My Dev Team - VS Code agent
An agentic AI dev team inside VS Code's native chat panel. Invoke it with
@devteam: it answers questions, drafts step-by-step plans, and carries them
out - reading, searching, creating, and editing files in your workspace and
running shell commands - while shell commands ask you first (file changes
apply directly to your Git-backed workspace). It runs on local models via
Ollama by default - nothing leaves your machine - and can
optionally use cloud models (OpenAI, Anthropic, Groq) when you add an API key.
It is built to work alongside GitHub Copilot, not just replace it (though
replacing it is one supported scenario). Sitting in the same chat panel, it
gives you a free local fallback: send the "dummy" or exploratory questions to a
local model so you do not spend your Copilot budget on them, and when that
budget runs out and Copilot chat is disabled, @devteam is still there to
answer.
Highlights
- Complements Copilot, on your budget. Live in the same VS Code chat next
to GitHub Copilot rather than instead of it. Route the cheap, throwaway, and
exploratory questions to a free local model to preserve your Copilot
allowance, and keep getting answers when that allowance is spent and Copilot
chat goes dark. Replacing Copilot outright is still a supported path - it is
just not the only one.
- Native chat experience. No custom UI to learn: attachments,
conversation follow-ups, slash commands (
/explain, /review, /plan,
/do, /fix, /test, /compact, /clear, /model), and 👍/👎 feedback all
work the way the rest of VS Code chat does. Your project's AGENTS.md or
CLAUDE.md is respected as standing instructions on every request. You can
also start from the editor: a Fix with Dev Team Quick Fix on a diagnostic,
an Explain action on a selection, and a write/repair-tests CodeLens on
test files.
- Pick a model, or let Auto choose. Run on local Ollama models or cloud
models (OpenAI, Anthropic, Groq) - choose one with
/model, or leave it on Auto
to route by capability among the models available to you. Every reply shows
which model answered. Cloud keys are stored securely, never in settings.
- A real multi-agent pipeline on local models. A triage agent routes each
request, a planner drafts a tool-aware checklist, an executor walks it in a
tool-calling loop over five workspace tools, and an answerer handles plain
questions - each on the model best suited to its job.
- Skills the agent loads on demand. Package reusable know-how (how to write
a commit message, format a changelog entry, follow a team convention) as a
named, described skill; the agent pulls in its full instructions only when a
task matches. A few ship built-in, and you add your own by dropping a
SKILL.md into .devteam/skills/ in your workspace.
- Capability-based model routing. Agents never name models: they declare
weighted capability requirements, and a router picks the best match from a
registry of scored models. Agents, models, tools, commands, and skills are
plain
.md config files - drop a file in to register one, no code change.
- You stay in control. Demanding work pauses to let you approve the plan
first - Approve, Cancel, or Revise it with a comment - and running a shell
command always asks before it runs, streaming approved commands live into a
read-only "Dev Team" terminal. File writes and edits apply directly (your
Git-backed workspace makes them easy to review and revert), and the agent can
only touch files inside your workspace.
- Built to split. The whole pipeline sits behind a wire-shaped engine
protocol: today an in-process engine, later a remote backend speaking the
same contract - without touching the UI or the tools.
Documentation
- QUICKSTART.md - the shortest path for a new user: what
to do right after installing the
.vsix to get a working chat.
- HOWTO.md - the full end-user guide: setup, first steps,
slash commands, approvals, settings, and troubleshooting.
- DESIGN.md - the developer documentation: architecture,
the request flow, the engine protocol, the configuration system, the model
router, the tools, development setup, and the roadmap.
- CONTRIBUTING.md - how to contribute.
- CHANGELOG.md - what changed, release by release.
Tech stack
- Mastra - agents and the orchestrating workflow
- Vercel AI SDK +
ollama-ai-provider-v2,
@ai-sdk/openai, @ai-sdk/anthropic, @ai-sdk/groq - model interface to
local Ollama and cloud (OpenAI, Anthropic, Groq) models
zod - structured-output, protocol, and config validation- VS Code Chat + Language Model Tools APIs - the front end and tool
surface
License
Apache License 2.0 - see LICENSE.
| |
