Agents chat. Workers ship.

Orchicon is a self-hosted control plane for autonomous AI work. It keeps the backlog, assigns each task to a team of AI Workers, runs them through review loops and human approvals, holds them to a budget, and picks up where they left off when something breaks.

Prefer a classic harness? Orchicon works like any other one too. Open Ask Orchicon in Iteration mode and pair with it directly in your repo, from the browser or the terminal. Add workflows, budgets and schedules only when you need them.

One command installs the whole stack: control plane, web app, terminal client, Postgres, NATS and a Tempo/Loki/VictoriaMetrics telemetry plane.

Desktop, terminal and web clients. Native on macOS, Windows and Linux.

Two full clients, one product: work in the web app or run orch in your terminal.

One product, three full clients

Desktop, terminal and web — every screen, every action, on macOS, Windows and Linux. The desktop app is a native shell that launches or connects to your plane; orch is read/write at parity over SSH or in a tmux pane; the web app ships inside the binary, so there is nothing else to host. Same data, same permissions, same seven workspaces.

localhost:8080 · Ask Orchicon
Ask Orchicon in the web app: a Brainstorm conversation showing tables of modes and platform pillars, with conversation folders on the right. Work items tree: epics, features and tasks with sequential order, branch names, merged pull requests and status. Visual workflow editor for the SDLC Non-human workflow, with versions and runs listed below. Schedules history: 367 past runs with durations, workflow and run identifiers. Idea Cloud: feature ideas from a research run, each with cited sources and Approve and Dismiss buttons.

    Seven workspaces in all three — one key each

    F1 Ask OrchiconConversations in folders, all three modes, with reasoning, diffs and live cost.
    F2 OverviewDashboard, telemetry and cost at a glance.
    F3 WorkTree and board views; create, sequence, and bulk-bind workflows.
    F4 ExecutionWorkflows, runs and executions, with a flow view of every step.
    F5 AutomationRecurring items, schedules and the Idea Cloud.
    F6 EnforcementPolicies, approvals and budgets.
    F7 ControlServers, adapters, settings and secrets.

    Keys you will use all day

    /
    Slash commands and the command palette
    ctrld
    Slide out the file diff sidebar
    space
    Mark items for a bulk action
    W
    Set workflow and runtime image on every marked item
    ctrly
    Stop a reply mid-stream
    • 43 themes in the terminal, including true transparency where your terminal shows through.
    • Mouse when you want it. Click to select, scroll, and click your own message to copy it.
    • Honest about cost. Context, tokens, cache hit rate and dollars sit under the composer.

    An agent waits for a prompt. A Worker shows up with a job.

    Agents are good at a single task in a single session, as long as someone is there to steer. A Worker has a role, a scope, tools it is allowed to use, a budget, and a process it answers to. The hard part is everything around the task: deciding what to build, checking the result, keeping costs sane, and not losing a day of work when a run falls over. That is the part Orchicon does.

    An agent on its ownWorkers under Orchicon
    Where work comes fromWhatever you type into the prompt today.A backlog of epics, features, tasks and subtasks with dependencies, run order and recurring schedules.
    How it gets doneOne model, one pass, one opinion.A workflow of specialised Workers (architect, engineer, reviewer, QA, DevOps) with review loops and human approval gates.
    Quality controlYou read the diff and hope.Reviewers can send work back, capped at a set number of loops. Every item gets an acceptance review recorded against the run.
    When it failsYou notice, then start over.Recovery captures, summarises, preserves, reviews, plans and resumes, escalating from L1 to L3 if it has to.
    CostYou find out on the invoice.Hard ceilings on tokens, dollars, wall-clock time and tool calls, rolled up by tenant, project, task and execution.
    SafetyThe prompt asks it nicely not to.Disallowed tools are withheld and refused at execution. Every dispatch passes a Rego policy with a full decision trace.
    IsolationRuns in your working copy.Each run gets its own container and git worktree, reset to pristine afterwards. Nothing leaks between runs.
    VisibilityTerminal scrollback.Traces, logs and metrics in the built-in telemetry explorer, an audit trail of every action, and a searchable history of every run.

    From an idea to a merged pull request

    This is the loop Orchicon was built around, and the one it uses to build itself. It is also the logo: six steps, and the last one lands. The work items and schedules in these screenshots are Orchicon's own backlog.

    1. 1

      Talk it through

      Brainstorm with Ask Orchicon. It reads your actual repository, so the plan is grounded in real files and line numbers.

    2. 2

      Turn it into work

      When a decision lands, it becomes a work item with scope and acceptance criteria detailed enough for a Worker to finish without coming back with questions.

    3. 3

      Bind a workflow

      Choose how it gets built: a full SDLC team with or without human gates, a local-models-only team, or a single quick job. Bind a whole selection in one gesture.

