Before the agent edits,see how far the change reaches.

Lantern IDE hands the model only the code each question needs, and for every edit the agent proposes it shows who calls that code, which tests cover it, and how risky it is, before you approve. You choose the model.

Download for Windows0.1.2 · 9.2 MB

Ask in Korean, and it still finds the English code.

The context engine first turns the question into search terms. A general developer glossary widens 중복 (“duplicate”) into duplicate and dedupe, so code without Korean comments still matches. The request below is the one this page follows.

Task

한국어 질문을 검색어로 바꾸는 곳에서 중복을 줄여줘

한국어
질문
검색어
바꾸
곳에서
중복
줄여줘
duplicate
dedupe

Words from the questionAdded by the glossary

Only the code it needs, within budget.

It ranks code from the symbol graph, full-text search, files that changed together in git, and project notes, then trims to the token budget you set. The inspector shows every file it sent, and why.

14 files picked for this question

crates/lantern-context/src/tokenize.rs900
app/src-tauri/src/agents.rs326
app/src-tauri/src/tasks.rs550
crates/lantern-context/src/assemble.rs2,214
eval/ab/run.mjs79
crates/lantern-context/src/main.rs356
8 more3,533

7,958 / 8,000 tokens · assembled in 6 ms

How many answer files each approach retrieves at the same budget

How many answer files each approach retrieves at the same budget	Keyword search	Lantern
Answer files retrieved (mean recall)	24%	67%
Questions with at least one answer file	35%	90%

Measured on one private project (1,772 code files, 20 hand-labelled questions, 8,000-token budget). A public benchmark on an open-source codebase is in progress.

What the agent read stays on the map.

Context sent to the model is teal, files the agent opened itself get a dashed ring, and edited files are indigo. Each task keeps its footprint, so you can come back later and see what it looked at before deciding.

context sent
read
edited

read_file crates/lantern-context/src/assemble.rsRead the caller first

edit_file crates/lantern-context/src/tokenize.rsProposing an edit

Before you approve, see how far this one line reaches.

For every proposed edit it finds the functions that change, then counts who calls them, who calls those callers, the tests in range, and files that usually change together. These numbers were computed on the Lantern repository when this page was built.

Edit filecrates/lantern-context/src/tokenize.rsSample edit

256/// 질문을 검색어 목록으로 바꾼다.

257pub fn query_terms(query: &str) -> Vec<QueryTerm> {

258 let mut seen = HashSet::new();

259 let mut out = Vec::new();

260 let mut push = |text: String, prefix: bool, expanded: bool, out: &mut Vec<QueryTerm>| {

261- if seen.insert(text.clone()) {

261+ if seen.insert(text.to_lowercase()) {

262 out.push(QueryTerm { text, prefix, expanded });

263 }

264 };

265

Impact radiusrisk high

query_terms4direct callers24indirect callers17files No tests

“No tests” means there is no separate test file in range. This function is only covered by unit tests inside its own file.

The real app: approval card with impact radius and Show on map

Isolate a task, and your project stays untouched until you apply.

Each task can get its own git worktree, and the agent edits only there. When it finishes you review the changes and apply them to the project or discard them. Undo asks first if files changed elsewhere in the meantime.

Workspace lantern/tsk-q7m2Sample

1 file changed · tokenize.rsApply to projectDiscard

Many repositories, one view.

It finds every git repository under the folder you open and shows each one’s branch, remote and changes. Fetch, pull, push, commit history and branch switching included. Pull only fast-forwards; if your branch has diverged, it tells you instead of merging.

REPOSITORIESSample workspace

api main↑12

web feat/search5

infra main↓3

docs main

FetchPullPushHistoryBranches

You pick the model, and you can see everything it was sent.

Lantern doesn’t sell a model. Connect the API key you already use, or a local model on your machine.

Anthropic: Claude (API key)
OpenAI-compatible: OpenAI, OpenRouter, vLLM, or any OpenAI-style endpoint
Local: Ollama, LM Studio. Your code never leaves the machine

No hidden system prompt
No telemetry
API keys in the OS credential store
Monthly cost limit

What it can’t do yet

This is a beta built by one person. Worth knowing before you try it.

Code signing: Not yet. On Windows, SmartScreen warns you; choose “More info → Run anyway”.
macOS · Linux: They build and pass tests, but haven’t been used day to day. macOS isn’t notarized, so you have to allow it the first time.
Merges · conflicts: No UI. If a pull diverges, choose merge or rebase in the terminal.
Extensions: VS Code extensions don’t run. It connects language servers (TypeScript, Rust, Python).
Public benchmark: Not yet. The recall numbers above come from one private project.

Open a project you know well, and hand it one small change.

Before approving, read the impact radius on the card and press Show on map. Tell us what worked and what got in the way; it shapes the next beta.

Windows 10 · 11x64 installerTested0.1.2 · 9.2 MB
macOS 11+Apple Silicon · IntelUntested0.1.2 · 23.4 MB
LinuxAppImage · debUntested0.1.2 · 86.2 MB

View on GitHub Send beta feedback

The map behind this page is the Lantern repository itself: 77 files, 956 links (commit 423187f).