Parth — AI Coding Agent for VS Code
Attach files, pick any AI provider and model, and describe the change. The agent reads your code, edits it, updates the files that depend on the change, checks its work with VS Code's diagnostics, and shows every edit for review.
🚀 Getting Started
- Open the Parth view in the Activity Bar (or run Parth: Focus on Chat View).
- Click Connect a provider, then Sign in or API key on a provider. Keys go into VS Code's password prompt and secure storage (your OS keychain), never into the chat.
- Pick a model in the message box. The first model of a newly connected provider is picked for you.
- Attach files (
@, 📎, or hold Shift and drop them on the chat), type what you want changed, and press Enter.
- Review the files changed card: click a file for its diff, then Keep or Undo.
✨ Features
Attach context your way. Type @ to search the workspace, click 📎, or hold Shift and drop files or folders anywhere on the chat. You can also right-click → Add to Parth Chat, or send a selection (Cmd/Ctrl+Alt+L). Code copied from an editor and pasted into the chat attaches as a pill with its lines, like app.ts (10-20). In a notebook, a selection or paste from a cell attaches as a pill for the notebook file with the cell's number, like analysis.ipynb [cell 3]:5-8. Code blocks in the answers have Copy and Add to Parth Chat buttons.
Three modes.
- Agent edits files and checks the result.
- Plan is read-only and writes a step-by-step plan, with a Build this plan button.
- Ask is read-only and answers questions.
Any provider, your choice of model:
| Provider |
Connect with |
| Anthropic (Claude) |
API key (Sign in opens the key page for you) |
| OpenAI |
API key (Sign in opens the key page for you) |
| OpenRouter |
Sign in or API key |
| Google Gemini API |
API key (Sign in opens the key page for you) |
| Google Vertex AI |
Google Cloud credentials, or an API key (express mode) |
| AWS Bedrock |
AWS profile or SSO, or a Bedrock API key |
| Azure OpenAI / Foundry |
API key or Microsoft Entra sign-in |
| Any OpenAI-compatible endpoint |
API key + base URL |
| Ollama, LM Studio |
Local, no account |
| GitHub Copilot models |
Your VS Code account |
Pick the model and, for models that support it, the reasoning Effort in the message box. A ring next to the send button shows how much of the model's context the chat uses.
Edits that reach every affected file. The agent uses VS Code's language services: go to symbol, find references, rename, diagnostics.
Review everything. Changes go to the editor buffers, not straight to disk. The files changed card above the message box lists each file: click it for a diff, then ✓ Keep or ↺ Undo, or Keep all / Undo all. In your files, the lines the agent changed are highlighted, with Keep and Undo buttons above each changed block, so you can review one block at a time without leaving the editor.
You stay in control. Choose the Permissions in the message box: Manual (every change and command asks), Auto Edit (edits apply for review, every command asks; the default) or Auto (tests, builds, linters, type checks and read-only commands run; anything risky asks, and the card says why). An approval card can allow a command once, for this chat, or always, and deleting files always asks. ■ Stop ends a running command too, not just the request. The agent can't read .env, keys or .git, and can't touch files outside the workspace.
Your organization's AI gateway. Add a company provider (an Anthropic- or OpenAI-compatible proxy) under ⚙ → Providers, with the credential taken from a sign-in command, a file, an environment variable or an API key. One ~/.parth/providers.json can be shared with a whole team. If the gateway bills per project, a Project picker next to the model says which one this repository's usage counts against.
Context from anywhere. Attach files and folders from outside the workspace (sent read-only), and screenshots or images for models with vision.
Keep typing while it works. Messages sent while the agent is busy are queued and sent in order. Click your last message to edit it and send it again.
Pick up where you left off. The 🕘 chat history button lists your earlier chats in this workspace, with a search box. Open one to carry on: the model still remembers it. Chats are saved on your computer only, in ~/.parth/chats.
Local models and MCP. Ollama and LM Studio run without an account. Servers listed in the project's .parth/mcp.json can be added for projects you trust.
📋 Requirements
- VS Code 1.95.0 or higher, or an editor built on it (Cursor, Windsurf, VSCodium, …) with the same minimum.
- A folder or workspace open. The agent can only see and change files inside it.
- An account with an AI provider (API key or sign-in), or a local model server such as Ollama. GitHub Copilot models also need the GitHub Copilot extension, signed in.
- Desktop VS Code. The agent runs as a local process, so vscode.dev and github.dev aren't supported.
- No Node.js install needed. The agent runs on VS Code's own runtime.
- Internet access to your provider's endpoint (not needed for local models).
⚙️ Extension Settings
| Setting |
What it does |
parth.approvalMode |
Permissions: manual (asks before every change and command), autoEdit (edits apply, still listed for Keep / Undo; commands ask; the default) or auto (commands that pass the safety check run; risky ones ask). Also in the chat's Permissions pill and /permissions |
parth.allowedCommands |
Command prefixes that may run without asking in every permissions mode, e.g. npm test. Matched on whole words; in a chain every part must be allowed; deleting commands always ask |
parth.maxToolRounds |
Maximum tool round trips per request. 0 (the default): no limit |
parth.highlightChanges |
Highlight the lines the agent added or changed in your files until you Keep or Undo them. On by default |
parth.reviewInEditor |
Show Keep / Undo above each change the agent made in your files, and Keep All / Undo All at the top of the file. Needs editor.codeLens. On by default |
parth.oauthClientIds |
OAuth client IDs for sign-in that needs a registered app, e.g. {"azure": "<entra-app-id>"} |
parth.providersFile |
The company and custom providers file. Empty: $PARTH_PROVIDERS_FILE, then ~/.parth/providers.json. Only use a file you trust: an entry can run a command to get its credential |
parth.mcp.enableWorkspaceServers |
Also start the MCP servers in .parth/mcp.json. Only for projects you trust |
📖 Usage
Type / at the start of a message for commands (/plan, /new, /keep, /undo, /models, /permissions, …). The ⚙ button opens Settings: models to show, providers, permissions and allowed commands, agent limits and the log.
New chat (the speech bubble with a plus, or /new) stops the current chat: its request is cancelled, an open approval card is denied, and a running command is ended. Its changes stay in the files changed card, and the chat itself stays in the history.
Chat history (🕘, left of New chat) lists this workspace's earlier chats, newest first. Click one to open it and carry on; point at a row and click the bin to delete it. Chats are saved in ~/.parth/chats, readable only by you; the newest 100 are kept.
In the Command Palette under Parth: New Chat, Clear Attachments, Keep All Changes, Undo All Changes, Set Provider API Key, Sign Out of a Provider, Restart Agent Core, Show Logs. Add to Parth Chat is in the Explorer and editor-tab context menus, and Add Selection to Parth Chat is in the editor context menu.
🛠️ Known Issues
Drag and drop doesn't work on VS Code 1.140.0
On VS Code 1.140.0 a dropped file isn't attached, and the "Drop to add to chat" outline doesn't show while you Shift-drag. Check your version in Help → About.
- Without downgrading: attach files with
@, 📎, or right-click → Add to Parth Chat (Explorer or editor tab).
- To fix drag and drop: go back to VS Code 1.139.0, and set
"update.mode": "none" in your user settings so it doesn't update again. Full steps are in the main guide.
🔒 Privacy & Security
Parth sends your prompts and the file contents the agent reads to the AI provider you choose, using your own account. It has no servers of its own and collects no telemetry.
📄 License
MIT
| |
