Bob the IT guy. Bob might already use opencode, he can setup agents and workflows and share them with his team. The only thing he needs is a way to share this. The way he does is by using OpenWork and creating "workpaces".
Susan in accounting
Susan in accounting doesn't use opencode. She certaintly doesn't paly aorund to create workflow create agents. She wants something that works. Openwork should be given to give her a good taste of what she can do. We should also eventually guide ther to:
- creating her own skills
- adding custom MCP / login into mcp oauth servers through ui)
- adding skills from a list of skills
- adding plugins from a list of plugins
- create her own commands
- knowledge worker: "Do this for me" workflows with guardrails.
- Mobile-first user: start/monitor tasks from phone.
- Power user: wants UI parity + speed + inspection.
- Admin/host: manages a shared machine + profiles.
- < 5 minutes to first successful task on fresh install.
-
80% task success without terminal fallback.
- Permission prompts understood/accepted (low confusion + low deny-by-accident).
- UI performance: 60fps; <100ms interaction latency; no jank.
OpenWork must feel like "OpenCode, but for everyone."
- A Task = a user-described outcome.
- A Run = an OpenCode session + event stream.
OpenWork provides a first-class plan UI:
- Plan is generated before execution (editable).
- Plan is updated during execution (step status + timestamps).
- Plan is stored as a structured artifact attached to the session (JSON) so it's reconstructable.
Implementation detail:
- The plan is represented in OpenCode as structured
parts(or a dedicated "plan message") and mirrored in OpenWork.
- Each tool call becomes a step row with:
- tool name
- arguments summary
- permission state
- start/end time
- output preview
Artifacts are user-visible outputs:
- files created/modified
- generated documents/spreadsheets/presentations
- exported logs and summaries
OpenWork lists artifacts per run and supports open/share/download.
Every run provides an exportable audit log:
- prompts
- plan
- tool calls
- permission decisions
- outputs
- premium, calm, high-contrast
- subtle motion, springy transitions
- zero "developer vibes" in default mode
- 60fps animations
- <100ms input-to-feedback
- no blocking spinners (always show progress state)
- bottom navigation
- swipe gestures (dismiss, approve, cancel)
- haptics for major events
- adaptive layouts (phone/tablet)
- WCAG 2.1 AA
- reduced motion mode
- screen-reader labels for steps + permissions
use the design from ./design.ts that is your core reference for building the entire ui
- Host vs Client selection
- workspace selection (Host)
- connect to host (Client)
- provider/model setup
- first-run "hello world" task
- create task
- plan preview and edit
- run with streaming updates
- pause/resume/cancel
- show artifacts and summaries
- clear prompts with "why"
- allow once/session
- audit of decisions
- save a task as a command
- arguments + quick run
- schedule command runs
- notify on completion
- User installs OpenWork.
- App launches.
- App shows "Choose mode: Host / Client".
- Host: start local OpenCode via SDK.
- Client: connect flow to an existing host.
- Welcome + safety overview.
- Workspace folder selection.
- Allowed folders selection (can be multiple).
- Provider/model configuration.
global.health()check.- Run a test session using
session.create()+session.prompt(). - Success + sample commands.
- User selects "Client".
- UI explains it connects to a trusted host.
- User scans QR code shown on host device.
- Client verifies connection with
global.health(). - Client can now list sessions and monitor runs.
- UI pings
global.health(). - If unhealthy:
- Host: attempt restart via
createOpencode(). - Client: show reconnect + diagnostics.
- Host: attempt restart via
- User types goal.
- OpenWork generates plan (structured).
- User approves.
- Create session:
session.create(). - Send prompt:
session.prompt(). - Subscribe to events:
event.subscribe(). - Render streaming output + steps.
- Show artifacts.
- Wizard collects goal, constraints, outputs.
- Plan preview with "risky step" highlights.
- Run execution with progress UI.
- User attaches files.
- OpenWork injects context into session.
- Execute prompt.
- Event indicates permission request.
- UI modal shows request.
- User chooses allow/deny.
- UI calls
client.permission.reply({ requestID, reply }). - Run continues or fails gracefully.
- User clicks "Stop".
- UI calls
client.session.abort({ sessionID }). - UI marks run stopped.
- User taps "Summarize".
- UI calls
client.session.summarize({ sessionID }). - Summary displayed as an artifact.
- UI calls
session.list(). - Tap a session to load
session.messages(). - UI reconstructs plan and steps.
- User searches:
find.text(). - Open file:
file.read(). - Show changed files:
file.status().
- Save a plan + prompt as a command.
- Re-run command creates a new session.
- separate profiles
- separate allowed folders
- separate providers/keys