tutorials

First captured request

Fifteen minutes, no application changes: sign up, send one request through the gateway, and read the recorded evidence in the dashboard. Everything else Understudy does builds on the loop you'll close here.

1 — Sign up and save the starter key

Sign in at app.understudylabs.com and name your workspace. The handoff screen shows your starter sk_* key — copy it now; the plaintext is never shown again. You land on the Setup page with the key prefilled.

shell

export UNDERSTUDY_API_KEY="sk_live_..."   # the starter key

2 — Turn capture on

Captures are opt-in per workload. Open Projects → rehearsal → Workloads and flip capture on for the main workload — the one unscoped traffic resolves to.

3 — Send the request

Pick your path on the Setup page (both end here identically). With an Anthropic key of your own, the BYO smoke test is:

byo smoke test

curl https://api.understudylabs.com/v1/messages \
  -H "x-api-key: $UNDERSTUDY_API_KEY" \
  -H "x-understudy-upstream-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-haiku-4-5","max_tokens":32,"messages":[{"role":"user","content":"Say ok."}]}'

Without a provider key, use managed mode against a catalog model instead (ids from GET /v1/models or the Models page):

managed smoke test

curl -i https://api.understudylabs.com/v1/chat/completions \
  -H "Authorization: Bearer $UNDERSTUDY_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"glm-5.1","max_tokens":32,"messages":[{"role":"user","content":"Say ok."}]}'

4 — Read the response headers

With -iyou'll see the gateway's legibility headers alongside the provider's normal response:

x-understudy-request-id: 0190...        ← the capture's name
x-understudy-mode: managed              ← whose credential paid
x-understudy-route: understudy          ← which arm served it
x-understudy-effective-model: glm-5.1   ← what actually ran

5 — Find the capture

Open rehearsal → Captures. Your request is the top row within seconds; the request id matches the header. Open it:

metadata — status, first-byte latency, workload, requested vs served model, mode.
request / response tabs — the raw bodies, byte for byte. This is the evidence evals and fine-tuning are built from.