/playground

Skill playground

Run one ACE gateway skill (PII redaction, prompt compaction, routing, guards) on a payload you paste and see what it would change. Free with a dev key; no provider is called.

Playground: one skill, run alone on the payload you paste. The production gateway runs skills together on live traffic, with session state, caching and real provider pricing, so results in production can differ. See each skill's docs for production behaviour.

What's simplified

One skill at a time (production chains several).
No session state (no fold latch, memos or cross-request cache).
No provider call, so no real cache counters.
Default params unless you pass params.
Learned detectors only where prod has them loaded.
Trajectory compaction runs each call as one fresh session: a caller that declares cache breakpoints is folded here, where production without a session id would skip it.
pii_ner redaction can raise the token count, because a placeholder can be longer than the text it replaces.
Semantic cache and knowledge graph do read-only lookups against your tenant's own data; nothing is written.

skill

example

loading examples…

payload (JSON: system, messages, tools, model, params…)0.2 KB / 2 MB

dev key

Kept in this page's memory only, never stored. No key? Get a free key. Evaluations are free and never call a provider.

Pick a skill, load an example or paste a payload, add your dev key and run.

MCPevaluate_skillon /mcp/admin · /mcp/onprem

Your tenant, no key needed, or pass dev_key. Contract: get_doc skill-evaluation.

over HTTP

The endpoint the Run button calls, with your key in ACE_DEV_KEY.

POSThttps://engine.acefleet.dev/v1/skills/pii_ner/evaluate

curl -sS 'https://engine.acefleet.dev/v1/skills/pii_ner/evaluate' \
  -H 'Authorization: Bearer '"$ACE_DEV_KEY" \
  -H 'content-type: application/json' \
  --data-raw '{
  "system": "You are a helpful assistant.",
  "messages": [
    {
      "role": "user",
      "content": "Hi, I am Jane Doe, email jane.doe@example.com"
    }
  ]
}'