AgentCode

Agentic AI coding assistant for VS Code — chat, edit, run tools, and get inline completions powered by Claude, GPT-4o, or Gemini. Open source, multi-model, and cost-aware.

Quick Start

1. Install the CLI

pip install agentcode-cli

2. Add your API key

Open VS Code Settings (Cmd+, / Ctrl+,) and search AgentCode, then paste your key:

Or add a .env file to your project root — AgentCode reads it automatically:

ANTHROPIC_API_KEY=sk-ant-...

3. Open the chat panel

Press Cmd+Shift+A (Mac) / Ctrl+Shift+A (Windows/Linux) or run AgentCode: Open from the Command Palette.

Features

Agentic Chat Panel

Ask AgentCode to read files, write code, run shell commands, search your codebase, and more — all from the side panel. It streams responses in real time, shows every tool call it makes, and asks for your approval before writing or running anything.

Inline Completions

Ghost-text suggestions appear as you type. Press Tab to accept. Toggle on/off anytime via AgentCode: Toggle inline completions in the Command Palette.

By default completions use Claude Haiku (requires agentcode.anthropicApiKey). You can swap in any OpenAI-compatible endpoint — e.g. a local Ollama running an open-weight coding model — by setting agentcode.inlineCompletions.endpoint and agentcode.inlineCompletions.model.

Example: run a small coding model locally with Ollama

ollama pull qwen2.5-coder:3b
ollama serve

Then in VS Code Settings:

• agentcode.inlineCompletions.endpoint = http://localhost:11434/v1
• agentcode.inlineCompletions.model = qwen2.5-coder:3b
• Leave agentcode.inlineCompletions.apiKey blank

Completions are now free, local, and private.

Multi-Model Support

Switch models mid-session from the dropdown in the chat panel. Works with:

Run the AgentCode 27B Locally (Free, Private)

AgentCode ships its own fine-tuned agentic model — a 27B Qwen-3.5 fine-tune that emits tool calls natively. Run it with Ollama and pay nothing.

Requirements: ~20 GB free RAM (Q4_K_M quantization). A 32 GB+ Apple Silicon Mac or any machine with a 16 GB+ GPU is comfortable.

# 1. Pull the GGUF from HuggingFace (~16 GB, one time)
ollama pull hf.co/Vigp17/agentcode-27b-gguf:Q4_K_M

# 2. Alias it to a clean name AgentCode expects
ollama cp hf.co/Vigp17/agentcode-27b-gguf:Q4_K_M agentcode-27b

Then pick Local → AgentCode 27B from the chat panel dropdown (or set agentcode.model to ollama/agentcode-27b). AgentCode parses the model's XML-style tool calls and strips its <think> reasoning automatically.

Don't have the hardware? Deploy the model on a HuggingFace Inference Endpoint (GPU, ~$0.50/hr, scale-to-zero) and point agentcode.model at it via litellm, or just use Claude/GPT/Gemini instead.

Cost-Aware Routing

Automatically routes each request to the cheapest model capable of handling it — heavy tasks go to powerful models, simple ones go to fast/cheap ones.

Active File Context

Your current file is attached to every message automatically. No copy-pasting needed.

Diff Viewer

When AgentCode wants to edit a file, VS Code's native diff viewer opens so you can review the change before it's applied.

Commands

Configuration

Keys can also be set via environment variables or a .env file in your project root — VS Code settings take priority.

Links

• GitHub — source code, issues, contributions welcome
• PyPI — CLI package

License

MIT