---

Free Repo Agent brings a Claude/Cursor/Cline-style agent into VS Code. It indexes your whole repository, writes and edits code with inline Keep / Undo, runs terminal commands with your approval, verifies its own work (typecheck/build → then fixes the errors), lets you switch models (DeepSeek V4 Pro / V4 Flash with thinking mode, or any OpenAI-compatible model), tunes effort (fast → max reasoning), reads images (paste a UI → build the component), extends itself with MCP plugins (local & remote), learns your codebase, remembers your project, and runs a staff-engineer pre-commit code review — all on the affordable DeepSeek API with no monthly subscription.

📚 Contents

⚡ Everything it does 🚀 Get started in 60 seconds 🧠 Models & thinking mode 📖 Learn this codebase 👁️ Vision — image → code

🧩 Plugins (MCP) 🧠 Project memory 🔍 Code review + custom rules 🧭 Repo indexing

⚙️ Commands 🔧 Settings 🆚 vs Copilot/Cursor/Claude ❓ FAQ · 🐞 Report a bug

⚡ Everything it does

💬 Repo-aware chat Ask anything about your codebase. It indexes your project and injects the most relevant files into every answer — no copy-pasting context.	🤖 Autonomous agent Give it a task — it reads files, edits across many files, and runs commands step by step. Every change is reviewed inline with ✓ Keep / ✗ Undo.
✅ Auto-verify & self-fix After editing it runs your typecheck/build, reads the errors, and fixes them before saying "done" — no more code that doesn't compile.	👁️ Vision (image → code) Paste a UI screenshot/mockup and say "build this component." Free by default via Google Gemini (or local Ollama). Falls back to OCR for screenshots of code/errors.
🧩 Plugins (MCP) One-click tools the agent can use — GitHub, web search, Postgres, Playwright, filesystem & more — via the Model Context Protocol. Add any custom server.	🔍 Staff-engineer code review Reviews staged changes / files / branches, reads the full files, runs a second audit pass, and reports Critical · Quality · Architecture · Testing. Enforces your own rules.
🧠 Model switcher + thinking mode Switch models from the footer, Claude-style: DeepSeek V4 Pro / V4 Flash with thinking mode (deep reasoning that works with tool calls), or point at OpenAI / OpenRouter / Gemini for GPT/Claude/Llama.	⚡ Effort control A dial from fast & cheap → max reasoning. Higher effort = more tool steps, longer answers, more context, and deeper V4 thinking — you decide speed vs. quality per task.
🧩 Plugins (MCP) — local & remote One-click tools — GitHub, web search, Postgres, Playwright, filesystem & 20+ more — via the Model Context Protocol. Install from npm/GitHub, a custom command, or a remote URL.	🔍 Staff-engineer code review Reviews staged changes / files / branches, reads the full files, runs a second audit pass, and reports Critical · Quality · Architecture · Testing. Enforces your own rules.
📖 Learn this codebase + 🧠 memory One click generates a knowledge brief of your project (architecture, conventions, key files) that's injected into every prompt — plus a memory.md for your own always-on rules.	👍 Learns from your feedback Every reply has Copy · 👍 · 👎. A 👎 saves what was wrong as a rule the agent follows from then on — and offers a one-click Retry. Works with no folder open too.

🚀 Get started in 60 seconds

1. Get a DeepSeek API key → platform.deepseek.com/api_keys — pay-as-you-go, a few dollars lasts a long time.

2. Add it to VS Code → Cmd/Ctrl + Shift + P → Agent: Set DeepSeek API Key (stored in VS Code SecretStorage, never in your repo).

3. Open chat → press Cmd/Ctrl + Shift + A, or click the Repo Agent </> icon to dock it beside your code — and start.

💡 Type / in the chat for quick commands: Plugins, Memory, Reindex, Auto-accept, Vision and more.

🤖 Autonomous agent — self-verifying

Describe a task and watch it work end-to-end:

"Add a /health route and a test for it, then run the tests."

1. 📖 Reads the relevant files (from the repo index).
2. ✍️ Proposes edits inline in your files with green highlights and ✓ Keep / ✗ Undo.
3. ▶️ Asks to Run any terminal command in a visible Repo Agent terminal.
4. ✅ Auto-verifies — runs your typecheck/build and fixes errors before finishing.
5. ☑️ A live checklist ticks each step pending → in-progress → ✓ done. Hit Stop anytime.

