Everything you need to ship AI agents to production.
A runtime, an SDK, an observability suite, and a governance layer. One platform — the boring parts done right.
Durable execution that survives anything.
Long-running agents need long-running guarantees. circuitnova checkpoints every step, replays on crash, and retries on the failures you choose. No glue code, no babysitting.
- Step-level checkpointing
- Configurable retries and fallbacks
- Resume from anywhere
- Horizontal scale by default
await runtime.run(researchAgent, {
retries: { max: 3, backoff: "exponential" },
checkpoint: "per-step",
timeout: "30m",
});Framework adapters, not framework prisons.
Use LangGraph for orchestration, CrewAI for personas, AutoGen for collaboration — or all three in one workflow. circuitnova wraps each as a first-class runtime primitive.
- LangGraph, CrewAI, AutoGen
- OpenAI, Anthropic, Google
- Raw HTTP / SDK agents
- Bring your own framework
agent({
framework: "crewai",
crew: writerCrew,
}).pipe(agent({
framework: "langgraph",
graph: editorGraph,
}));Traces, replays, diffs — built in.
Click any run. See every prompt, tool call, and token. Replay it with new inputs. Diff two runs side by side. The debugging surface you wish OpenTelemetry shipped.
- Per-step traces with token counts
- Run replay with input overrides
- Side-by-side run diffs
- Cost attribution by agent, tool, and run
# replay a failed run with patched inputs
$ circuitnova replay run_8c2a \
--override 'plan.depth=2'
✓ replayed in 3.8s • $0.0091 cheaperGovernance ready for the audit you knew was coming.
Role-based access, audit logs, PII redaction, model allow-lists, and policy hooks. SOC 2 in flight. EU AI Act compliance scaffolding ships in Q3.
- RBAC + SSO (SAML, OIDC)
- Immutable audit log
- Model and tool allow-lists
- EU AI Act risk classification
policy({
models: ["anthropic/claude-*", "openai/gpt-4*"],
redact: ["email", "credit_card"],
maxCostPerRun: "$0.50",
});Open standards. No proprietary tax.
Native Model Context Protocol (MCP) and Agent-to-Agent (A2A) support. Any tool, any agent, any vendor. We bet on the ecosystem because lock-in is the problem we're solving.
- MCP server + client built in
- A2A inter-agent messaging
- OpenTelemetry export
- Webhooks for everything
// expose any agent as an MCP server
mcp.serve(researchAgent, {
port: 8080,
capabilities: ["tools", "resources"],
});