task, cost-aware, so the hard reading and the routine fix can go to different models. Pin a model if you want it for every step.
Does every step work on my key today?
Not yet. Apps on your own key are shipping. Until then, your agent can use its own tools for these steps, and Murage runs them today with apps built in. The steps marked Live work today.
What will it cost me?
The costs on this page are illustrative, not a quote. flux-auto bills each model step at the Standard rate, or the Fast rate when it sends the step to the Fast lane and a Fast model answers. Never more than Standard. Pin a model and you pay that model’s own rate instead. Web, image, voice and app steps have their own prices.
Anything to know about Claude Code?
settings.json wins over your shell. If both set the same variable, the settings.json env block takes precedence. Claude Code turns off MCP tool search on a custom host. Set ENABLE_TOOL_SEARCH=true to bring it back. On a brand new install, the settings.json env block can fail to load on the very first run. Run it again, or export the variables in your shell instead.
More jobs your agent can do
Give your agent everything else.
One key for every model and tool in this example.