ORCA is the orchestration mesh that coordinates LLM agents across long-running operational workflows. Persistent context, structured handoffs, full observability, running on your own infrastructure.
Spawn a sub-agent and the parent loses sight of what happened. Findings disappear. Decisions get re-litigated. Tasks duplicate.
Agents return whatever prose they feel like. The orchestrator has to re-parse and re-interpret. Brittle, ad-hoc, unreproducible.
Existing orchestration platforms force cloud dependencies and proprietary state stores. Self-hosters end up with shell scripts and prayer.
ORCA treats agents as specialized workers with declared scope, explicit handoff contracts, and shared persistent context. The orchestrator routes work, never re-parses.
Each agent declares scope, capabilities, and an explicit return contract. The orchestrator dispatches based on declared fit, not prompt-engineering luck.
agent: woocommerce-ops
scope: order-flow,inventory,refunds
returns: { status, payload, evidence }
context: persistentShared context layer persists across agent boundaries. Findings from one agent become inputs to the next without re-discovery cost.
context.write(agent_a, finding)
context.read(agent_b, scope)
→ replayable, observable, durableRuns on a Tailnet, your KVMs, your hardware. No proprietary state store. Bring your own LLM endpoints (OpenAI, Anthropic, local Ollama, anything OpenAI-compatible).
Full trace of which agent ran when, with what inputs, what context it pulled, what it returned. Replayable. Auditable. Debuggable.
Coordinate inventory, order, support, and marketing agents across multiple WooCommerce / Shopify stores from a single mesh.
Agents that audit, patch, and report across your Tailnet hosts. No cloud agent platform required.
Multi-agent investigations with persistent ledgers, verifier passes, and durable on-disk artifacts.
ORCA started as internal tooling for running a multi-store ecommerce operation and a self-hosted infra stack. Existing agent platforms were either too cloud-dependent, too generic, or too brittle for production ops work.
We rebuilt it from first principles: specialists with declared scope, persistent context that survives handoffs, full observability, and zero proprietary lock-in. Currently in private beta with a small group of self-hosted operators.
We are onboarding a small number of self-hosted operators each month. Tell us what you would orchestrate.