AnythingLLM for VS Code

A production-ready Visual Studio Code extension for AI Chat & AI Agent, tightly integrated with your own AnythingLLM instance. Run RAG (Retrieval-Augmented Generation) over your private documents, upload files straight from the editor, and chat with AnythingLLM workspaces — all without leaving VS Code.

✨ Features

💬 Chat Panel v2 (Webview UI)

A custom webview-based chat with a WhatsApp-style layout — user messages on the right, AI on the left:

• Bubble layout: user messages on the right (accent bubble), AI answers on the left (subtle bubble)
• Workspace selector dropdown
• Slash command palette: Ask / Summarize / Explain / Search / Upload / Agent
• Streaming responses with a typing cursor animation
• Collapsible citations with source documents
• Per-code-block actions: Copy, Insert at Cursor, Apply as Diff
• Follow-up suggestion buttons
• Progress indicator while a request is in-flight
• Cancel button to abort streaming
• Full Markdown rendering (headings, tables, code, blockquotes, lists, links)

💭 Collapsible <think> Blocks

When the underlying model emits a reasoning trace (e.g. <think>...</think> tags or any chain-of-thought output), the panel renders it as a collapsible section:

• Collapsed by default (configurable via anythingllm.thinkBlocksCollapsed)
• Click the 💭 Thinking process header to expand and inspect the model's reasoning
• The reasoning stream is captured live, so you can expand mid-response and watch it grow

⚙️ Comprehensive Settings Modal

Click the ⚙️ button in the panel header to open settings with 6 tabs:

• General — default workspace, chat mode (chat/query), show citations, agent mode
• API — base URL, API key (SecretStorage), test connection, timeout, retries
• Theme — auto / dark / light + accent color (blue / purple / green / orange / pink) with live preview
• Agent & MCP — planner strategy, max iterations, tool permissions, agent run history, MCP server management
• Usage — token & cost stats, budget warning, reset
• Advanced — telemetry toggle, multi-turn context, <think> collapse, status bar, danger zone (clear API key)

🤖 Tier 3 — Agentic Mode

An autonomous AI Agent for complex goals. Instead of chatting directly, the agent will:

1. Build a plan based on the configured planner strategy
2. Execute tool calls in a ReAct (Reason → Act → Observe) loop
3. Synthesize a final answer when the goal is satisfied or the iteration budget is reached

Planner strategies (configurable via anythingllm.agentPlanner):

• react — ReAct loop, the most flexible (default)
• heuristic — fast pattern-match intent detection
• llm — ask the LLM to produce a JSON plan
• native — use the model's native function-calling API

Built-in tools (13): | Tool | Description | |---|---| | vector_search | Semantic search over AnythingLLM workspace documents | | file_read | Read a file from the VS Code workspace | | file_write | Create or overwrite a file (with permission gate) | | grep_search | Regex search using ripgrep | | find_references | Find symbol references in the active project | | terminal_exec | Run a shell command (with permission gate) | | web_fetch | Fetch a URL and return text content | | run_diagnostics | Pull VS Code diagnostics (errors / warnings) | | git_status | Show working tree status | | git_diff | Show unstaged or staged diff | | open_file | Open a file in the editor | | list_directory | List entries in a directory | | mcp_call | Invoke a tool from an external MCP server |

Tool permission gate: destructive tools (file_write, terminal_exec) require explicit user confirmation. Users can Allow once, Always allow this session, or Deny. Run AnythingLLM: Reset Agent Tool Permissions to clear approvals.

Usage:

• Toggle the 🤖 button in the Chat Panel header, or
• Pick "Agent" from the slash command palette, or
• Ctrl/Cmd+Shift+G for a global toggle, or
• @anythingllm /agent <goal> in the VS Code Chat view

Example agent goals:

• "Find documents about the leave policy and summarize the key points"
• "Upload the active file to the workspace and explain its contents"
• "Based on the code in the editor, find related docs and explain how to use them"

🧠 Chat Participant (@anythingllm)

Invoke in the VS Code Chat view (Ctrl/Cmd+L):

• @anythingllm or @anythingllm /ask — default chat with RAG over the active workspace
• @anythingllm /summarize — summarize the active file or selection
• @anythingllm /explain — explain the active code/text for developers
• @anythingllm /search — vector search documents (no LLM)
• @anythingllm /upload — upload the active file to a workspace
• @anythingllm /agent <goal> — Tier 3 Agent mode