    4. 4

      Dispatch safely

      Reconcilers check policy and budgets, pull a warm container, check out a worktree and start the first Worker. No cold starts.

    5. 5

      Review, loop, approve

      Reviewers and QA send work back until it passes. You approve at the gates you chose. Every step transition streams live.

    6. 6

      Land it

      The branch is pushed, the PR is merged, and the acceptance review is saved on the work item. The next item in sequence starts on its own.

    Workflows: design the team that builds each task

    A workflow is a graph of Workers, approval gates and loop decisions. Reviewers can send work back a set number of times, humans step in only where you ask them to, and every published version is locked so a run always knows exactly which process it followed.

    Visual workflow editor with the SDLC Non-human workflow, a palette of steps and a properties panel.
    Drag it together in the browser, with undo, versions and a list of runs.
    The SDLC Human workflow as an eleven-step list in the orch terminal client.
    Read, edit and publish the same graph as a step list in orch.

    Ask Orchicon: three modes, enforced by the platform

    Ask Orchicon is the colleague who holds the whole system in their head. Its modes are not prompt suggestions. A mode's disallowed tools are withheld from the model and refused at execution, so Brainstorm cannot be talked into editing your files. The rule is adapter-agnostic, so every new runtime inherits it.

    Iteration is also how you use Orchicon like a regular harness: one Worker, your repo, you in the loop, with nothing else to set up.

    ModeWhat it doesWhat you get back
    BrainstormArchitecture, trade-offs, research and planning, grounded in your codebase. Read-only.Scoped work items with acceptance criteria
    IterationBranches, edits, runs tests and commits with you in the loop.Working commits and a pull request
    Quick WorkSpins up a temporary Worker and workflow, fires one run, cleans up after itself.The finished job, with nothing left behind

    Autonomous market research from a scheduled Worker

    A recurring Worker scans the market on a schedule, checks what it finds against real sources, and files evidence-backed feature ideas into the Idea Cloud. You spend five minutes triaging instead of an afternoon researching.

    1. 1Schedule itA recurring item with Outputs: ideas fires on your interval, inside your work window.
    2. 2Research itA research Worker surveys competitors, articles and releases, then verifies each claim.
    3. 3File itEach idea lands with its sources, its reasoning and the run that produced it.
    4. 4Triage itApprove turns it into a real work item. Dismiss sends it to Rejected, kept as history so it never resurfaces.
    The Idea Cloud: three feature ideas from one research run, including tenant-level spend governance, external secrets-manager integration and sub-agent delegation, each with cited sources and Approve and Dismiss buttons.

    Ideas you can check, not vibes

    • Cited. Every idea quotes its outside evidence with dates and links, so you can verify it in a click.
    • Traceable. Each one names the recurring item and run that spawned it.
    • Kept apart. Ideas live on their own board until you approve them, so your backlog stays yours.
    • Remembered. Rejected ideas stay readable and are never proposed again.

    The three ideas in the screenshot came from a single run of Orchicon researching its own market.

    Budgets: cost is a constraint, not a report

    Every execution runs under ceilings 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. Cache re-reads do not count as work, so a long task cannot trip its budget by re-reading what it already knows.

    TokensPer execution, per task, per project
    DollarsRolled up to the tenant, streamed live
    Wall-clockRunaway Workers stop outright
    Tool callsHang and loop guards on top

    Everything a crew of Workers needs

    Orchicon owns the parts that make autonomous work dependable and leaves the actual execution to a runtime you choose.

    AreaWhat you get
    Work itemsEpic, Feature, Task and Subtask form a dependency graph with cycle detection. Sequences, reordering, bulk actions, board and tree views, archive, and recurring fires.
    WorkersVersioned definitions with system prompt, permissions, gated tools and budgets. Lifecycle runs draft, published, deprecated, retired. The model is pinned by a human, with no silent failover.
    WorkflowsStep graphs with task, approval, decision, parallel, loop and recover steps. Visual editor, edit locks, immutable versions, retry-in-place and force-progress.
    Projects & contextproject_dir scopes every Worker to a repo. context_files injects files or whole directories into the prompt. Git strategy per project, worktrees always isolated.
    ExecutionReconcilers create executions and dispatch them to runtime adapters — in-process on the host, or in a per-run container from a warm pool.
    Plane controlA Servers page starts, stops and restarts the plane with all four connection states rendered honestly, over the same local mechanism the CLI uses.
    RuntimesA built-in engine that needs nothing installed, plus pluggable adapters over gRPC such as OpenCode. Buildable images from apt packages and toolchains, chosen per work item.
    RecoveryFailures are recoverable by default: capture, summarise, preserve, review, plan, resume. Checkpoint replay, bounded auto-relax and L1 to L3 escalation.
    GovernanceOPA/Rego policy at the dispatch decision, approvals, budgets, a full audit trail, AES-256-GCM secrets and Postgres row-level security per tenant.
    ReliabilityStall detection, liveness probes and nudges, repetition and text-loop guards, tool-hang redirects and an execution reaper.
    ObservabilityOpenTelemetry into Tempo traces, Loki logs and VictoriaMetrics metrics. Cost rolls up from execution to tenant, streamed live.
    IntegrationsA curated MCP catalog (filesystem, GitHub, Postgres, Playwright, Sentry, Slack and more) with one-click install and ${SECRET} references.
    AutomationRecurring items on interval schedules with day toggles and work windows. A research Worker proposes evidence-backed features into the Idea Cloud for you to approve or dismiss.
    DeploymentThe plane and its services are host processes — native on macOS, Linux and Windows, no WSL2 and no Docker for the plane. Containers exist only for workflow runs, on Docker or podman.

