Docs.

Surfaces

The product exposes the same authorization through two surfaces. Both run inside one Next.js app and share the compiler service code for retrieval, context-pack assembly, and source-ACL filtering.

• MCP — Model Context Protocol server at /mcp. Designed for agent frameworks that already speak MCP.
• REST — /api/v1. Everything MCP exposes plus admin operations (audit events, sync status, sub-resource lookups).

MCP tools

• search_company_memory — Hybrid search over normalized artifacts, ACL-filtered.
• get_entity — Read a single entity with its readable evidence.
• get_entity_graph — Traverse the entity graph; explanations sanitized.
• get_current_facts — Current typed facts with conflicts and provenance.
• get_service_context — Service-level owners, reviewers, related artifacts.
• get_skill — Latest published skill version the caller is authorized to read.
• create_context_pack — Compile a task-specific context pack with citations.
• propose_memory_update — Create a draft memory proposal — never auto-promoted.
• record_task_outcome — Attach an outcome to an agent run; citations re-checked.

REST endpoints

• POST /api/v1/context-packs — Compile a context pack for a task brief.
• GET /api/v1/context-packs/:id — Read a previously compiled context pack.
• GET /api/v1/entities/:id — Read entity context (same authorization as MCP).
• GET /api/v1/entities/:id/graph — Entity graph traversal with readable evidence.
• GET /api/v1/entities/:id/timeline — Entity timeline of artifact events.
• GET /api/v1/facts?entityId=:id — Current facts and conflict groups for an entity.
• GET /api/v1/skills/:name — Latest published skill version.
• POST /api/v1/memory-proposals — Create a draft memory proposal.
• POST /api/v1/agent-runs — Open an agent run bound to a context pack.
• POST /api/v1/agent-runs/:id/trace/steps — Append a step trace; raw prompts rejected.
• POST /api/v1/agent-runs/:id/trace/tool-calls — Append a sanitized tool-call trace.
• POST /api/v1/agent-runs/:id/outcome — Record the outcome of an agent run.
• GET /api/v1/audit-events — Tenant-scoped audit feed.

What every response guarantees

• Public payloads never include tenant IDs, source-connection IDs, external IDs, provider tokens, raw settings, or hidden record counts.
• Every claim, fact, artifact, and citation is filtered through source ACLs before it leaves the API.
• Redactions are reported as generic counts and reasons. We do not leak restricted titles, URLs, or snippets through error messages.
• Write-shaped tools (propose_memory_update, source-write candidates) only create draft records. Promotion and execution require owner review.

Want the full spec?

The schema reference, auth flow, error model, and end-to-end Linear-issue-to-PR-review walkthrough land with the pilot. Book a call to request early access.