🗂️ Sidebar

• Workspaces — browse and select AnythingLLM workspaces
• Threads — view conversation threads in the active workspace

📎 Editor Integration

• Upload the active file to AnythingLLM via the editor context menu (Right-click → AnythingLLM)
• Auto-embed documents into the active workspace after upload

🌟 v0.3.0 Feature Set (18 additions)

Agentic (P1)

1. LLM-based JSON planner
2. ReAct (Reason → Act → Observe) loop executor
3. 11+ new agent tools (file_read, file_write, grep_search, find_references, terminal_exec, web_fetch, run_diagnostics, git_status, git_diff, open_file, list_directory, mcp_call)
4. Tool permission / confirmation gate
5. Native function-calling API integration

UX (P2) 6. Chat history persistence (per-workspace, via globalState) 7. Multi-turn context (sends last N messages with each request) 8. Code block enhancement — syntax highlighting + copy + insert + apply diff 9. Workspace document management UI (list / delete / re-embed) 10. Slash command autocomplete popup

Nice-to-have (P3) 11. MCP (Model Context Protocol) client — auto-start, tool invocation 12. Export chat to Markdown or JSON 13. Pin / bookmark messages (persists across Clear) 14. Status bar item (active workspace, agent mode, today's token count) 15. Welcome walkthrough 16. Image attachment (drag & drop or paste for multimodal models) 17. Token / cost tracking with daily budget warning 18. Agent run history persisted to disk (audit trail)

🔒 Production-Ready

• ✅ Streaming responses via SSE (Server-Sent Events)
• ✅ Secure API key storage via VS Code SecretStorage (never written to settings.json)
• ✅ Auto-retry with exponential backoff for 5xx & rate-limit responses
• ✅ Timeout & cancellation support
• ✅ Granular error handling (auth, not-found, rate-limit, network)
• ✅ Local telemetry (request count, latency, error rate, token usage) — no third-party calls
• ✅ Editor context awareness (active file + selection)
• ✅ Citations for source traceability
• ✅ Follow-up suggestions for smoother UX
• ✅ Theme customization (auto / dark / light + 5 accent colors)

💜 Support / Donate (non-intrusive)

This extension is free and open source. If it saves you time, consider supporting it. Donations are voluntary — every feature works without paying a cent.

Donate buttons are placed in 3 non-intrusive locations:

1. Chat Panel header — small 💜 heart icon next to the ⚙️ Settings button
2. Settings modal → About tab — Saweria & PayPal buttons
3. Sidebar Workspaces view — a 💜 "Support this project" row at the bottom, plus a heart icon in the view title bar

All three trigger the same AnythingLLM: Support / Donate command, which opens a QuickPick to choose between Saweria (IDR) and PayPal (USD).

📦 Prerequisites

• VS Code 1.90.0 or newer
• A reachable AnythingLLM instance (self-hosted or cloud)
• An AnythingLLM API key (generate at AnythingLLM → Settings → API Keys)

🚀 Quick Start

1. Build & Run in development mode

# Clone or navigate to the extension folder
cd anythingllm-vscode

# Install dependencies
npm install

# Compile (TypeScript check + lint + esbuild bundle)
npm run compile

To test in VS Code:

1. Open the extension folder in VS Code
2. Press F5 (pick the "Run Extension" launch config)
3. VS Code opens a new Extension Development Host window with the extension active

2. Configure

Set Base URL (if not the default)

Command Palette (Ctrl/Cmd+Shift+P) → "AnythingLLM: Set Base URL"

Default: http://localhost:3001/api — change this if your instance lives elsewhere.

Set API Key

Command Palette → "AnythingLLM: Set API Key"

The API key is stored securely in VS Code SecretStorage (OS keychain), never in plain config.

Pick a Workspace

Open the AnythingLLM sidebar in the Activity Bar, click any workspace → "Set as Active". Or via Command Palette:

"AnythingLLM: Select Workspace"

3. Start Chatting

Open the VS Code Chat view (Ctrl/Cmd+L) and type:

@anythingllm what's in the documents I just uploaded?
@anythingllm /summarize
@anythingllm /explain the main responsibility of this class
@anythingllm /search leave policy
@anythingllm /upload

Or open the dedicated Chat Panel with Ctrl/Cmd+Shift+A for the full UI experience (workspace selector, slash command bar, agent toggle, settings, export, history, document manager).

