The unified stack
Layer-by-layer
Face — SINT Avatar
3D WebGL avatar with voice I/O, Conversation Compiler, 18 widget types. React Three Fiber.
Brain — SINT Console
31 feature modules, visual workflow builder, real-time agent orchestration.
Hands — SINT Operators
LinkedIn automation, reply classification, CRM routing. Production agentic loops.
Content — CMO Operator
Video-to-content pipeline with 18 skills, multi-channel publishing.
Spine — SINT Protocol
Capability tokens, policy gateway, evidence ledger, 12 transport bridges.
OS — OpenClaw runtime
Agent runtime: WebSocket gateway, channel plugins, MCP router, session manager, sandbox.
The Protocol core
Every product and agent action flows through the same governance core. This is not an optional middleware — it’s the only path to physical actuation.1
Capability tokens
Ed25519-signed tokens encode resource scope, allowed actions, physical constraints, time bounds, and a delegation chain of max 3 hops. Attenuation-only delegation: a child token’s scope is always a subset of its parent’s.
2
Policy Gateway
A single
intercept() entry point. Every request validates schema, token, scope, rate limit, physical constraint, and tier assignment before reaching a transport bridge.3
Evidence Ledger
SHA-256 hash-chained append-only log. Every decision is recorded with a prevHash pointer, providing tamper evidence without requiring blockchain infrastructure.
4
Transport bridges
Twelve bridges translate protocol-specific actions (MCP tools, ROS 2 topics, MAVLink commands, OPC UA writes, etc.) into a unified policy evaluation format.
Cross-system policies
The Protocol’s differentiating feature is cross-system policies — invariants that span multiple transport bridges simultaneously.
No single-protocol safety system can enforce these. The Protocol can because every bridge publishes state into the same gateway.
Example: voice command to deploy
What happens when a user says “Jarvis, deploy the staging build”:1
Voice → text
Qwen3.5-Omni or OpenClaw Voice Wake transcribes input to text.
2
Text → agent session
OpenClaw gateway routes to the active SINT agent session.
3
Agent → tool call
Agent decides:
exec("railway up --environment staging").4
Tier classification
@sint/openclaw-adapter classifies the tool call as T2 (deploy action).5
Cross-system check
Is
robot.moving active? No → continue. Is drone.armed active? No → continue.6
Policy Gateway
Token scope check passes. Rate limit passes. Physical constraint not applicable. Decision: ALLOW.
7
Evidence Ledger
SHA-256 entry written with prevHash linkage. TEE signature if configured.
8
OpenClaw executes
The shell command runs. Output captured.
9
Result → Conversation Compiler
Agent output rewritten for natural speech: “Staging deploy is live.”
10
Avatar reacts
Smile expression, head-nod animation, status widget (“Staging” → success, 3s).
11
Evidence HUD
New entry appears in real-time feed.
Example: voice command to move a robot
“Jarvis, move the robot two meters forward”:1
Voice → agent → tool call
Agent proposes
nodes.invoke("move_forward", distance=2).2
Tier classification
T3 — physical, irreversible in the sense that a collision cannot be undone.
3
Cross-system check
All clear.
4
Policy Gateway → escalate
T3 requires human approval.
approvalId = "apr-456" is queued.5
Avatar reacts
Surprised expression, thinking animation, action widget (“Approve 2m forward?”, 30s timeout).
6
Human approves
One click on the widget. Gateway resolves approval.
7
Node action executes
Robot moves 2m forward. StateTracker activates
robot.moving.8
During motion
File writes, shell execution, and deploys are automatically denied by cross-system policies.
9
Motion complete
robot.moving deactivates. Evidence ledger: full chain from request → approval → execution → completion.Where to read more
SINT Protocol overview
Detailed treatment of the Protocol layer.
SINT OS guide
How the products integrate through os-core.
Tiers and escalation
T0–T3 and the Δ-factor escalation function.
Bridges
All twelve transport bridges, per-resource state machines, profiles.