Skip to main content
Provider selection is controlled by runtime config v2 under providers, with backward-compatible mirrors in auth, billing, storage, llm, database, and capabilities. Use none plus the matching feature flag when a provider family must be disabled. Disabled features are hidden from UI and blocked by server-side gates with capability_disabled responses.

Core Providers

Enterprise Provider Families

Disable External Processors

For a private deployment, disable high-risk external processors explicitly:
For strict deployments, combine this with compliance.allowExternalProcessors=false and allowlist only approved processors such as models:openai or objectStorage:s3. Vector search is opt-in. For deployments that do not need knowledge retrieval, keep:
To enable memory and knowledge search, use convex vector search plus one supported embeddings provider (ai-gateway or openai). Convex performs indexing, model-versioned embeddings, scoped hybrid retrieval, and cleanup inside the deployment; no separate worker is required.

MCP server credentials

People can add their own MCP servers. Their tokens are encrypted with a key you provide. Generate it once:
Store it as MCP_CREDENTIAL_ENCRYPTION_KEY on the web runtime. During rotation, place the old value in MCP_CREDENTIAL_ENCRYPTION_KEY_PREVIOUS, deploy the new primary key, rewrite saved credentials through the management API, then remove the previous key. MCP responses expose only hasAuth; bearer tokens and custom header values are AES-256-GCM encrypted at rest and are never returned by list APIs. Production MCP URLs must use HTTPS and pass Overlay’s SSRF validation. Each server has a default execution policy of allow, approval_required, or deny, plus optional per-tool overrides. Use approval_required as the enterprise default when a server contains write-capable tools. Tool name, argument hash, policy decision, result status, duration, and request context are persisted for audit without storing raw tool arguments.

Deep search retrieval

TinyFish Search and Fetch provide source evidence; Overlay’s model synthesizes it with citations and can request follow-up searches. This does not call TinyFish’s separately billed Research API. The fast, standard, and deep effort levels cap search queries at 2, 4, and 6 and fetched pages at 3, 6, and 20, respectively. Deep defaults to at most 10 pages unless maxResults requests more. Fetch requests contain at most 10 URLs; per-page time budgets are 15, 25, and 45 seconds, with a 60-second request deadline. Text budgets and domain/date filters still apply, and maxAgeSeconds maps to Fetch cache freshness. Partial results retain readable pages and clearly label unfetched snippets. If no usable pages are available or TinyFish is unavailable, the tool falls back to Elo Parallel. Completed Search and Fetch requests count toward the existing per-call tool charge, including completed requests preceding an Elo fallback. At most six Search requests, two Fetch batches, and one Elo fallback fit within the existing reservation. Failed overall operations retain the reconciliation path.