🧠 Models & thinking mode — generate the best code

Switch the model from the 🧠 chip in the chat footer (or the ⚙︎ Options menu), just like Claude:

Model	Best for
DeepSeek V4 Pro (deepseek-v4-pro)	Hardest code & reasoning — top quality
DeepSeek V4 Flash (deepseek-v4-flash)	Fast & affordable, 1M context — the default
R1 / V3 (deepseek-reasoner / deepseek-chat)	Legacy models

• Thinking mode — on the V4 models the agent uses DeepSeek's built-in deep reasoning that works with tool calls, so it thinks before it edits. It's wired to the ⚡ Effort dial: off at Low (fast/cheap), then low → high → max reasoning at Medium → High → Max.
• Not just DeepSeek — the extension speaks an OpenAI-compatible API, so set a custom model id + point repoAgent.deepseek.baseUrl at OpenAI, OpenRouter, Gemini, or any provider to run the agent on GPT-4o, Claude, Llama, Qwen and more.

📖 Learn this codebase

Run Agent: Learn This Codebase (or / → Learn this codebase) and it indexes your repo and writes a thorough knowledge brief — purpose, stack, architecture, key files, data models, conventions, build/run/test — to .repo-agent/knowledge.md. That brief is injected into every prompt, so the agent knows your project instantly. Re-run anytime to refresh.

👁️ Vision — paste an image, build the component

DeepSeek is text-only, so image understanding uses a separate, free-by-default vision model.

/  →  👁️ Vision  →  choose a provider:
     • Google Gemini — FREE (no credit card)   ← default
     • Ollama / LM Studio — local, no key
     • OpenRouter (free models) · OpenAI · Custom

Then paste a screenshot/mockup into the chat and ask "build this as a React component." The vision model turns the image into a precise implementation spec, and the agent builds the files. No key? Images fall back to OCR — perfect for screenshots of errors or code.

🧩 Plugins (MCP) — give the agent superpowers

Open the 🧩 Plugins panel (or type / → Plugins) and add tools with one click. Built on the Model Context Protocol — the same plugin standard Claude uses — so the entire MCP ecosystem works here. Add from the built-in catalog, npm/GitHub, a custom command, or a remote URL.

📖 Full guide: how to use plugins →

	Plugin	What the agent can do
📁	Filesystem	Read/write/search files in a folder you allow
🐙	GitHub	Search repos/issues, read files, open PRs & issues
🦊	GitLab	Repos, issues, merge requests
🔎	Web Search (Brave)	Search the web for up-to-date info
🦅	Tavily	AI-grade web search & page extraction
🌐	Fetch	Fetch a URL as clean text/markdown
🎭	Playwright	Drive a real browser: click, fill, scrape, screenshot
🐘	Postgres · 🗃️ SQLite	Query your databases
🪜	Sequential Thinking	Structured step-by-step reasoning
📚	Context7	Up-to-date docs & snippets for any library
🧠	Memory	A knowledge graph the agent reads/writes across tasks
🗺️	Google Maps · 💬 Slack	Places/directions · channels & messages

➕ Add any MCP server — from npm/GitHub (paste @scope/server or owner/repo), a custom command, or a remote URL (hosted servers over Streamable HTTP, with an optional bearer token). Keys are stored securely in VS Code SecretStorage. Full guide → docs/PLUGINS.md

🧠 Project memory — memory.md

Type / → Memory to create a memory.md at your repo root. Its contents are injected into every chat and agent prompt, so the agent always remembers your project:

# Project Memory

## Project
- Stack: React + TypeScript + Vite, Node/Express API, Postgres
- Run: `npm run dev` · Tests: `npm test`

## Conventions
- 2-space indent, single quotes, named exports.
- Always add types and a test. Never commit secrets or use `any`.

Delete the file to turn memory off. (You can also drop a .agent.md / .claude.md — it reads those too.)

🔍 Staff-engineer code review + your own rules

Stage changes (git add …) → Agent: Review Staged Changes (also in the Source Control title bar). Findings appear in the Problems panel, the Review Results sidebar, and inline — each with Fix with AI, Ignore, and Commit Anyway.

Make it enforce YOUR standards. Run Agent: Create Review Guidelines to scaffold a system-design.md. The reviewer reads it before every review and flags any change that breaks a rule — quoting the rule:

# Review Guidelines & System Design

## Architecture
- UI → services → data access. UI must not import data-access directly.

## Review rules
- No `console.log` in committed code — use the logger.
- Every new exported function ships with a test.
- No secrets in source; read them from config/secret storage.

