OpenWork Architecture

Design principle: Predictable > Clever

OpenWork optimizes for predictability over "clever" auto-detection. Users should be able to form a correct mental model of what will happen.

Guidelines:

Prefer explicit configuration (a single setting or env var) over heuristics.
Auto-detection is acceptable as a convenience, but must be:
- explainable (we can tell the user what we tried)
- overrideable (one obvious escape hatch)
- safe (no surprising side effects)
When a prerequisite is missing, surface the exact failing check and a concrete next step.

Example: Docker-backed sandboxes (desktop)

When enabling Docker-backed sandbox mode, prefer an explicit, single-path override for the Docker client binary:

OPENWORK_DOCKER_BIN (absolute path to docker)

This keeps behavior predictable across environments where GUI apps do not inherit shell PATH (common on macOS).

Auto-detection can exist as a convenience, but should be tiered and explainable:

Honor OPENWORK_DOCKER_BIN if set.
Try the process PATH.
On macOS, try the login PATH from /usr/libexec/path_helper.
Last-resort: try well-known locations (Homebrew, Docker Desktop bundle) and validate the binary exists.

The readiness check should be a clear, single command (e.g. docker info) and the UI should show the exact error output when it fails.

Minimal use of Tauri

We move most of the functionality to the openwork server which interfaces mostly with FS and proxies to opencode.

Filesystem mutation policy

OpenWork should route filesystem mutations through the OpenWork server whenever possible.

Why:

the server is the one place that can apply the same behavior for both local and remote workspaces
server-routed writes keep permission checks, approvals, audit trails, and reload events consistent
Tauri-only filesystem mutations only work in desktop host mode and break parity with remote execution

Guidelines:

Any UI feature that changes workspace files or config should call an OpenWork server endpoint first.
Local Tauri filesystem commands are a host-mode fallback, not the primary product surface.
If a feature cannot yet write through the OpenWork server, treat that as an architecture gap and close it before depending on direct local writes.
Reads can fall back locally when necessary, but writes should be designed around the OpenWork server path.

Agent authority map

When OpenWork is edited from openwork-enterprise, architecture and runtime behavior should be sourced from this document.

Entry point	Role	Architecture authority
`openwork-enterprise/AGENTS.md`	OpenWork Factory multi-repo orchestration	Defers OpenWork runtime flow, server-vs-shell ownership, and filesystem mutation behavior to `_repos/openwork/ARCHITECTURE.md`.
`openwork-enterprise/.opencode/agents/openwork-surgeon.md`	Surgical fix agent for `_repos/openwork`	Uses `_repos/openwork/ARCHITECTURE.md` as the runtime and architecture source of truth before changing product behavior.
`_repos/openwork/AGENTS.md`	Product vocabulary, audience, and repo-local development guidance	Refers to `ARCHITECTURE.md` for runtime flow, server ownership, and architectural boundaries.
Skills / commands / agents that mutate workspace state	Capability layer on top of the product runtime	Should assume the OpenWork server path is canonical for workspace creation, config writes, `.opencode/` mutation, and reload signaling.

Agent access to server-owned behavior

Agents, skills, and commands should model the following as OpenWork server behavior first:

workspace creation and initialization
writes to .opencode/, opencode.json, and opencode.jsonc
OpenWork workspace config writes (.opencode/openwork.json)
workspace template export/import, including shareable .opencode/** files and opencode.json state
workspace template starter-session materialization from portable blueprint config (not copied runtime session history)
share-bundle publish/fetch flows used by OpenWork template links
reload event generation after config or capability changes
other filesystem-backed capability changes that must work across desktop host mode and remote clients

Tauri or other native shell behavior remains the fallback or shell boundary for:

file and folder picking
reveal/open-in-OS affordances
updater and window management
host-side process supervision and native runtime bootstrapping

If an agent needs one of the server-owned behaviors above and only a Tauri path exists, treat that as an architecture gap to close rather than a parallel capability surface to preserve.

Reload-required flow

OpenWork uses a single reload-required flow for changes that only take effect when OpenCode restarts.

Key pieces:

createSystemState() owns the raw queued-reload state.
reloadPending() means a reload is currently queued for the active workspace.
markReloadRequired(reason, trigger) queues the reload and records the source that caused it.
app.tsx exposes reloadRequired(...sources) as a small helper for UI filtering. It is used to decide whether the shared reload popup should show for a given trigger type.

