01Open-source AI workspace
Your open workspace for AI agents.
Connect your models and tools. Plan, act, and review every step in one desktop workbench.
No installer has been published yet. The build is public and reproducible — clone the repository and run the desktop build, or watch releases to be notified when the first one ships. Watch releases
The task on this page
- Repository
- acme/checkout-service
- Branch
- release/2.4.0
- Failing check
- unit-tests
Plan
StateWaiting for approval
- Re-run the failing check and read the assertionDone
- Round per currency in the total helperDone
- Re-run the check against the changed filesIn progress
- Draft the launch notes from the merged changesNot started
Open source
AGPL-3.0-or-later, public repository
Bring your models
Local runtimes, your own keys, your own subscription
Permissioned actions
Tools declare what they touch; you confirm
Desktop first
Local files, terminal, long-running tasks
02One task, end to end
One task. Every step visible.
Plan, tools, approvals, tests, and artifacts stay in one reviewable thread.
TaskReview this release, fix the failing check, and prepare the launch notes.
Step 1 of 6
It reads the repository before it proposes anything.
The agent opens the project, the diff under review and the instructions the project carries. What it read is listed, so you can tell whether it looked at the right thing.
repository · project instructions · release diff
Read
src/checkout/total.ts
The change under review
03Your workbench
Everything the task needs, in one workbench.
Chat, agents, workflows, knowledge and plugins are not five products with a shared login. They read and write the same working context.
Chat & context
The thread that started it, with the files and instructions it was given.
- You · Agent
- acme/checkout-service
Agent task
The task holds the plan, the steps, and the state each one reached.
- DoneRe-run the failing check and read the assertion
- DoneRound per currency in the total helper
- In progressRe-run the check against the changed files
- Not startedDraft the launch notes from the merged changes
Workflow
The repeatable version of the same work, runnable on a schedule.
- terminal → editor → terminal → artifact
Knowledge
What the workspace remembers between tasks, and where it came from.
- src/checkout/total.ts
- src/checkout/total.test.ts
- AGENTS.md
Plugins
The tools this task is allowed to call, and what each one may touch.
- git push
- pnpm test --filter checkout
Artifact
What the task produced, kept in the workspace rather than in a message.
- launch-notes.md · 2.4.0
04Desktop first
A workspace that stays close to the work.
The work is on your machine: the files, the terminal, the repository, the tools. So is the workspace.
Files and projects
Open a real project directory. The agent reads and writes the same files you do.
Integrated terminal and editor
Run the check, read the output, edit the file — without leaving the task.
System entry points
Reach the workspace from a shortcut, and let long tasks report back through notifications.
Long-running tasks
Work that takes minutes keeps running while you do something else.
Local actions that ask first
Anything touching your machine outside the project names itself and waits.
$ git push05One workspace, many entry points
Start on the desktop. Decide from wherever you are.
The task lives on your machine. The desktop runs it, and your phone, the chat platform you already use, a terminal and the browser all reach the same thread, each with only the authority you gave that device.
How one task moves between devices: started on the desktop, approved from the phone, reported in chat, resumed from the terminal, and fed a page from the browser.
Desktop
Runs the task
macOS, Windows and Linux. The files, the terminal and the agent are on the same machine.
Phone
Approves the checkpoint
The paired iOS or Android app shows the approval that is waiting, and nothing the pairing did not grant.
Chat
Receives the result
Progress and questions arrive where you already are. A reply that would change anything still asks first.
Terminal
Resumes the session
The same agent loop as the desktop, from a shell, picking up the session where it left off.
Browser
Adds a page
The companion extension captures a page you chose and hands it to the workspace over a paired, loopback-only channel.
Chat platforms with an adapter in the repository
- Telegram
- Slack
- Discord
- Lark
- DingTalk
- WeCom
- WeChat Official Accounts
- Matrix
- OneBot
Interface reconstruction. Every surface above is rebuilt in this page from the same demo task. Each device holds only the capability the host granted it.
06Run it your way
Choose the model. See the boundary.
Different work deserves different runtimes. Each option states what leaves your device, who receives it, and what needs your confirmation.
Local runtime
A model served from your own machine or a service you operate.
- What leaves the device
- Nothing, when the endpoint is on your machine
- Who receives it
- The runtime you configured
- Tools it can call
- The tools you grant the session
- Needs confirmation
- Writes outside the project, and any external call
Bring your own key
Your provider account and your credentials, held in the system keychain.
- What leaves the device
- The prompt and the context you attach to it
- Who receives it
- The provider you chose
- Tools it can call
- The tools you grant the session
- Needs confirmation
- Writes outside the project, and any external call
Existing subscription
Reuse a coding subscription you already pay for, through its own sign-in.
- What leaves the device
- The prompt and the context you attach to it
- Who receives it
- The subscription's provider
- Tools it can call
- The tools that subscription's agent supports
- Needs confirmation
- Writes outside the project, and any external call
Fallback and routing
A second model takes over when the first is rate-limited or unavailable, on rules you set.
