Local Models
Where local models fit Coven: the fastembed embedding model used by coven-memory, and local-model harnesses through adapter manifests.
1 min read
Local models fit Coven in two places today, and neither one turns the daemon into a model gateway.
Local embeddings: shipping now
The coven-memory archival layer runs a local embedding model: nomic-embed-text-v1.5 (768-dimensional) via fastembed. The model downloads once on first run and is cached; every ingest and search after that is fully local and air-gapped. No provider account or API key is involved. See Semantic memory for the full stack.
Local-model agents: through the harness boundary
A local model that powers an agent (Ollama, llama.cpp, or a custom local model runtime) still arrives as a harness. Any CLI that fronts a local model can be registered through an adapter manifest today — the same mechanism the trusted grok and hermes recipes use — as long as it meets the adapter bar.
Posture
Local models are not a special daemon credential feature. The daemon does not grow provider-specific model routing just because the provider is local.
The same rules apply:
- Stable harness id.
- Fixed adapter command shape.
- Project-root and working-directory validation.
- PTY supervision.
- Session ledger and event log.
- No secret fields in the socket API.
For adapter requirements, see Custom adapters and Models.