Use this flow when a change mutates startup-loaded OpenCode inputs, for example:

opencode.json
.opencode/skills/**
.opencode/agents/**
.opencode/commands/**
MCP definitions or plugin lists that OpenCode only loads at startup

Do not invent a separate reload banner per feature. New UI that needs restart semantics should:

perform the config or filesystem mutation
call markReloadRequired(...)
rely on the shared reload popup to explain and execute the restart path

Current examples that should use this shared flow include MCP changes, auto context compaction, default model changes, authorized folder updates, plugin changes, and other opencode.json writes.

opencode primitives

how to pick the right extension abstraction for @opencode

opencode has a lot of extensibility options: mcp / plugins / skills / bash / agents / commands

mcp use when you need authenticated third-party flows (oauth) and want to expose that safely to end users good fit when "auth + capability surface" is the product boundary downside: you're limited to whatever surface area the server exposes
bash / raw cli use only for the most advanced users or internal power workflows highest risk, easiest to get out of hand (context creep + permission creep + footguns) great for power users and prototyping, terrifying as a default for non-tech users
plugins use when you need real tools in code and want to scope permissions around them good middle ground: safer than raw cli, more flexible than mcp, reusable and testable basically "guardrails + capability packaging"
skills use when you want reliable plain-english patterns that shape behavior best for repeatability and making workflows legible pro tip: pair skills with plugins or cli (i literally embed skills inside plugins right now and expose commands like get_skills / retrieve)
agents use when you need to create tasks that are executed by different models than the main one and might have some extra context to find skills or interact with mcps.
commands / commands that trigger tools

These are all opencode primitives you can read the docs to find out exactly how to set them up.

Core Concepts of OpenWork

uses all these primitives
uses native OpenCode commands for reusable flows (markdown files in .opencode/commands)
adds a new abstraction "workspace" is a project folder and a simple .json file that includes a list of opencode primitives that map perfectly to an opencode workdir (not fully implemented)
- openwork can open a workpace.json and decide where to populate a folder with thse settings (not implemented today

Repository/component map

/apps/app/: OpenWork app UI (desktop/mobile/web client experience layer).
/apps/desktop/: Tauri desktop shell that hosts the app UI and manages native process lifecycles.
/apps/server/: OpenWork server (API/control layer consumed by the app).
/apps/orchestrator/: OpenWork orchestrator CLI/daemon. In start/serve host mode it manages OpenWork server + OpenCode + opencode-router; in daemon mode it manages workspace activation and OpenCode lifecycle for desktop runtime.
/apps/opencode-router/: first-party messaging bridge (Slack/Telegram) and directory router.
/apps/share/: share-link publisher service for OpenWork bundle imports.
/ee/apps/landing/: OpenWork landing page surfaces.
/ee/apps/den-web/: Den web UI for sign-in, worker creation, and future user-management flows.
/ee/apps/den-controller/: Den controller API that provisions/spins up worker runtimes.
/ee/apps/den-worker-proxy/: proxy layer that keeps Daytona API keys server-side, refreshes signed worker preview URLs, and forwards worker traffic so users do not manage provider keys directly.
/ee/apps/den-worker-runtime/: worker runtime packaging (including Docker/runtime artifacts) deployed to Daytona sandboxes.

Core Architecture

OpenWork is a client experience that consumes OpenWork server surfaces.

OpenWork supports two product runtime modes for users:

desktop
web/cloud (also usable from mobile clients)

OpenWork therefore has two runtime connection modes:

Mode A - Desktop

OpenWork runs on a desktop/laptop and can host OpenWork server surfaces locally.
The OpenCode server runs on loopback (default 127.0.0.1:4096).
The OpenWork server also defaults to loopback-only access. Remote sharing is an explicit opt-in that rebinds the OpenWork server to 0.0.0.0 while keeping OpenCode on loopback.
OpenWork UI connects via the official SDK and listens to events.
openwork-orchestrator is the CLI host path for this mode.

Mode B - Web/Cloud (can be mobile)

User signs in to hosted OpenWork web/app surfaces (including mobile browser/client access).
User launches a cloud worker from hosted control plane.
OpenWork returns remote connect credentials (/w/ws_* URL + access token).
User connects from OpenWork app using Add a worker -> Connect remote.

