MindConnect for VS Code

Runs the MindConnect agent server on your machine and puts its agents into the VS Code chat as @mindconnect — with the workspace folder, file, selection and problems you are looking at.

What it does

The default agent is vscode-assistant, which the extension creates on the first start: a copy of the seed data's coding-assistant with file_edit / file_write swapped for the reviewed VS Code tools. /agent or MindConnect: Select Agent picks another, /new starts a fresh session.

How the back channel is wired

@mindconnect ──REST/SSE──▶ MindConnect server ──LLM──▶ tool call vscode_edit_file
                                  │
                                  └─ MCP gateway ──streamable HTTP, bearer token──▶ extension (127.0.0.1)
                                                                                     └─ diff ▶ Accept ▶ WorkspaceEdit + save

• The registration is a file the extension drops into the managed server's data directory (<data>/local/system/mcp-servers/vscode.json); the gateway re-reads that directory, no restart needed. With an external server there is no back channel, and the chat falls back to coding-assistant.
• The gateway gives up on a tool call after 60 s, so a proposal waits 50 s for the decision; after that the agent hears “still awaiting review”, and a later Accept is still applied.
• The Admin UI is framed through a small reverse proxy in the extension: the server sends X-Frame-Options: SAMEORIGIN, and its session cookie would not travel in an iframe under a vscode-webview:// page. The proxy drops the header and keeps the cookies itself — one local user, one jar.
• The tool server is stateless (a fresh MCP server per request), so either side can restart without an MCP session to recover.

Security

The managed server runs with authentication off — it is meant for one user on this machine. What keeps other parties out:

Host theme (Admin UI inside VS Code)

The Admin UI view loads the UI with ?mc-host=vscode and hands it VS Code's colours by postMessage, so a theme in the UI can follow the editor's theme:

// webview → Admin UI frame, on load, on request and on every theme change
{ type: "mc-host-theme", version: 1, host: "vscode",
  kind: "dark" | "light" | "high-contrast" | "high-contrast-light",
  vars: { "--vscode-editor-background": "#1f1f1f", "--vscode-foreground": "#cccccc", … } }

// Admin UI frame → webview, when its listener is in place
{ type: "mc-host-ready" }

Previewing the theme without VS Code

dev/theme-preview/ holds a prototype of the Admin UI side — vscode.css (the theme overlay) and bridge.js (the receiver) — and a page that plays VS Code with the colours of Dark Modern, Light Modern and High Contrast, at side bar and editor width. It proxies any running server and injects both files, so nothing on the server changes:

node dev/theme-preview/serve.mjs http://127.0.0.1:<server-port>

Then open http://localhost:18780/__preview/. The files are read on every request — edit, reload.

Install without the Marketplace

Every mc-clients release carries the extension as mindconnect-vscode-<version>.vsix (releases): Extensions view → ⋯ → Install from VSIX…, or code --install-extension mindconnect-vscode-<version>.vsix, then Developer: Reload Window. Its version is the release's — release.sh sets it in package.json. A local build to try a change: raise the version (npm version patch --no-git-tag-version, not committed — VS Code caches icons by version), npx vsce package --no-dependencies, install the file.

Run it from source

npm install
npm run compile

Open this folder in VS Code and press F5 — an Extension Development Host starts with the extension loaded. Open the chat and type @mindconnect.

Needs Java 21+ (JAVA_HOME, the PATH, or mindconnect.server.javaHome). LLM configs are set in the Admin UI (MindConnect: Open Admin UI) — without one the agent answers with a “no API key” error.

Settings

The managed server gets MINDCONNECT_TOOLS_WORKINGDIRROOT set to the file system root, so a project outside the home directory can be the working directory too. Its MINDCONNECT_ENCRYPTION_SECRET_KEY (it encrypts stored LLM credentials and has no default) is generated once and kept in VS Code's secret storage.

Layout

src/
  extension.ts             activation, commands, status bar menu
  server/serverManager.ts  start / stop / readiness / status bar / log
  server/releases.ts       Maven Central download, Java 21 lookup
  api/client.ts            REST + SSE client for /api/**
  context/editorContext.ts what the agent learns about the editor
  chat/participant.ts      @mindconnect: sessions, streaming, approvals
  chat/chatView.ts         the Chat view and tab — sessions, streaming, approvals
  chat/session.ts          what both chats share: the session, the agent setting
  webview/chat.ts          the chat page (marked + DOMPurify), bundled separately
  backchannel/
    toolServer.ts          MCP server with the VS Code tools
    proposals.ts           diff view, Accept / Reject, WorkspaceEdit
    provisioning.ts        gateway registration, vscode-assistant
    backChannel.ts         sets it up whenever a managed server comes up
  server/envStore.ts       environment: settings + secret storage
  ui/serverView.ts         the Server view — form over the settings
  ui/adminUiView.ts        the Admin UI view and editor tab
  ui/adminUiProxy.ts       frames the Admin UI: header, cookies, redirects
  test/                    integration tests in a real VS Code

Tests

npm test starts a separate VS Code (downloaded into .vscode-test/ on the first run) on test-fixtures/workspace and runs the back-channel suite: an MCP client plays the gateway, the test clicks Accept / Reject. The stable VS Code only runs tests from the command line while no other instance is open — with VS Code open, use the Extension Tests launch configuration (F5) instead.

Next steps

• Several windows: the registration names one tool server, so with two VS Code windows the last one started receives the proposals. Needs the session (or its working directory) to reach the MCP call.
• Client-side tools in the runtime (clientTools, concept K7) would let the tools travel with the chat request instead of a registration file.
• MCP server definition provider: register MindConnect so Copilot's agent mode can use its agents and workflows as tools.
• Language model chat provider: the server's LLM configs in VS Code's model picker.
• Smaller download: the exec jar is ~360 MB; a slimmer local server distribution or a platform-specific VSIX with a jlink runtime.
• Tests with @vscode/test-electron; Marketplace and Open VSX publishing.