- What leaves the device
- The same context, to whichever route is selected
- Who receives it
- The next provider in your route
- Tools it can call
- Unchanged by the switch
- Needs confirmation
- Unchanged by the switch
Local, offline, self-hosted and private are four different things. Each row above says which one it is.
07Connections with consequences
Connect tools without losing the task.
A connection is only useful if you can say what it reads, what it may do, and when it has to ask.
Repository
- Reads
- Code, issues and the diff under review
- Can act
- Propose a branch, a commit, a pull request
- Requires approval
- Anything pushed to the remote
MCP tool
- Reads
- The inputs you pass to the call
- Can act
- Run the tool the server exposes
- Requires approval
- First use, and any call marked sensitive
Plugin
- Reads
- Only the capabilities its manifest declares
- Can act
- Produce or update an artifact in the workspace
- Requires approval
- Every capability outside its declared set
Chat and notifications
- Reads
- The messages addressed to it
- Can act
- Report progress and ask for a decision
- Requires approval
- Acting on a reply that changes anything
The full provider, MCP, plugin and connector catalogs live in the documentation rather than on this page.
Agent interop
Run these agents inside the workspace over ACP, or import their existing session history. Each runs with the tools and permissions you grant it, and its steps land in the same reviewable thread.
- Claude CodeRunImport history
- CodexRunImport history
- Gemini CLIRunImport history
- OpenCodeRunImport history
- Cursor CLIRun
- Copilot CLIRun
- Qwen CodeRun
- KiroRun
- AiderImport history
- ContinueImport history
08The whole instrument
One codebase. Every part in the open.
What is inside the workspace, counted from the repository when this page was built rather than claimed in a slide.
Counted at build time
- In-tree plugins
- 66
- Chat platform adapters
- 11
- Workflow node kinds
- 236
- Rust crates
- 59
- Workspace packages
- 33
- Recorded design decisions
- 195
- Test files
- 10,803
Directory listings and file counts over the checkout. Any of them can be reproduced with a single ls.
Work
Where a task is planned and run.
- ChatThe thread that starts a task, with the files, mentions and instructions it was given.Learn more
- Agent tasksA plan you approve, steps with the tools each one needs, and the state each reached.Learn more
- SquadsSeveral agents on one task, with roles, a shared record and one place to review their work.Learn more
- WorkflowsThe repeatable version of work already done once, drawn as a graph and run on a trigger.Learn more
- SchedulerTasks and workflows on a calendar, with a ledger of every run and what it produced.Learn more
- Background runsWork that takes minutes keeps going while you do something else, and reports back.Learn more
Remember
What the workspace keeps between tasks.
- Knowledge and repo wikiDocuments, a generated wiki of the repository, and cited search over both.Learn more
- Long-term memoryFacts the workspace learned, reviewable before they are kept and traceable to their source.Learn more
- Artifacts and canvasWhat a task produced, kept as a versioned file in the workspace rather than in a message.Learn more
- OCRText lifted from screenshots, scans and PDFs, on the machine, before it reaches a model.Learn more
Reach
What it can touch, and on what terms.
- PluginsWebAssembly, TypeScript and Python plugins that declare every capability they use.Learn more
- MCP serversTools exposed by any MCP server, granted per session and confirmed on first use.Learn more
- Chat connectorsInbound messages and outbound reports over the platforms listed above.Learn more
- Browser companionAn extension that sends a page you chose into a task, and nothing it did not ask for.Learn more
- Computer useThe agent drives applications on your desktop, each action visible and interruptible.Learn more
Control
How every action stays accountable.
- Permission modelTools declare what they touch. Anything that reaches outside the project stops and names itself.Learn more
- Traces and costEvery call, token and tool result recorded against the task that made it.Learn more
- Source controlDiffs, commits and stacked branches from inside the workspace, pushed only when you say.Learn more
- Integrated terminalReal shells beside the thread, shareable to a paired device with the roster in view.Learn more
09Trust receipts
Built in the open. Controlled in use.
Four things you can check yourself, without taking a marketing page's word for any of them.
Source
The whole application is public and licensed AGPL-3.0-or-later — including the parts that talk to your files, your keys and your models.
Read the sourceData
Where your work is stored, what a model call carries, and which runtime receives it — stated per option rather than as one promise.
Data and storagePermission
Tools declare the capabilities they need. Anything that reaches outside the project stops and names the action before it runs.
Permission modelRecord
Steps, tool calls, approvals and results stay attached to the task that produced them, so a result can be traced back or undone.
How records work
From the repository
- Stars
- 54
- Contributors
- 3
- License
- AGPL-3.0
- Latest release
- None published yet
as of 2026-09-29
10Start here
Bring your agents into one open workspace.
No installer has been published yet. The build is public and reproducible — clone the repository and run the desktop build, or watch releases to be notified when the first one ships. Watch releases
Desktop first. Open source. Built for real work.
What you get today
- License
- AGPL-3.0
- Platforms
- macOS · Windows · Linux
- Latest release
- None published yet
- Changes waiting
- 1480entries since the last release