Integrations / Langfuse
Langfuse
Loop-level evidence: which step of which workflow spends the money.
Why connect it
At scale the bill is the sum of agent loops. A trace store already records every model call in every conversation with tokens, model, and cost. GitDoctor reads it and answers the question a usage export never can: cost per conversation, per workflow, per step — and which step to change.
What you'll see
- Cost per run for every workflow, ranked
- The step that dominates each workflow's cost
- Context growth within a run (is the history re-sent every turn?)
- Cache rate and heavy-model share per workflow
- Findings that name the step and the fix
What the credential can and cannot do
Where to get it: Langfuse → Project settings → API keys. Self-hosted: your own host URL.
It grants: Read on traces and observations for that project.
It cannot: GitDoctor never writes traces or scores.
Whatever you connect lives in an encrypted cookie in your browser and nowhere else. Sign out and it's gone; revoke it at the provider and it stops working immediately.