One workspace for the whole loop
Chat, edit, inspect source control, run terminals, review files, and coordinate agents without moving the project into a hosted workspace.
Projects · threads · editor · terminalsLocal-first desktop workspace
AI coding, under your control.
ADDOM brings multiple providers, guarded tools, an integrated editor, memory, and delegated agents into one desktop workspace. Your projects stay local.
WinGet
winget install --id 9N3MG7BJRSRB --exact --source msstorev0.1.2-alpha · Signed and updated by Microsoft Store · MIT licensed
ProvidersChoose the model for the work
GuardrailsReview consequential actions
ContinuityKeep project context attached
DelegationRoute work to specialist agents
A complete working surface
More than a model in a text box.
ADDOM is built around the repository, the evidence, and the decisions that move real work forward—not around isolated prompt sessions.
Bring the provider that fits
Use OpenAI, Anthropic, Google, OpenRouter, and other supported providers from a common workspace instead of committing the project to one model path.
Provider choice · per-thread modelsAutonomy with a visible boundary
Permission modes, tool approvals, execution evidence, and review surfaces keep machine-side work inspectable while letting you decide how much the system may do.
Permissions · approvals · evidenceSpecialists, not duplicate chat tabs
Create agent roles with their own provider, model, skills, and working scope. ADDOM routes bounded tasks and brings their results back into the parent workflow.
Roles · routing · parallel workThe real product
One workspace. Several useful modes.
The website uses the actual application—not speculative mockups—to show how conversation, code, providers, and agents fit together.
01 / Coordinate
Give each specialist a clear role.
Assign each role its own provider, model, skills, and working scope before ADDOM routes bounded tasks.
- Provider and model per role
- Bounded working scope
- Parallel and nested work
02 / Build
Read, edit, run, and review in context.
The integrated editor, source control, terminal sessions, and file review keep implementation evidence next to the conversation that produced it.
- Project-aware editor
- Visible terminal sessions
- Reviewable file changes
03 / Choose
Use the model that fits the task.
Configure supported providers once, select providers and models per thread, and use OpenRouter visibility without reshaping the workspace around one vendor.
- Direct provider connections
- OpenRouter model catalog
- Local and account-backed paths
Local by product shape
Your work is not our dataset.
Projects, settings, and continuity remain anchored to the device. A remote boundary appears only when you deliberately choose a configured service.
Read the privacy boundaryProjects and threads
Stored locally with the desktop workspace.
Credentials
Saved on-device and used only for the provider path you configure.
Telemetry
No analytics SDK, behavioral tracking, or developer data collection.
Remote models
Requests go to the selected provider when you choose a remote model.
Alpha, stated plainly
What works. What doesn't yet.
Kept honest on purpose. ADDOM is developed in the open by one person, and the current alpha is available as a packaged download or as source you can inspect and build.
Working today
- Signed Microsoft Store package with Store-managed updates
- Multi-provider chat with per-thread provider and model choice
- Guarded file, search, command, and terminal tools with approvals
- Integrated editor, change review, and artifact history
- Delegated agent roles with bounded working scope
- Local threads, project memory, attachments, and settings
Not there yet
- GitHub-hosted installers remain unsigned
- Terminal sessions are Windows-first; macOS and Linux need polish
- Windows is locally verified; macOS and Linux need broader testing
- Expect rough edges and fast-moving changes between releases
From the developer
I built ADDOM alone over six months because I wanted an AI coding workspace I could actually trust on my own machine — one that shows its boundaries instead of asking for faith. It is written with AI assistance and held to the standard any code should be: tests, a written threat model, and everything open to inspection. If something breaks, tell me — I read every issue.
1 developer · 4,200+ integration tests · 15 locale variants · 0 telemetry
Open source · MIT
See how ADDOM is built.
Download the alpha, inspect the source, follow development, or report issues on GitHub.