Anthropic — BYOK Setup

Anthropic is the recommended provider for premium-quality evaluate and evolve. Default model is claude-sonnet-4-6; claude-opus-4-7 is significantly more expensive but worth it for high-stakes evaluations.

Anthropic does not offer an embeddings endpoint. If you choose Anthropic as your only credential, Engramia automatically falls back to the lightweight local LocalEmbeddings (sentence-transformers, 384-dim) for /v1/recall. For production semantic search at scale, add a separate OpenAI or Gemini credential with purpose=embedding.

1. Create an API key in Anthropic

1. Sign in at console.anthropic.com.
2. Settings → API Keys → Create Key.
3. Name it descriptively (e.g. engramia-prod).
4. Optional: scope to a specific Workspace if you have multiple.
5. Copy the key now — Anthropic shows it exactly once. The key begins with sk-ant-.

2. Set a spending limit

1. Plans & billing → Spending limit in the Anthropic console.
2. Set a monthly cap. Anthropic enforces this hard — once you hit it, API calls 402 until the next billing cycle starts or you bump the cap.

When the limit is reached, Engramia surfaces the 402 to the caller and marks the credential invalid after the first failure. Subsequent calls fall back to demo mode.

3. Add the key in Engramia

In the dashboard:

1. Settings → LLM Providers → Add provider.
2. Provider: Anthropic.
3. Paste the key (starts with sk-ant-).
4. Default model: leave blank to use claude-sonnet-4-6, or set explicitly:
5. claude-haiku-4-5 — cheap, fast, good for high-volume eval
6. claude-sonnet-4-6 — default, good balance
7. claude-opus-4-7 — premium, ~5× the cost of Sonnet
8. Click Validate & save. Engramia pings https://api.anthropic.com/v1/models (5 s timeout).

REST equivalent:

curl -X POST https://api.engramia.dev/v1/credentials \
  -H "Authorization: Bearer engramia-prod-..." \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "anthropic",
    "purpose": "llm",
    "api_key": "sk-ant-...",
    "default_model": "claude-sonnet-4-6"
  }'

Note purpose: "llm" (not "both") — Anthropic has no embeddings.

4. Add an embeddings credential (production)

For real semantic search you need an embeddings provider. Two options:

OpenAI for embeddings (recommended — best price/quality):

curl -X POST https://api.engramia.dev/v1/credentials \
  -H "Authorization: Bearer engramia-prod-..." \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "openai",
    "purpose": "embedding",
    "api_key": "sk-...",
    "default_embed_model": "text-embedding-3-small"
  }'

Gemini for embeddings:

curl -X POST https://api.engramia.dev/v1/credentials \
  -H "Authorization: Bearer engramia-prod-..." \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "gemini",
    "purpose": "embedding",
    "api_key": "AIza...",
    "default_embed_model": "gemini-embedding-001"
  }'

Engramia's resolver will use the LLM credential for /v1/evaluate and the embedding credential for /v1/recall automatically.

5. Per-role routing (Business tier)

curl -X PATCH https://api.engramia.dev/v1/credentials/{id} \
  -H "Authorization: Bearer engramia-prod-..." \
  -H "Content-Type: application/json" \
  -d '{
    "role_models": {
      "eval": "claude-haiku-4-5",
      "evolve": "claude-opus-4-7",
      "default": "claude-sonnet-4-6"
    }
  }'

Useful pattern: cheap Haiku for eval (3× per evaluation), premium Opus for the rare evolve calls.

Troubleshooting

400 category=auth_failed : Anthropic rejected the key. Confirm the key is exact (no leading whitespace from the copy/paste) and that the workspace it belongs to is still active.

recall returns lower-quality matches than expected : You may be using the local 384-dim embeddings fallback. Add an OpenAI or Gemini embedding credential per step 4 above and run Settings → Embedding reindex (Pro+ tier) to rebuild the HNSW index with the new dimensionality. See embedding-reindex.md.

Anthropic returns 529 Overloaded : The Anthropic API is under temporary load. Engramia retries with exponential back-off up to 3 attempts. If all attempts fail, the call returns 502 to your client. The credential remains active — no need to re-validate.