Runs on your backend
Connect Selfcoder to LM Studio, Ollama, vLLM or another OpenAI-compatible backend such as llama.cpp. No hosted account and no telemetry.
Local AI coding agent for VS Code.
Selfcoder brings a local coding assistant into VS Code. Connect your own LM Studio, Ollama, vLLM, or another OpenAI-compatible backend, pick the model that fits your hardware, and use Chat, Plan, or Agent mode. No hosted account. No telemetry.
Works with LM Studio, Ollama, vLLM, and OpenAI-compatible servers · Qwen, DeepSeek, Llama, Gemma, and more
-
Chat with a local model right inside VS Code. -
Agent mode reads, edits, and runs commands in your workspace. -
Automatic context: selections, diagnostics, git diff, and pinned files. -
Pick any model your backend serves, straight from the sidebar. -
Use @Selfcoder in native VS Code chat with your local model.
Features
Bring local models into your workspace
Chat, Plan, Agent mode, workspace context, and native VS Code integration - powered by the model you choose.
Plan and Agent modes
Explore your repository without making changes in Plan, then continue in Agent to edit files, run commands, and review the work through inline diffs.
Model choice
Use any chat or coding model your backend supports - Qwen, DeepSeek, Gemma, GLM, Kimi and more.
Smart context
Selfcoder automatically selects and organizes relevant workspace context, helping the model focus on the code, changes, and details that matter for the task.
VS Code integration
Work directly inside VS Code through native chat or the dedicated panel, with streaming responses, Markdown rendering, code blocks, and persistent chat history.
Permission controls
Choose how much the agent can do on its own. Autonomous, Elevated, Balanced, and Restricted presets keep sensitive actions under control.
Bring your model stack
Use the self-hosted backend you already trust.
Selfcoder does not bundle a hosted model. Point it at a backend you control, pick the model that fits your hardware, and keep the coding loop within your own infrastructure.
LM Studio
GUI-friendly: download, manage, and serve models with a click.
Download LM StudioOllama
CLI-driven local runtime with a simple ollama pull workflow.
vLLM
GPU-focused inference server for self-hosted deployments and shared development machines.
View setup guideOpenAI compatible
Connect llama.cpp or another server that exposes /v1/models and streaming /v1/chat/completions.
- Qwen
- DeepSeek
- Gemma
- Nemotron
- Llama
- GLM
- Kimi
Why Selfcoder
A self-hosted Copilot alternative
Local
Runs on hardware you control
Models run through LM Studio, Ollama, vLLM, or another OpenAI-compatible backend you control - no hosted account and no telemetry.
Capable
Coding tasks
Go beyond simple answers. Turn coding requests into visible, reviewable work inside your VS Code workspace.
Custom
Bring own model
Choose the model that fits your machine, tune the system prompt, and set how much the agent may do on its own.
FAQ
Frequently asked questions
Is Selfcoder free?
Yes. Selfcoder is free, with no subscription, account, trial, ads, or telemetry. If it saves you time, sponsorship is welcome but never required.
Does my code ever leave my machine?
With a local setup, Selfcoder sends prompts, code context, and attachments only to the backend you configure on your machine. If you point Selfcoder at a remote or shared backend address, requests go there instead. Selfcoder itself does not add a cloud account or telemetry layer.
Which backend do I need?
Selfcoder needs a model server. Choose LM Studio for a friendly GUI, Ollama for a simple CLI workflow, vLLM for GPU-focused self-hosted serving, or the OpenAI-compatible option for servers such as llama.cpp.
Which models can I use?
You can use any model supported by your selected backend, as long as it fits your hardware. This includes models from families such as Qwen, DeepSeek, Llama, Codestral, Gemma, Mistral, and others.
Can I control what gets sent to the model?
Yes. Selfcoder can automatically select active files, selected code, diagnostics, git diffs, pinned files, workspace instructions, attachments, and recent conversation history based on the available context window. You can disable automatic context selection in settings when you want tighter control.
Does Selfcoder work with native VS Code Chat?
Yes. Selfcoder can register a native VS Code chat participant, @Selfcoder, and can also expose eligible local models to VS Code’s model picker. This is best suited for stronger hardware and models with large context windows, because native VS Code Chat manages context differently and does not use Selfcoder’s full context budgeting. For longer coding tasks, Selfcoder’s built-in chat gives you more predictable context control.
Can I attach files or images?
Yes. Selfcoder supports text files and image attachments when vision support is enabled and the selected model can understand images. This is useful for screenshots, UI bugs, diagrams, and mockups.
Can I point Selfcoder at another machine on my network?
Yes. Set Selfcoder.backendAddress to a reachable LM Studio, Ollama, vLLM, or OpenAI-compatible server on your LAN. If the server requires bearer authentication, set Selfcoder.backendAuthorizationToken. Treat that backend as the privacy boundary for prompts, attachments, and workspace context.
Does Selfcoder support project-specific instructions?
Yes. Selfcoder can read workspace instruction files such as local-instruction.md, copilot-instructions.md, or CLAUDE.MD and include them automatically with future requests.
Can Selfcoder run commands without asking?
That depends on your permission settings. Selfcoder includes permission controls so you can choose how much autonomy the agent has and keep destructive or sensitive actions gated.
What model should I start with?
Start with a model that fits in your VRAM for the best performance. Local models can also run from system RAM, but they are usually much slower. Smaller models are better for quick, focused questions, while larger models with bigger context windows are better for broad refactors, debugging, and multi-file work.