MUSCLE Flex matches each supported coding-agent task to the best qualified model path for that work. Your integration stays fixed while the frontier keeps moving.
Different tasks, different strengths
A quick explanation, a tool-heavy search across a repository and a multi-file repair reward different model strengths. A production agent may also stream, call tools, continue for many turns and cross several cost boundaries in one session.
So Flex looks at the route around the task and the workflow, not at a one-line benchmark prompt. The question it answers is narrow and practical: which path can do this request properly, and which of those is the sensible spend?
Qualification comes first
Before any route is compared, Flex removes the ones that cannot meet the request's hard requirements:
- the tool shape the request uses
- how the response streams
- how much context the work needs
- your account's policy
- the balance available to fund it
Only then does the routing policy compare the paths that remain and choose the best qualified fit for the task. A low score cannot make an incompatible route safe. Hard requirements stay hard.
Your application keeps the tools
Flex returns function calls in the supported API shape. Your application decides whether to run them, runs them in its own environment and sends the results back.
Routing the model call gives Flex no shell access and no hold on your repository. Approvals, sandboxes and permissions stay where you put them.
Prepaid, with a hard stop
Beta accounts fund a balance before use. Flex holds enough of it before admitting new paid work and settles measured usage afterward. When the available balance is too low, new work is declined rather than run on credit.
The portal shows the remaining balance and recent activity, so spend is something you read, not something you discover. How beta billing works.
New models earn their place
Flex does not send your work to a model because it is new or popular. Candidate paths are tested with controlled coding tasks, real SDK calls, tool continuations, concurrent requests, cancellation and transport cases, funding denial and accounting replay. A path stays out of production until the relevant gates pass. When something better qualifies, the roster changes and your integration does not.
Explicit boundaries
Flex supports selected OpenAI-shaped and Anthropic-shaped workflows. It does not claim every endpoint, field, hosted tool, media type or background mode.
The compatibility matrix is the contract. Each behavior is marked as passed, failed, excluded or pending, against a named client version. See compatibility.