Best IDE Agent

An agentic coding assistant for VS Code powered by local LLMs via LM Studio (or any OpenAI-compatible endpoint). No cloud, no Copilot dependency.

This is the MVP extension for a future VS Code fork — the agent core under src/core/ is editor-agnostic and designed to be embedded into the fork later. See ROADMAP.md for planned features.

Features

• Chat sidebar with streaming responses from your local model
• Agentic tool loop: file/dir search, regex + semantic code search, targeted search_replace, diagnostics/symbol lookup, git status/diff, file writes, and terminal commands
• Optional MCP tool bridge: connect stdio MCP servers and call external tools from agent runs
• Chat defaults to the right-hand auxiliary sidebar
• Editor-native file review: agent edits open as pending diffs with Accept/Reject/Accept All, plus one-click revert of the last accepted turn
• Approval flow: run_command requires explicit pre-execution approval
• Ask/Agent/Composer chat modes: use Ask for read-only analysis, Agent for standard tool use, or Composer for structured plan-and-apply multi-file tasks
• Project guidance via .bestide/rules.md, plus task-scoped skills loaded with @skill:name from .bestide/skills/
• Graceful step-limit resume: continue or stop when bestIde.maxSteps is reached
• Optional Tab/ghost-text code completion via VS Code inline completions
• Model picker backed by GET /v1/models
• Works with any OpenAI-compatible server (LM Studio, Ollama, llama.cpp server, vLLM, ...)

Getting started

1. Set up LM Studio

1. Install LM Studio
2. Download a model with tool-use support. Recommended starting points:

• google/gemma-4-e4b - runs well on macbooks with 32 gb+ ram
• qwen2.5-coder-7b-instruct (good balance of speed and tool-calling quality)
• qwen3-8b or larger Qwen3 variants
• llama-3.1-8b-instruct

3. Open the Developer tab and start the local server (defaults to http://localhost:1234)
4. Load the model into the server

2. Install the extension from source

The extension isn't published to the marketplace yet. To build and install it locally:

1. Clone and build a VSIX package:

   git clone <repo-url>
   cd best-ide
   npm install
   npm run vsix
   

   This produces best-ide-agent-<version>.vsix in the project root.

2. Install the VSIX (either way works):

   • CLI: code --install-extension best-ide-agent-0.1.0.vsix (if code isn't on your PATH, run "Shell Command: Install 'code' command in PATH" from the VS Code command palette first)
   • UI: open the Extensions view, click the ... menu in its top-right corner, choose Install from VSIX..., and select the file

3. Configure it: reload VS Code, open Settings, search for bestIde, and run Best IDE: Set API Key from the command palette if your LM Studio server has authentication enabled (LM Studio > Developer > API tokens). The key is stored in VS Code Secret Storage.

To update after pulling new changes, re-run npm run vsix and install the new VSIX over the old one.

3. Run from source instead (Development Host)

For hacking on the extension itself, skip the VSIX: run ./setup_and_run.sh (or npm install && npm run build), open this folder in VS Code, and press F5 to launch the Extension Development Host.

Settings

Use Best IDE: Set API Key and Best IDE: Clear API Key from the command palette to manage the default API token in VS Code Secret Storage.

Privacy and offline story

• By default, Best IDE is local-first: point bestIde.baseUrl (or your local backend profile) at a local OpenAI-compatible server such as LM Studio.
• Extension data stays on your machine: conversation threads are saved in local VS Code extension state, and agent file edits happen in your workspace.
• Nothing is sent to Best IDE-managed cloud services because this project does not run one. Network traffic goes only to the model/MCP endpoints you configure.
• If you add non-local backends (for fallback, cost, or quality routing), requests for those operations are sent to those configured providers by design.

Rules and skills

• Add project-wide guardrails in .bestide/rules.md (automatically merged into the system prompt).
• Add reusable task instructions under .bestide/skills/.
• Reference a skill in chat with @skill:name (for example @skill:test-driven loads .bestide/skills/test-driven.md if it exists).
• Skills can also be folders containing SKILL.md (for example .bestide/skills/refactor/SKILL.md).

Development

npm run test        # unit tests (Vitest)
npm run coverage    # coverage report (90% threshold on src/core)
npm run typecheck   # tsc --noEmit
npm run watch       # rebuild on change

To run a live closed-loop test of the agent core against a running LM Studio server (writes to a throwaway temp dir, never your workspace):

LM_API_TOKEN=your-token npx esbuild scripts/live-e2e.ts --bundle --platform=node \
  --outfile=dist/live-e2e.cjs --log-level=error && node dist/live-e2e.cjs

The codebase follows red/green TDD on src/core — the OpenAI client, SSE parser, tool registry, and agent loop are all pure TypeScript with no vscode imports, tested against mocks.

Architecture

webview/         React chat UI (runs in webview sandbox)
src/extension/   VS Code glue: webview provider, message bridge, WorkspaceHost impl
src/core/        Editor-agnostic agent: OpenAI client, agent loop, tools

The webview talks to the extension host over postMessage. The extension host runs the agent loop, which calls LM Studio over HTTP and dispatches tool calls through a WorkspaceHost interface.

Known limitations (MVP)

• Tool-calling quality varies a lot between local models; small models may emit malformed calls
• Single conversation at a time; history is in-memory only
• run_command depends on VS Code terminal shell integration; if shell integration is unavailable, command execution is blocked