⚙️ Settings

Access via File → Preferences → Settings → search "AnythingLLM":

🏗️ Architecture

anythingllm-vscode/
├── src/
│   ├── extension.ts            # Entry point & activation
│   ├── chatParticipant.ts      # @anythingllm handler + slash commands
│   ├── chatPanel.ts            # Webview Chat Panel provider
│   ├── anythingllmClient.ts    # HTTP client with retry, timeout, SSE parser
│   ├── agent.ts                # Tier 3 Agent orchestrator (ReAct / planner)
│   ├── agentTools.ts           # 13 built-in agent tools
│   ├── agentPermissions.ts     # Tool permission / confirmation gate
│   ├── agentRunHistory.ts      # Agent run audit log on disk
│   ├── mcpClient.ts            # MCP (Model Context Protocol) client
│   ├── historyStore.ts         # Per-workspace chat history (globalState)
│   ├── tokenTracker.ts         # Token & cost estimation
│   ├── statusBar.ts            # Status bar item
│   ├── config.ts               # Settings.json wrapper
│   ├── secretStore.ts          # API key storage via SecretStorage
│   ├── state.ts                # State manager (active workspace, cache)
│   ├── treeViews.ts            # Sidebar Workspaces & Threads
│   ├── commands.ts             # All command registrations
│   ├── logger.ts               # OutputChannel logger
│   ├── telemetry.ts            # Local telemetry
│   └── types.ts                # Type definitions
├── media/                      # Icons, SVG assets, chat-ui.js, chat-ui.css
├── .vscode/                    # launch.json & tasks.json for debugging
├── package.json                # Extension manifest (participants, views, commands, config, walkthroughs)
├── tsconfig.json
├── esbuild.js                  # Bundler config
└── README.md

🔌 AnythingLLM API Endpoints Used

🛠️ Development

Watch mode (auto-rebuild on save)

npm run watch

Type check only

npm run check-types

Lint

npm run lint

Production build (minified)

npm run package

Package as .vsix (for distribution)

npm install -g @vscode/vsce
vsce package
# Output: anythingllm-vscode-0.3.0.vsix

Install the .vsix:

code --install-extension anythingllm-vscode-0.3.0.vsix

🐛 Debugging

View logs

Command Palette → "AnythingLLM: Show Output Log"

Or open the Output panel (Ctrl/Cmd+Shift+U) → select the "AnythingLLM" channel.

View telemetry

After "Show Output Log", the tail of the log shows a summary:

Stats: X requests, Y errors, avg Zms, N tokens

Verify API key & connectivity

Command Palette → "AnythingLLM: Refresh Workspaces"

If it fails, check:

1. The API key is valid (AnythingLLM → Settings → API Keys)
2. The Base URL is correct and reachable
3. Network / firewall isn't blocking the request

📝 Usage Tips

• Slash commands are more accurate than free-form prompts for specific tasks
• For large documents, use /upload first and then ask via /ask — don't paste huge content into the chat box
• /search is great for quick exploration without spending LLM tokens
• Set "anythingllm.chatMode": "query" when you only need retrieval (saves tokens)
• Toggle 🤖 Agent mode for multi-step research or code-modification tasks
• Pin important answers with 📌 so they survive "Clear Chat"
• Export the conversation with 💾 before clearing if you want to keep a transcript

🗺️ Roadmap

Ideas for future releases:

• [ ] Long-term cross-session memory (persistent user preferences)
• [ ] Auto-summarization when context window grows too large
• [ ] Self-correction loop driven by VS Code diagnostics
• [ ] Inline diff preview before applying agent-proposed edits
• [ ] Multi-model routing (pick a model per task type)
• [ ] Workspace-aware chat threads in the panel UI

📄 License

MIT

🔗 Links

• AnythingLLM Assistant
• AnythingLLM
• VS Code Chat Participant API
• AnythingLLM API Docs
• Model Context Protocol

☕ Support Development

This project is open source and free. If you find it useful, buy me a coffee to keep me motivated to keep coding! 😄

 

⭐ Don't forget to star the repository if you like this project!