Skip to main content

ClawQLStreams

Event-driven autonomous agents you own — not a metered lab runtime.

World events enter. WORM records everything. Significance filters decide when agents wake — with full MCP tools, virtual-key budgets, and Durable Object scale.

Three delivery modes

One event loop. Three ways to respond.

The same subscription can audit continuously, keep agents ambiently aware, and escalate only when the world crosses a threshold — without a human opening a chat.

Reactive

Every event writes to WORM immediately — compliance-grade recording whether an agent wakes or not.

Ambient

Events buffer in NATS and surface on the next MCP tool call — awareness without spawning a subprocess.

Autonomous

A significance filter escalates only what matters into a full agent session with ClawQL tools and budget.

Protocol Fabric

Anything in. Anything out. MCP as the IR.

ClawQL Core turns any API into MCP. mcp-api-adapter turns MCP back into OpenAPI, GraphQL, gRPC, CLI, or WebSocket. Streams wraps that fabric in an event loop so the bus can act without a human at the console.

OpenAPIGraphQLgRPCWebSocketCLIClawQL Core
MCP · common IR
mcp-api-adapterOpenAPIGraphQLgRPCWebSocketCLI

ESB lesson, agent era: N×M protocol integrations collapse to N+M. Streams is how the fabric reacts to webhooks, NATS topics, cron, and live sockets — not only interactive MCP clients.

Durable Objects · Kubernetes

Infinite session scale without losing the audit trail

SubscriptionDO holds the event buffer and significance state. AgentSessionDO runs the wake — with Audit, Inference, and TrainingData sidecars. Same mental model on Cloudflare Workers or Kubernetes HPA.

  • Virtual-key budgets

    Keys bind on session create and expire on destroy — every autonomous wake has a hard spend ceiling through clawql-inference.

  • WORM on every action

    Reactive ingest never waits on the model. Forensic reconstruction is a hash chain you own, not a vendor console export.

  • Training emission

    Optional RTP + OpenBenchTrace export with consent scopes — the Intelligence Flywheel without surrendering sovereignty.

Positioning

The pattern labs built for themselves — as a platform

Stripe Minions, Anthropic Managed Agents, and OpenAI Agents SDK all implement event → context → reason → tools → audit. ClawQL Streams is that pattern with operator-owned infrastructure.

ClawQL Streams compared to managed agent runtimes
CapabilityStripe MinionsManaged AgentsAgents SDKClawQL Streams
SovereigntyInternal onlyAnthropic-hostedManaged runtimeSelf-host or Cloudflare
TriggersSlack reactionsCron / APIPipeline / APIWebSocket · NATS · webhook · cron
ToolsCustom ToolshedBuilt-in + customBuilt-in + customAny MCP surface via Protocol Fabric
AuditInternalProvider logsProvider logsOperator-owned WORM Merkle trail
ScaleCustomSession hoursManagedDurable Objects or K8s HPA
Specifications

Draft specs live on docs.clawql.com

Streams is shipping as an opt-in coordination layer over packages you already deploy — inference, payments, Ouroboros, NATS, and WORM memory.

Put an event loop under your agents

Start a trial on the Agentic Gateway today. When Streams lands, the same MCP endpoint, vault, and WORM trail become the autonomous runtime — not a second product.