Read the code first. Decide second.

Three surfaces. One engine.

Raptor sits between your AI proposals and the actions/responses your users see. The AI proposes; the deterministic core disposes. You keep upstream flexibility (any model, any prompt). You get a fixed downstream process — trust-boundary metadata on every response, governance gates on every action, hash-chained audit events on every meaningful state change.

Built on TypeScript/Node + Go (control plane), Neon Postgres with RLS, Clerk for identity, Cloudflare for edge, unified Provider abstraction over Anthropic / OpenAI / Gemini / Together AI.

The canonical end-user product surface at /app demonstrates the full trust-boundary rendering, governed actions, and progressive disclosure — built on the same API the SDK consumes.

Pick the shape that fits your stack.

What integration looks like.

The TypeScript SDK is generated from the OpenAPI spec. Trust-boundary types preserved through the SDK.

// SDK source at sdks/typescript/; npm publication under @raptor/sdk is pending.
// During evaluation: install directly from the repo or vendor the generated client.
import { RaptorClient } from '@raptor/sdk';

const raptor = new RaptorClient({ apiKey: process.env.RAPTOR_API_KEY });

const decision = await raptor.decisions.classify({
  nestId,
  userMessage,
  history,
});

// decision.decisionType is one of:
//   'respond' | 'ask' | 'propose_action' | 'execute_confirmed_action'
// decision.proposedAction exists when decisionType === 'propose_action'
// decision.requiresConfirmation reflects governance gate state

Auth is via Clerk for production tenants (Google SSO, SAML/OIDC ready); bearer tokens for programmatic access. OAuth 2.1 for MCP is in active development.

What's solid. What's in development. What's not built.

Surface	State
REST API + OpenAPI spec	Solid. 85 documented routes.
TypeScript SDK	Solid. Generated from OpenAPI; trust-boundary types preserved.
Provider abstraction (4 providers)	Solid. Cross-model eval validates parity.
Hash-chained audit log	Solid. 23 immutable tables with DB-level enforcement.
MCP server (token auth)	Solid. 4 tools, 4 resources.
MCP server (OAuth 2.1)	In active development.
Decision support backend confidence mapping	In active development.
Temporal trust lifecycle backend	In active development.
Self-hosted deployment	Not built. SaaS-first.
MCP registry publication	Deferred. Available on direct connection during evaluation.
Webhooks for governance events	Queued.
Self-serve API keys + sandbox	Queued.

Observed, not aspirational.

The friction you'll actually hit during integration.

• Engineering team is one founder + AI senior staff. Support response is direct and fast for active prospects. Not 9-to-5 in any time zone.
• Single SaaS deployment. No regional hosting yet.
• 14-day Workspace trial available. Infrastructure customers start with $10 minimum credit purchase (approximately 833 governed responses at $0.012 each).
• Documentation surface is thinner than the implementation.
• Provider rate limits (Together AI serverless, others).
• No webhooks yet. Move 8 is queued.

Read these five files in order.

1. docs/thesis/raptor-canonical-v1.md — the canonical product thesis. Locks in the architectural framing.
2. docs/thesis/raptor-interaction-model-v1.1-addendum.md — trust-bearing content semantics.
3. internal/engine/dmc_shadow.go — the shadow-extraction pattern. Real code demonstrating the proposal/commit architecture.
4. cmd/control-panel/decision_engine.go Process() — the orchestration spine.
5. docs/adr/ADR-021-phase-b-managed-open-weight.md — the multi-provider bridge architecture and its three amendments. How the team actually reasons about architectural decisions.

If the code reads cleanly and the ADR reasoning matches how your team thinks, the substrate is for you. If the code looks unfamiliar in patterns you'd disagree with, integration friction is going to be high regardless of what marketing claims.