    Orchicon orchestrates. Runtimes execute.

    The control plane is a single Go binary. Reconcilers keep the world converged on what you declared, the way Kubernetes does for containers. Runtimes never touch Postgres or NATS directly; everything streams over gRPC.

    To run work somewhere new, implement orchicon.adapter.v1. The built-in engine is one implementation of it, not a special case. External adapters are mounted from your host and never bundled, which keeps Orchicon redistributable whatever an adapter's licence says.

    YouWeb app, orch terminal client, CLI and API (Protobuf + Connect: gRPC, REST, streaming)
    Control planeOne Go binary with the embedded web app
    ReconcilersSchedulerRego policyBudgetsRecoveryAudit
    State and signalsPostgres with row-level security, NATS, and an OpenTelemetry collector feeding Tempo, Loki and VictoriaMetrics
    RuntimesBuilt-in engine or adapters over orchicon.adapter.v1, each run in its own warm-pool container, run as your user, torn down at completion
    Native OrchiconOpenCodeClaude

    Pricing

    It is the same software in every column. What changes is the licence to use it commercially, and who you can ask for help.

    Personal

    Free

    For personal, non-commercial use.

    • The whole product: web app and terminal client
    • Workers, Workflows, Work Items, Recovery and the Idea Cloud
    • Runs entirely on your own hardware, under your own models and keys
    • No limits imposed by us on projects, runs or Workers
    Install Orchicon

    Small & medium business

    Get in touch

    For commercial use in a small or medium sized business.

    • Everything in Personal, licensed for commercial use
    • Support when you need a hand
    • Email us and we will get you set up
    [email protected]

    Enterprise

    Get in touch

    For larger organizations.

    • Everything in Small & medium business
    • We will talk through how you want to run it
    • Email us to start that conversation
    [email protected]

    Install in one command

    The installer downloads the orchicon binary and brings this machine's plane up as a host process: Postgres, NATS and the telemetry backends included. Containers are provisioned for workflow runs, not for the plane. Re-run it to update — it converges an existing instance and refuses a downgrade.

    $ curl -fsSL https://orchicon.dev/install | bash

    macOS: nothing to unblock — curl sets no quarantine bit, so this install runs without a Gatekeeper prompt. Only a browser download is marked; clear it on that one file with xattr -d com.apple.quarantine /path/to/orchicon. That does not disable Gatekeeper. Every release publishes SHA256SUMS and both installers verify the download against it.

    Requires curl and tar. No Docker needed for the plane; a container engine is used only for workflow runs.

    Common commands — add them to the line above

    • --instance <dev|test|prod> — which stack to act on: each has its own ports, data and host state. Every command that changes something asks unless you pass this.
    • --version <tag> — a specific released version instead of the latest.
    • --dry-run — show what would happen. Changes nothing.
    • --no-setup — the binary only: no services, no daemon, no plane.
    • --uninstall — stop this instance and remove the binaries. Your data is kept.
    • --purge — uninstall and destroy this instance's data. Irreversible.
    • --remove-binary — just the orchicon and orch executables; services, data and state are untouched.
    • --clean — upgrade: stop, remove the old binary, install the latest. Data kept.
    • --force-clean — rebuild from scratch for this instance. All of its data is lost.
    1. 1
      InstallRun the command for your platform, then add the flags you want — the common ones are listed beside it.
    2. 2
      Open the appThe control plane and web app come up on localhost:8080. Sign in with the built-in dev identity provider.
    3. 3
      Create a projectPoint it at a repository. Every Worker is scoped to that directory.
    4. 4
      Ask, then dispatchBrainstorm in Ask Orchicon, turn the result into a work item, bind a workflow and start it. Watch it in the browser or in orch.

    Building Orchicon itself? git clone, then make build and make run. Needs Go 1.26+ and Node 22+; a container engine (podman or Docker) is optional and only used for workflow runtime containers.

    Contact & community

    Follow along, ask a question, or show us what you built. These are the places we post, answer and ship from.