Lantern IDE

An AI code editor that shows you what the agent sees, what it changes, and how far the change reaches.

한국어 · Windows beta · Apache-2.0 · Bring your own model (Claude, GPT, Gemini, DeepSeek, Grok, Mistral, Groq, OpenRouter, Ollama, LM Studio, any OpenAI-compatible API)

Beta. Lantern is young software built by one developer. Windows is the tested platform; macOS and Linux build in CI but have not been used day to day. Expect rough edges and please report them.

Why another AI editor?

Most AI editors are organised around files and a chat box. Once you let an agent edit code, the questions change: what did it look at, what did it touch, and what else will break? Lantern is organised around those questions.

• Code map. The centre of the window switches between the editor and a live map of your code (folders → files → symbols, calls, and files that usually change together).
• Impact radius on every edit. Before you approve an agent's edit, the card shows the symbols it touches, direct and indirect callers, the tests that cover them (or that there are none), and a risk level. One click shows it on the map.
• Agent footprint. Each task records what was sent as context, what the agent read, and what it edited, and paints it on the map.
• Project memory map. Team conventions live in .lantern/memory/*.md; Lantern links each rule to the code it talks about.
• Tasks, not tabs. A task keeps its conversation, context, changed files, and undo together. Tasks survive restarts, and can run in an isolated git worktree so nothing touches your working tree until you apply it.
• Nothing hidden. No hidden system prompt; the inspector shows every file, token, and cent sent to the model. No telemetry.

Context engine

The core is a local context engine (crates/lantern-context, Rust + tree-sitter + SQLite FTS5). For every question it assembles relevant code from the symbol graph, full-text search, git co-change history, and project memory within a token budget.

Languages the engine understands (symbols, calls, map, impact radius): Rust, Python, TypeScript/JavaScript, Java, Go, C#. Other files still open with syntax colors, but have no symbols.

Public benchmark. 60 real commits from Flask (Python), Hono (TypeScript), and Apache Commons Lang (Java). The question is the commit message, the answer files are the code files that commit changed, and each question is measured at the commit's parent. Nobody hand-picks the answers. Same 8,000-token budget, no model calls:

	Keyword search (grep, read whole files)	Keyword search (grep, snippets around matches)	Lantern
Answer files retrieved (mean recall)	13%	41%	88%
An answer file among the first 3 files	15%	30%	65%
Same questions in Korean: recall / first 3	13% / 13%	38% / 27%	76% / 43%

Picking as many files at random as Lantern returns would hit an answer file 10% of the time. Korean questions are human translations of the same commit messages. Reproduce with node eval/bench/run.mjs (results, method). Bring your own project and questions with eval/retrieval.mjs.

The engine also runs on its own as a CLI and an MCP server, so you can use it from other agents:

cargo build --release -p lantern-context
./target/release/lantern -C <project> context "where is the session cookie issued?"
./target/release/lantern -C <project> graph impact src/auth/session.ts --lines 40-60   # blast radius as JSON
claude mcp add lantern -- <path-to-lantern> mcp <project>

Install

Download from Releases. Builds are not code-signed yet:

• Windows (*-setup.exe): when SmartScreen warns you, choose More info → Run anyway.
• macOS (*.dmg, Apple Silicon and Intel, untested): drag Lantern to Applications, then run xattr -cr /Applications/Lantern.app or allow it in System Settings → Privacy & Security → Open Anyway.
• Linux (*.AppImage, *.deb, untested).

On first run, Connect an AI Model walks you through it (later: Settings → Models): pick a cloud provider (Anthropic, OpenAI, Google Gemini, DeepSeek, xAI, Mistral, Groq, OpenRouter), paste its key, and choose a model from its model list; or use a local model through Ollama / LM Studio (free, and your code never leaves the machine). Any other OpenAI-compatible API works from Settings → Models. The model button in the chat box switches between the models you've connected. Keys go to the OS credential store or environment variables, never to the config file.

Use your ChatGPT subscription or Google account (external agents)

Lantern can drive official agent CLIs over the Agent Client Protocol. The CLI signs in by itself, so a ChatGPT subscription or a Google account works without an API key, and Lantern never sees your credentials.

Agent	Install	Sign-in
Codex	npm install -g @zed-industries/codex-acp	ChatGPT (paid plan), or an OpenAI API key
Gemini CLI	npm install -g @google/gemini-cli	Google account, or a Gemini API key

Pick Codex (external) or Gemini CLI (external) in the chat's agent selector. The first time, a card offers the agent's own sign-in (usually in your browser). Everything else stays Lantern: each question carries Lantern's context (and the agent gets Lantern as an MCP server), edits come to the approval card with the impact radius, and changes can be undone. External agents run only in trusted folders. Any other ACP agent can be added in config.toml with [acp.<name>] command = "…".

Claude subscriptions (Pro/Max) can't be used this way: Anthropic's terms limit subscription sign-in to its own apps. Use an Anthropic API key instead.

Build from source

Requirements: Rust (stable), Node.js 22, and on Windows WebView2 (preinstalled on Windows 11). Linux needs libwebkit2gtk-4.1-dev and friends (see CI).

cd app
npm install
npx tauri dev            # development
npx tauri build          # installer in target/release/bundle

Tests

cargo test --workspace                     # context engine + app backend
cargo clippy --workspace --all-targets -- -D warnings
cd app && npm run typecheck && npm test    # frontend
cd app && npx tauri build --no-bundle && npm run e2e   # drives the real app (Windows)
python scripts/i18n_check.py               # missing English strings

Project layout

Path	What
crates/lantern-context	Context engine library, lantern CLI, MCP server
app/	The IDE: Tauri v2 + CodeMirror 6 frontend, Rust backend in app/src-tauri
eval/	Retrieval evaluation and model A/B harness
docs/	Design notes (Korean): context engine, IDE, the task-and-map design, release guide

Language

The UI ships in Korean and English. Source comments and design docs are mostly Korean; issues and pull requests in English or Korean are both welcome.

Contributing

See CONTRIBUTING.md. Security issues: SECURITY.md.

License

Apache-2.0. Third-party notices: THIRD_PARTY_NOTICES.md.