🧭 Repo-aware indexing & retrieval

On first run, Repo Agent indexes your repository — symbols, structure, and optional AI per-file summaries — into a local cache (.agent-cache/). Every question then pulls in the most relevant files automatically, so answers fit your codebase. Re-run anytime with Agent: Reindex Repository, and tune what's indexed via repoAgent.indexing.* and how many files are injected via repoAgent.retrieval.maxFiles.

⚙️ Commands

Cmd/Ctrl + Shift + P → type "Agent:"

Command	Description
Agent: Chat / Open Chat (beside editor)	Open the chat panel
Agent: Generate Code	Describe code to generate
Agent: Run Agent (autonomous)	Run a task that edits files & runs commands
Agent: Review Staged Changes / Review Current File	Pre-commit / file review
Agent: Create Review Guidelines	Scaffold a system-design.md of review rules
Agent: Learn This Codebase	Generate a persistent project knowledge brief
Agent: Edit Memory	Edit the always-on memory.md
Agent: Choose Vision Provider	Pick a vision backend (Gemini free / Ollama / …)
Agent: Reindex Repository / Explain Architecture	Index & architecture tools
Agent: Set / Clear DeepSeek API Key	Manage your key

🔧 Settings

All settings live under repoAgent.* — click to expand

Setting	Default	Description
deepseek.model	deepseek-v4-flash	Active model (switch it from the 🧠 footer chip)
effort	medium	Speed vs. reasoning depth: low · medium · high · max
deepseek.baseUrl	https://api.deepseek.com	API base URL — point at any OpenAI-compatible provider
deepseek.maxTokens	4096	Max tokens per response (auto-raised for V4)
vision.baseUrl / vision.model	Gemini free	OpenAI-compatible vision endpoint
review.guidelinesFile	system-design.md	Your custom review-rules file
review.deepAudit	true	Second audit pass for fewer false positives
indexing.maxFiles	1500	Index size cap
indexing.generateSummaries	true	AI per-file summaries (set false for $0 indexing)
retrieval.maxFiles	8	Relevant files injected per prompt
plugins.autoRun	true	Run plugin tools without a confirm prompt
review.autoOnSave	false	Auto-review files on save

🆚 vs Copilot, Cursor & Claude

	Free Repo Agent	Copilot	Cursor	Claude / Cline
Price	🟢 Free (BYO key)	💲 Subscription	💲 Subscription	💲 Sub / BYO
Repo-aware multi-file agent	✅	✅	✅	✅
Runs terminal commands	✅	✅	✅	✅
Self-verifies (build/typecheck)	✅	—	◐	—
Vision: screenshot → code	✅ free	—	✅	✅
MCP plugins	✅	—	✅	✅
Code review + custom rules	✅	◐	—	—
Project memory file	✅	◐	◐	✅

A free alternative to GitHub Copilot, Cursor, Claude, Codex, Cline, Continue, Codeium, Windsurf, Tabnine, Cody, Aider & Roo Code — a lightweight, bring-your-own-key agent that edits code, self-verifies, and reviews diffs inside VS Code.

❓ FAQ

Is it really free?
Yes — the extension is free and open. You only pay DeepSeek's low pay-as-you-go API usage (your own key). Is it a good Copilot / Cursor / Claude / Cline alternative?
Yes — repo-aware chat, an autonomous agent that edits files and runs commands, visible reasoning, vision, MCP plugins, and pre-commit code review — the same core workflow, no subscription. Does it fix its own mistakes?
Yes — after editing it runs your typecheck/build, reads the errors, and fixes them before finishing. Can I add plugins / MCP servers?
Yes — open the 🧩 Plugins panel and add GitHub, web search, Postgres, Playwright and more with one click, or any custom MCP server (needs Node/npx). Does my code leave my machine?
Only the context needed for a request is sent to the DeepSeek API when you ask. Your key never leaves SecretStorage; the index is cached locally under .agent-cache/.

🐞 Report a bug / request a feature

Found a bug or have an idea? Open an issue → — you'll get a guided bug report or feature request form. Please include your extension version, VS Code version, steps to reproduce, and any error from View → Output → "Repo Agent". Browse existing reports on the Issues page.

⭐ Star history