Everything here runs against the same public API your agent uses — the console has no private endpoints. ← home
No signup? Try instantly with the shared public trial key ak_trial_taedri_public_notebook —
100 requests/day shared across everyone, 25/day per address, read-only. The
example Kaggle notebook uses it.
curl -s -X POST https://www.taedri.dev/v1/signup -d '{"email":"you@example.com"}'claude mcp add --transport http taedri https://www.taedri.dev/mcp \ --header "Authorization: Bearer YOUR_API_KEY"
curl -s -X POST https://www.taedri.dev/v1/agent -H "Authorization: Bearer YOUR_API_KEY" \
-d '{"action":"retrieval.search","args":{"query":"parse csv header"}}'primitive_search — Retrieve reusable capability primitives by intent.
primitive_get — Retrieve one primitive by immutable id after search.
find_reuse — Guard against reinventing a solved capability.
capability_compose — Compose a route of existing primitives for a natural-language request instead of generating code.
primitive_corpus_status — Read measured corpus coverage: unique searchable IDs, four-field presence, versioned source-grounded/usefulness-passing descriptions, embeddings, source cursors, and the latest 100M-loop state.
primitive_description_gaps — Return a bounded cursor page of immutable primitive IDs that still lack an effective title, blackbox, typed input, or typed output after core/overlay dedupe and accepted descriptor revisions.
record_reuse_outcome — Record metadata-only feedback that a searched primitive was accepted, actually reused, dismissed, or ignored.
capability_free — 200 requests / day
capability_pro — 20,000 requests / day
Prices live at /v1/pricing (draft until billing goes live). Hit the limit? The
429 carries Retry-After + X-RateLimit-* headers, an upgrade pointer, and the
share-to-continue lane: POST /v1/contribute grants +5 requests per accepted
anonymized card (up to +200/day), candidate-only until reviewed.
401 — missing/invalid key; sign up or check the Bearer header. 429 — daily plan
limit; see above. Cold wake — the first search after a machine wake loads the full governed index;
the server warms it at boot and /healthz reports index_warm, but a search
that lands mid-warm waits instead of failing. Every response carries candidate /
serves_truth — results are governed candidates until promoted, never silently upgraded.
Home · News · Status · Test console