My_AI Relay — VS Code Extension

English · Français

Agentic chat with your self-hosted My_AI assistant from VS Code, over an end-to-end encrypted tunnel — like Claude Code, but running on the local LLM on your own machine.

The extension is a remote client. The LLM, history, and file processing live on the PC that hosts My_AI; the extension delegates filesystem and shell operations on the workspace side, scoped to your VS Code workspace by default. This means you can run a beefy LLM on a desktop at home and use it from any laptop, anywhere — with full read/write/exec parity with Claude Code's developer experience, and the existing mobile web UI still works exactly as before.

Highlights

• 🤖 Agentic mode — the model can read, edit, and create files, run shell commands, and search the workspace via nine built-in tools. Every tool call shows up as a Claude-Code-style foldable card rendered inline in narration order (text → tool card → text → tool card → final answer), with status, input, and captured output.
• 🛡️ Workspace sandbox — all paths are resolved against the open workspace folder. Out-of-workspace access requires an explicit modal approval per path, never auto-approvable.
• ✋ Approval flow — read-only tools run silently; write_file, edit_file, and run_command ask for confirmation, with options to allow once, allow for the same file, or allow the tool for the rest of the session.
• 🔐 End-to-end encryption — AES-256-GCM over the public tunnel. The Cloudflare/serveo/localhost.run server only ever sees ciphertext.
• 🌐 Multi-tunnel failover — cloudflared, serveo, and localhost.run run in parallel on the host; the extension picks the first reachable one and falls over automatically when one drops.
• 💾 Persistent connection — credentials stay encrypted in VS Code SecretStorage. The extension reconnects automatically when the host Relay comes back online, and disconnects when it shuts down.
• 🧩 Workspace integration
  • Auto-attach the active editor file to every message (toggle).
  • "Send selection to My_AI" (Command Palette and editor context menu).
  • "Send active file to My_AI".
  • One-click Insert at cursor / Copy on every code block in replies.

Requirements

• VS Code ≥ 1.85.
• A My_AI host PC running the Relay (any machine where you can launch My_AI). The Relay button is in the GUI's left sidebar.
• An Internet connection on both ends (the encrypted tunnel goes through Cloudflare / serveo / localhost.run).

Quick start

1. Start the Relay on the host PC

1. Launch My_AI on the PC where the model runs.
2. Click the Relay button at the top-left of the sidebar.
3. Wait until at least one tunnel is "active" (status turns green).

2. Copy the connection string

In the same Relay popup, click 🧩 Copier pour l'extension VS Code. The string copied looks like:

https://gonicolas12.github.io/My_AI/router.html#d=…

It contains the list of public tunnels, the auth token, and the AES-256-GCM key — all in the URL fragment, so nothing leaks to the tunnel server. Treat it like a password.

3. Connect from VS Code

1. Install the My_AI Relay extension.
2. Open the My_AI Relay view in the Activity Bar (the small assistant icon, by default on the left). On first activation the extension will ask if you want to move the chat to the Secondary Side Bar (right side, like GitHub Copilot Chat) — accept and VS Code will remember the placement.
3. Click Paste connection string… (or run My_AI Relay: Connect from the Command Palette) and paste the string from step 2.

The chat is now live. The same conversation is visible on the host's GUI and on any mobile that has scanned the QR code.

Tip — moving the chat later. VS Code's extension API only allows custom views to be contributed to the Activity Bar or Panel; the Secondary Side Bar is reserved for user-driven layout. If you missed the initial prompt, run My_AI Relay: Move Chat to Secondary Side Bar from the Command Palette, or right-click the My_AI Relay icon → "Move To" → "Secondary Side Bar".

Agentic mode

When the extension connects, it tells the Relay it is a VS Code client. From that point on, every prompt goes through an agentic loop on the host: the local LLM can emit tool calls, the extension executes them inside your workspace, and the result is fed back to the model until it gives a final answer.

Available tools

⚠ = requires explicit user approval (modal dialog).

Sandbox & approvals

• Workspace-scoped by default. Any path is resolved against the first workspace folder. A path that escapes the root triggers a modal asking for one-shot permission — there is no "remember" option for out-of- workspace access.
• Read-only tools (read_file, list_dir, glob, grep, get_active_editor, open_file) run silently.
• Mutating tools (write_file, edit_file, run_command) prompt for approval. The dialog offers three options besides denial:
  • Allow once — single-call exception.
  • Allow for this file — auto-approves the same tool on the same path for the rest of the session (not available for run_command).
  • Allow all this session — auto-approves the tool everywhere until disconnect/reconnect.
• Approvals reset on every reconnect.

Tool cards in the chat

Each tool call renders as a foldable card with a colored left border:

• Orange — running.
• Indigo — awaiting approval.
• Green — completed.
• Red — failed.
• Gray — denied by the user.

Click the header to expand the input (JSON args) and the output (stdout/stderr for shell commands, file diffs for edits, etc.).

Conversation memory

The agentic context is preserved for the whole WebSocket session. You can say "now edit the file you just read" and the model will remember the previous tool calls. Reconnecting starts a fresh agentic conversation.

Workspace features

• Auto-attach active file — toggle in the chat header. Every message you send will silently upload the currently active editor file before delivery. Useful when you want the model to keep up with what you're working on.
• Send selection — Command Palette → My_AI Relay: Send Selection to My_AI, or right-click in the editor. Fences the selection with the current language for clean rendering.
• Send active file — Command Palette → My_AI Relay: Send Active File to My_AI. Uploads the whole file as an attachment (PDF, DOCX, code, image…).
• Insert at cursor / Copy — every code block in an AI reply gets these two buttons on hover.

Commands

Settings

Security model

• The connection string contains the AES key. Anyone with it can decrypt every message of that Relay session. Don't share it.
• Credentials are persisted in VS Code's SecretStorage — encrypted by the OS keychain, scoped per machine + user.
• The tunnel server (Cloudflare / serveo / localhost.run) sees ciphertext only. The IP of your host PC is hidden behind the tunnel.
• Stopping the Relay on the host PC invalidates the encryption key and the tunnels. The extension detects this within ~30 s and shows "offline".
• Workspace isolation. The agentic mode runs all tools client-side inside the VS Code extension. The host's full PC access (the local MCPs used by the desktop GUI and the mobile UI) is not reachable from the VS Code chat. The model can only touch files reachable through the workspace tools, which are themselves bounded by the workspace root and guarded by the approval flow.

Troubleshooting

• "No tunnel reachable" — the host PC is offline, the Relay was stopped, or your local network blocks all three providers. Try restarting the Relay.
• "E2EE decrypt failed" — the connection string is from a previous Relay session (the AES key is regenerated on every Relay start). Copy a fresh string and reconnect.
• Health check too sensitive — raise myaiRelay.healthCheckIntervalSeconds if you have a flaky link.

Development

cd vscode_extension
npm install
npm run watch       # esbuild in watch mode
# Press F5 in VS Code to launch an extension host

Package a VSIX:

npm run package

License

MIT © gonicolas12