This model keeps the user experience consistent across self-hosted and hosted paths while preserving OpenCode parity.

Mode A composition (Tauri shell + local services)

/apps/app/ runs as the product UI; on desktop it is hosted inside /apps/desktop/ (Tauri webview).
/apps/desktop/ exposes native commands (engine_*, orchestrator_*, openwork_server_*, opencodeRouter_*) to start/stop local services and report status to the UI.
Runtime selection in desktop:
- openwork-orchestrator (default): Tauri launches openwork daemon run and uses it for workspace activation plus OpenCode lifecycle.
- direct: Tauri starts OpenCode directly.
In both desktop runtimes, OpenWork server (/apps/server/) is the API surface consumed by the UI; it is started with the resolved OpenCode base URL and proxies OpenCode and opencode-router routes.
Desktop-launched OpenCode credentials are always random, per-launch values generated by OpenWork. OpenCode stays on loopback and is intended to be reached through OpenWork server rather than exposed directly.
opencode-router is optional in desktop host mode and is started as a local service when messaging routes are enabled.

/apps/app UI
    |
    v
/apps/desktop (Tauri shell)
    |
    +--> /apps/orchestrator (daemon or start/serve host)
    |          |
    |          v
    |        OpenCode
    |
    +--> /apps/server (OpenWork API + proxy surface)
    |          |
    |          +--> OpenCode
    |          +--> /apps/opencode-router (optional)
    |
    +--> /apps/opencode-router (optional local child)

Mode B composition (Web/Cloud services)

/ee/apps/den-web/ is the hosted web control surface (sign-in, worker create, upcoming user management).
/ee/apps/den-controller/ is the cloud control plane API (auth/session + worker CRUD + provisioning orchestration).
/ee/apps/den-worker-runtime/ defines the runtime packaging and boot path used inside cloud workers (including Docker/snapshot artifacts and openwork serve startup assumptions).
/ee/apps/den-worker-proxy/ fronts Daytona worker preview URLs, refreshes signed links with provider credentials, and proxies traffic to the worker runtime.
The OpenWork app (desktop or mobile client) connects to worker OpenWork server surfaces via URL + token (/w/ws_* when available).

/ee/apps/den-web
    |
    v
/ee/apps/den-controller
    |
    +--> Daytona/Render provisioning
    |        |
    |        v
    |      /ee/apps/den-worker-runtime -> openwork serve + OpenCode
    |
    +--> /ee/apps/den-worker-proxy (signed preview + proxy)

OpenWork app/mobile client
    -> Connect remote (URL + token)
    -> worker OpenWork server surface

OpenCode Router (Messaging Bridge)

OpenWork can expose a workspace through Slack and Telegram via opencode-router.

opencode-router is a local bridge that receives messages from messaging adapters and forwards them into OpenCode sessions.
The core routing key is (channel, identityId, peerId) -> directory.
Bindings and sessions are persisted locally so a chat can continue against the same workspace directory.
The router keeps one OpenCode client per directory and one event subscription per active directory.

Runtime shape

Telegram bot / Slack app
          |
          v
+---------------------------+
| adapter per identity      |
| (telegram/slack, id)      |
+---------------------------+
          |
          v
 (channel, identityId, peerId)
          |
          v
+---------------------------+
| router state              |
| - bindings                |
| - sessions                |
| - identity defaults       |
+---------------------------+
          |
          v
+---------------------------+
| directory resolution      |
| binding -> session ->     |
| identity dir -> default   |
+---------------------------+
          |
          v
+---------------------------+
| OpenCode client per dir   |
+---------------------------+
          |
          v
   OpenCode session.prompt
          |
          v
 reply back to same chat

Directory scoping

OpenWork optimizes for a predictable routing boundary.

The router starts with a single workspace root (serve <path> or OPENCODE_DIRECTORY).
Routed directories may be absolute or relative, but they must stay inside that root.
Relative chat commands like /dir foo/bar resolve against the router root.
Directories outside the root are rejected instead of silently accepted.

This keeps the mental model simple: one router instance owns one root, and chat routing stays inside that tree.

Local HTTP control plane

The router exposes a small local HTTP API for host-side configuration and dispatch:

/health
/identities/telegram
/identities/slack
/bindings
/send

OpenWork server proxies /opencode-router/* and /w/:id/opencode-router/* to that local router API.

