CodeSetu

Open-source AI coding assistant for VSCode and JetBrains. Multi-provider, self-hostable, made in India.

A Copilot / Cursor alternative designed for Indian developers, enterprises, public sector teams, and air-gapped deployments. Bring your own model — CodeSetu works with Sarvam, Hugging Face (any served chat model, dedicated endpoints, or self-hosted TGI), OpenAI-compatible APIs (Ollama, vLLM, OpenRouter, SGLang), and local self-hosted deployments. AI chat, repo-aware context, selected-code actions, inline FIM completions, tool-calling, and an extensible plugin and skill system across VSCode and JetBrains.

Highlights: AI chat in IDE · Repo-aware context · Selected-code actions · Inline (FIM) code completions · Agent Mode (reads, edits, and runs commands with your approval) · Plan Mode (plan-then-implement) · AI Skills with slash palette (/plan, /explain, /refactor, /test, /indic) · Voice dictation (Sarvam Saarika, browser SpeechRecognition, OpenAI-compatible Whisper, HuggingFace) · Provider setup and diagnostics · Workspace skills/checks · Air-gapped friendly · Hindi / Indic-aware · Plugin + skill SDK · 100% open-source (Apache 2.0)

Status

This repository is a pnpm + Gradle monorepo organized as:

• apps/vscode — VSCode extension with chat, repo-aware context, selected-code actions, setup diagnostics, and inline completions
• apps/jetbrains — JetBrains plugin with chat, selected-code actions, provider settings, and diagnostics
• packages/core — shared providers, provider factory, and tool-call registry (@codesetu/core)
• packages/plugin-sdk — plugin and skill type contracts for first- and third-party extensions (@codesetu/plugin-sdk)
• skills/ — AI skill manifests (SKILL.md per skill) loaded by hosts at activation
• plugins/ — first-party plugins built on @codesetu/plugin-sdk
• docs/ARCHITECTURE.md — full architecture, layout, and deployment notes

Agent Mode

Toggle Agent in the chat composer to let CodeSetu act, not just chat: it reads, edits, and runs commands in a tool-calling loop until the task is done, then you review the result. Plain chat is unchanged when Agent is off.

Tools: read_file, write_file, edit_file, bash, plus read-only helpers list_dir, glob, grep, todo_write, and IDE-native get_diagnostics (with find_symbol / find_references in VSCode).

Approval & safety: file edits and shell commands require your approval (Approve / Approve for session / Deny), and edits show a diff before they run. Read-only tools run without prompting. Hit Stop to cancel a run.

Project policy — drop a committable .codesetu/agent.json in the repo to share one approval policy across your team:

{
  "maxIterations": 16,
  "autoApproveCommands": ["^git (status|diff|log)\\b", "^npm (test|run lint)$"],
  "denyCommands": ["rm\\s+-rf", "\\bsudo\\b", "\\bcurl\\b"]
}

Patterns are regular expressions matched against the shell command. denyCommands block a command outright (deny wins), autoApproveCommands run without a prompt, and anything else asks. maxIterations caps the agent loop.

Prerequisites

• Node.js 18+
• pnpm 9+
• A provider — Sarvam API key, a Hugging Face token (hf_…), OpenRouter key, or a local OpenAI-compatible endpoint (Ollama, vLLM, SGLang)

Setup

corepack pnpm install
cp .env.example .env
corepack pnpm build
corepack pnpm test

Set SARVAM_API_KEY in .env before making real hosted Sarvam requests. For local OpenAI-compatible servers, set CODESETU_PROVIDER=openai-compatible, CODESETU_BASE_URL, and CODESETU_MODEL.

For local install and provider examples, see INSTALL.md. For Marketplace, Open VSX, and private VSIX hosting, see docs/PUBLISHING.md.

VSCode Settings

The VSCode extension reads these settings:

• codesetu.provider - sarvam, openai-compatible, or huggingface
• codesetu.baseUrl - optional base URL (e.g. https://router.huggingface.co/v1, a dedicated HF endpoint, or a local server)
• codesetu.model - optional model name (for Hugging Face, the model repo id, e.g. meta-llama/Llama-3.3-70B-Instruct)
• API key — set via the CodeSetu: Setup Provider command (stored in the OS secret store), or the SARVAM_API_KEY / HF_TOKEN / CODESETU_API_KEY environment variables
• codesetu.inlineCompletions.enabled - enable FIM inline completions
• codesetu.chat.maxTokens / codesetu.chat.temperature

Development

corepack pnpm lint
corepack pnpm build
corepack pnpm test

Open this repository in VSCode and press F5 from the extension workspace to launch an Extension Development Host after the scaffold is expanded with debug configuration.

Community

• Discord — chat with users and maintainers
• GitHub Discussions — design questions, RFCs, longer-form
• GitHub Issues — bugs and feature requests
• Security: see SECURITY.md for private vulnerability reporting

License

Apache License 2.0. See LICENSE and NOTICE.