Coordinate autonomous AI work
like a control plane should.

Projects & work

Tenants, projects, work items as a DAG (Epic → Feature → Task → Subtask), with cycle-detected dependencies.

Workers & dispatch

Versioned Worker definitions (system prompt, permissions, gated tools, budgets). Reconciler-driven dispatch to healthy adapters.

Workflows

Step-DAG workflows: task, decision, approval, parallel, recover. Visual React Flow editor. Real-time step transitions.

Recovery engine

Default 6-step recovery workflow (capture → summarize → preserve → review → plan → resume) with checkpoint replay and bounded auto-relax.

Rego policy

OPA v1 at the dispatch decision point. Narrowest-scope-first evaluation, full Rego trace per decision for explainability.

Telemetry & cost

OTel → Grafana stack (Tempo + Loki + VictoriaMetrics). Per-tenant cost roll-up (Tenant → Project → Task → Execution) with streamed usage events.

Single container + per-workflow runtime containers

The whole stack — Postgres, NATS, telemetry, control plane — runs in one container with orchicon container (PID-1 supervisor). Dev + prod as two containers, your data preserved.

Worker executions run in an isolated container per workflow run, drawn from a warm pool (orchicon runtime-daemon on the host) so dispatch never cold-starts — and reset to pristine after each run so no state crosses boundaries. Running as your user — no root, fully ephemeral, torn down at completion. A run mounts only the adapter it actually needs: a plane using the built-in engine mounts nothing at all, and the host's own adapter install is mounted when a run asks for one — never baked into an image.

A complete terminal client

orch is not a companion viewer — it is the whole product in a terminal, with read and write parity across every domain. Create, edit, publish, approve, cancel, retry, run bulk operations and watch live streams without leaving the keyboard. Boxed multi-line composer with slash commands and a command palette, a slide-out diff sidebar, click-to-copy on your own messages, and 43 themes — including true transparency, where your terminal shows through and the text adapts to it so it stays readable.

Ask Orchicon — three modes that are actually enforced

Brainstorm designs with you, Iteration works alongside you, Quick Work hands a defined job to a worker. The boundary is enforced by the platform rather than requested in a prompt: a mode's disallowed tools are withheld from it and refused at the point of execution, so "Brainstorm will not edit your files" is not something a model can be talked out of. The policy is adapter-agnostic, so a new runtime inherits it automatically.

Automation that proposes work, and a place to judge it

Put Orchicon on a schedule and a research crew surveys the world on its own — planning, verifying against external evidence, and synthesizing what it finds into feature proposals. Proposals land in the Idea Cloud, a triage board separate from real work, each carrying its evidence and the run that produced it. Promote one to turn it into work items, or dismiss it and it is remembered as rejected so it never comes back.

Cost is a constraint, not a report

Every execution runs under configurable ceilings — tokens, dollars, wall-clock, tool calls — with defaults calibrated from real usage. When a session gets expensive, Orchicon trims its own context and keeps working instead of re-sending an ever-growing history, and genuinely hard limits stop a runaway worker outright. Cache re-sends no longer count as work, so a long task cannot trip its budget by re-reading what it already knows.

Work that cannot be silently stranded

Work items form a DAG (Epic → Feature → Task → Subtask) with cycle-detected dependencies, and every run is a workflow — so an item with no workflow bound cannot be scheduled, and the platform says so when you create it rather than failing later. When you do need to unblock a backlog, binding a workflow and a runtime image across a whole selection is one gesture, not a form per item.

Recovery, policy and telemetry around every run

Failures are recoverable by default: the recovery flow captures, summarises, preserves, reviews, plans and resumes, with bounded auto-relax and escalation. Policy is evaluated in Rego at the dispatch decision point, with a full trace per decision. Telemetry streams over OpenTelemetry into the Grafana stack, and cost rolls up Tenant → Project → Task → Execution.