Workspace-scoped behavior in OpenWork

OpenWork server treats messaging identities as workspace-scoped.

Each workspace maps to a normalized router identity id.
When the app/server upserts a Telegram or Slack identity for a workspace, it also persists that workspace path as the identity default directory.
Binding and send APIs are filtered through that workspace identity, so the UI talks about "this workspace" even though the underlying router can track multiple directories.

Client UI / phone / web
          |
          v
+---------------------------+
| OpenWork server           |
| /workspace/:id/...        |
| /w/:id/opencode-router/*  |
+---------------------------+
          |
          | workspace-scoped identity id
          | workspace.path as default dir
          v
+---------------------------+
| local opencode-router     |
| HTTP control plane        |
+---------------------------+
          |
          v
+---------------------------+
| OpenCode                  |
+---------------------------+

CLI mental model

opencode-router serve /root/workspaces

message arrives
  -> lookup (channel, identityId, peerId)
  -> resolve directory
  -> ensure directory is under /root/workspaces
  -> reuse/create OpenCode session for that directory
  -> stream reply back to Slack or Telegram

Multiple workspaces: what works today

There are two layers here, and they matter:

The router core can multiplex multiple directories.
The current desktop embedding still runs a single router child process with a single root.

Current desktop shape

workspace A ----\
workspace B -----+--> OpenWork server knows all workspaces
workspace C ----/

                   but desktop starts one router child
                   with one configured root

                 +-------------------------------+
                 | opencode-router               |
                 | root: runtime active workspace|
                 | clients: many dirs under root |
                 +-------------------------------+

Practical consequences:

If multiple workspaces live under one shared parent root, one router can serve them all.
If workspaces live in unrelated roots, directories outside the active router root are rejected.
OpenWork server is already multi-workspace aware.
Desktop router management is still effectively single-root at a time.
On desktop, the file watcher follows the runtime-connected workspace root, not just the workspace currently selected in the UI.

Terminology clarification:

selected workspace is a UI concept: the workspace the user is currently viewing and where compose/config actions should target.
runtime active workspace is a backend concept: the workspace the local server/orchestrator currently reports as active.
watched workspace is the desktop-host/runtime concept for which workspace root local file watching is currently attached to.
These states must be treated separately. UI selection can change without implying that the backend has switched roots yet.
In practice, selected workspace and runtime active workspace often converge once the user sends work, but they are allowed to diverge briefly while the UI is browsing another workspace.

Desktop local OpenWork server ports:

Desktop-hosted local OpenWork server instances do not assume a fixed 8787 port.
Each workspace gets a persistent preferred localhost port in the 48000-51000 range.
On restart, desktop tries to reuse that workspace's saved port first.
If that port is unavailable, desktop picks another free port in the same range and avoids ports already reserved by other known workspaces.

Shared-root case

router root: /Users/me/projects

  /Users/me/projects/a   OK
  /Users/me/projects/b   OK
  /Users/me/projects/c   OK

Unrelated-root case

router root: /Users/me/projects/a

  /Users/me/projects/a   OK
  /Users/me/other/b      rejected
  /tmp/c                 rejected

This is intentional for now: predictable scoping beats clever cross-root auto-routing.

Cloud Worker Connect Flow (Canonical)

Authenticate in OpenWork Cloud control surface.
Launch worker (with checkout/paywall when needed).
Wait for provisioning and health.
Generate/retrieve connect credentials.
Connect in OpenWork app via deep link or manual URL + token.

Technical note:

Default connect URL should be workspace-scoped (/w/ws_*) when available.
Technical diagnostics (host URL, worker ID, raw logs) should be progressive disclosure, not default UI.

Web Parity + Filesystem Actions

The browser runtime cannot read or write arbitrary local files. Any feature that:

reads skills/commands/plugins from .opencode/
edits SKILL.md / command templates / opencode.json
opens folders / reveals paths

must be routed through a host-side service.

In OpenWork, the long-term direction is:

Use the OpenWork server (/apps/server/) as the single API surface for filesystem-backed operations.
Treat Tauri-only file operations as an implementation detail / convenience fallback, not a separate feature set.

This ensures the same UI flows work on desktop, mobile, and web clients, with approvals and auditing handled centrally.

