Mira Coding Agent for VS Code

Chat with Mira — an open-source coding agent that runs on your machine — right from the VS Code Chat sidebar.

Mira reads your code, edits files, runs commands, reviews diffs, and delegates subtasks to specialised subagents. It talks to any OpenAI-compatible provider (OpenRouter, OpenAI, Anthropic, Groq, DeepSeek, Ollama, LM Studio). Your key, your model, your machine.

Preview. Building in the open. @mira chat is stable; approvals in manual mode and inline diffs are on the near roadmap.

What you get

@mira chat participant in the built-in VS Code Chat sidebar — same surface as Copilot Chat, Continue, Cursor's assistant.
Streaming responses — tokens land as they're generated, tool calls render inline as fenced code blocks with a one-line arg summary so you can see what Mira actually did (read this file, ran that command, edited that region).
Any OpenAI-compatible provider — Anthropic, OpenAI, OpenRouter, Groq, DeepSeek, Ollama, LM Studio. Configured once in mira.yaml, swappable per-turn via /model.
Permission modes — plan / manual / auto / edit / yolo decide how freely the agent edits and runs commands. Switchable via /mode.
File and selection references — #file:foo.ts, #selection, or drag-and-drop; the extension flattens VS Code's native chat references into the prompt.
Follow-active-folder — Mira's working directory tracks the VS Code workspace folder, so a fresh chat lands in the repo you're looking at.

Requirements

Mira runs as a local backend the extension talks to over WebSocket. Install and start the backend once, then this extension takes care of the rest.

Install Mira (macOS / Linux):

brew install runmira/tap/mira
# or, no Homebrew:
curl -fsSL https://raw.githubusercontent.com/runmira/mira/main/install.sh | bash

Configure a provider (either via env vars or by editing ~/.mira/mira.yaml):

export MIRA_API_KEY=sk-or-v1-...                 # e.g. an OpenRouter key
export MIRA_BASE_URL=https://openrouter.ai/api/v1
export MIRA_MODEL=google/gemini-2.5-flash

Start it:

mira serve --port 8787

Open VS Code's Chat sidebar (⌃⌘I on macOS, Ctrl+Alt+I on Windows/Linux), type @mira, and go.

Slash commands

Command	What it does
`/reset`	Start a fresh session (drops in-memory history).
`/mode plan\\|manual\\|auto\\|edit\\|yolo`	Set the permission mode.
`/model <name>`	Hot-swap the model for the next turn.

Configuration

Setting	Default	What it does
`mira.baseUrl`	`http://127.0.0.1:8787`	Where `mira serve` is listening. Loopback by default.
`mira.showToolCalls`	`true`	Render each tool call the model makes (name + args + result). Turn off for prose-only.
`mira.followActiveFolder`	`true`	Sync the Mira session's cwd to the open workspace folder on each chat send.

Change any of these in VS Code's Settings under Mira.

What this extension doesn't do (yet)

Doesn't auto-spawn mira serve. Run it yourself; a follow-up release will start it on demand when the binary is on PATH.
Doesn't show approval modals for manual mode. If the model requests permission, open the web UI to approve, or set /mode auto to skip. Inline showQuickPick approvals are on the near roadmap.
Doesn't render the subagent / review / pull-request / plugins panels. Those live in the browser UI — run the Mira: Open Web UI in Browser command from the palette to reach them.

Mira on GitHub — the backend, CLI, and web UI live here. Apache-2.0.
Report an issue — file extension bugs against the main repo, tagged vscode.

License

Apache-2.0. See LICENSE.

Mira Coding Agent

runmira