Shipped capabilities

This page summarizes implemented behavior, not planned roadmap items.

Platform and compatibility

• one deployable Go engine for API traffic, workers, scheduling, gateway, and browser administration;
• OpenAI-compatible chat completions, streaming, models, tools, and errors;
• Anthropic-compatible Messages, streaming, tools, token estimates, and prompt cache usage passthrough;
• hosted and self-hosted providers through runtime connectors;
• end-to-end X-Request-ID correlation;
• public liveness, readiness, and human-readable status.

Identity and tenancy

• tenant-signed RS256, ES256, and EdDSA JWTs;
• engine-issued organization, role-ceiling, and user-bound API keys;
• platform, organization-admin, and organization-member authority;
• organization onboarding without database edits;
• forced PostgreSQL row-level security and multi-store tenant isolation;
• engine-owned teams with live roster-backed grants.

Models and commercial controls

• runtime connector registration, discovery, verification, update, key rotation, priority, weight, and disablement;
• platform connectors plus organization bring-your-own-key LLM connectors;
• default-deny model grants and caller-filtered discovery;
• heuristic and optional classifier-assisted model: "auto" routing;
• runtime integer pricing, margin, cache-aware billing, and model tiers;
• append-only credit ledger, transactional holds, top-ups, monthly limits, usage reporting, and reconciliation.

Knowledge and application state

• private and shared datasets with user, role, and team grants;
• durable file ingestion, website crawling, scheduled refresh, and job history;
• hybrid BM25/vector retrieval with optional cross-encoder reranking;
• CSV/XLSX ingestion and bounded read-only Text-to-SQL;
• projects with instructions, hidden knowledge stores, and dataset references;
• governed knowledge-agent answers with citations and bounded outcomes;
• stateful conversations, rolling compaction, semantic memory, explicit remember/forget, opt-out, retention, and right-to-forget cleanup.

Automation and tools

• one-shot agents and durable multi-turn runs;
• user-, team-, and organization-owned agents;
• tool grants, attached datasets, sampling defaults, and agent guardrails;
• server-side tool execution with idempotency and bounded results;
• human approval, denial, cancellation, budgets, deadlines, and durable events;
• one-time and POSIX-cron schedules;
• signed webhook delivery with retry and pollable event history;
• Gmail and Outlook OAuth connections with portable email tools;
• external MCP server registration, snapshot synchronization, user attachment, and namespaced agent tools;
• a read-only Deeplinq MCP server.

Governance and operations

• platform, organization, and agent guardrail layers;
• safety, PII, topic, and prompt-injection checks;
• append-only minimized audit events with hash-chain verification and signed anchors;
• OpenTelemetry traces and structured logs;
• backup/restore, billing reconciliation, key rotation, egress, and vulnerability-response procedures;
• fail-closed configuration, bounded ingress, graceful shutdown, cursor pagination, capacity ceilings, and durable recovery.

Deliberate boundaries

• the public tenant embeddings API is not shipped; /v1/embeddings is reserved for the authenticated memory sidecar;
• the Deeplinq MCP server is read-only;
• external MCP tool snapshots refresh only on explicit sync;
• user identity profiles remain in the tenant's system;
• payment-service-provider processing remains outside the engine;
• a configured dependency used by a request fails closed rather than silently degrading governance.

Use the API surface for route-level inventory and the task-oriented guides for implementation.