OpenCode Integration (Exact SDK + APIs)

OpenWork uses the official JavaScript/TypeScript SDK:

Package: @opencode-ai/sdk/v2 (UI should import @opencode-ai/sdk/v2/client to avoid Node-only server code)
Purpose: type-safe client generated from OpenAPI spec

Engine Lifecycle

Start server + client (Host mode)

Use createOpencode() to launch the OpenCode server and create a client.

import { createOpencode } from "@opencode-ai/sdk/v2";

const opencode = await createOpencode({
  hostname: "127.0.0.1",
  port: 4096,
  timeout: 5000,
  config: {
    model: "anthropic/claude-3-5-sonnet-20241022",
  },
});

const { client } = opencode;
// opencode.server.url is available

Connect to an existing server (Client mode)

import { createOpencodeClient } from "@opencode-ai/sdk/v2/client";

const client = createOpencodeClient({
  baseUrl: "http://localhost:4096",
  directory: "/path/to/project",
});

Health + Version

client.global.health()
- Used for startup checks, compatibility warnings, and diagnostics.

Event Streaming (Real-time UI)

OpenWork must be real-time. It subscribes to SSE events:

client.event.subscribe()

The UI uses these events to drive:

streaming assistant responses
step-level tool execution timeline
permission prompts
session lifecycle changes

Sessions (Primary Primitive)

OpenWork maps a "Task Run" to an OpenCode Session.

Core methods:

client.session.create()
client.session.list()
client.session.get()
client.session.messages()
client.session.prompt()
client.session.abort()
client.session.summarize()

Files + Search

OpenWork's file browser and "what changed" UI are powered by:

client.find.text()
client.find.files()
client.find.symbols()
client.file.read()
client.file.status()

Permissions

OpenWork must surface permission requests clearly and respond explicitly.

Permission response API:
- client.permission.reply({ requestID, reply }) (where reply is once | always | reject)

OpenWork UI should:

Show what is being requested (scope + reason).
Provide choices (allow once / allow for session / deny).
Post the response to the server.
Record the decision in the run's audit log.

Config + Providers

OpenWork's settings pages use:

client.config.get()
client.config.providers()
client.auth.set() (optional flow to store keys)

Extensibility - Skills + Plugins

OpenWork exposes two extension surfaces:

Skills (OpenPackage)
- Installed into .opencode/skills/*.
- OpenWork can run opkg install to pull packages from the registry or GitHub.
Plugins (OpenCode)
- Plugins are configured via opencode.json in the workspace.
- The format is the same as OpenCode CLI uses today.
- OpenWork should show plugin status and instructions; a native plugin manager is planned.

Engine reload (config refresh)

OpenWork server exposes POST /workspace/:id/engine/reload.
It calls OpenCode POST /instance/dispose with the workspace directory to force a config re-read.
Use after skills/plugins/MCP/config edits; reloads can interrupt active sessions.
Reload requests follow OpenWork server approval rules.

OpenPackage Registry (Current + Future)

Today, OpenWork only supports curated lists + manual sources.
Publishing to the official registry currently requires authentication (opkg push + opkg configure).
Future goals:
- in-app registry search
- curated list sync (e.g. Awesome Claude Skills)
- frictionless publishing without signup (pending registry changes)

Projects + Path

client.project.list() / client.project.current()
client.path.get()

OpenWork conceptually treats "workspace" as the current project/path.

Optional TUI Control (Advanced)

The SDK exposes client.tui.* methods. OpenWork can optionally provide a "Developer Mode" screen to:

append/submit prompt
open help/sessions/themes/models
show toast

This is optional and not required for non-technical MVP.

Folder Authorization Model

OpenWork enforces folder access through two layers:

OpenWork UI authorization
- user explicitly selects allowed folders via native picker
- OpenWork remembers allowed roots per profile
OpenCode server permissions
- OpenCode requests permissions as needed
- OpenWork intercepts requests via events and displays them

Rules:

Default deny for anything outside allowed roots.
"Allow once" never expands persistent scope.
"Allow for session" applies only to the session ID.
"Always allow" (if offered) must be explicit and reversible.

Open Questions

Best packaging strategy for Host mode engine (bundled vs user-installed Node/runtime).
Best remote transport for mobile client (LAN only vs optional tunnel).
Scheduling API surface (native in OpenCode server vs OpenWork-managed scheduler).

FilesExpand file tree

ARCHITECTURE.md

Latest commit

History