LemonBeam generates GitHub repo onboarding guides with verifiable citations

What LemonBeam does

LemonBeam, a newly launched open-source tool, scans a public GitHub repository and produces an onboarding guide in which every statement is backed by a citation pointing to the exact file and lines it came from. According to the launch post on dev.to, the project was built by Kanami Anderson, John DeLong, and Tanisha Thompson-Jones and is released under the MIT License.

The problem it targets

Anyone joining an unfamiliar codebase normally has to assemble a mental model from READMEs, package scripts, configuration files, and source code. The shortcut many people take — asking a general-purpose chatbot — often produces answers that look right but cannot be checked. The developers write that when they audited those answers, the citations were invented, pointed hundreds or even thousands of lines away from the relevant code, or were missing entirely. Their conclusion: an answer you cannot verify is still a guess.

How it works

LemonBeam's central design decision is that deterministic rules, not the model, decide what evidence enters the prompt. The tool downloads a repository at a specific commit, then filters, parses, and chunks the code using Tree-sitter, scoring each chunk against a per-section token budget. A single AI call then writes the guide using only that pre-selected evidence.

Citation checking runs after generation. A separate validator strips out any citation that does not match the material the model actually received, and anything that could not be read or included is surfaced in a dedicated Uncertainties section rather than quietly omitted.

Every guide follows the same six-part structure: Project Overview, Setup/Installation, Running Locally, Project Structure, Testing, and Uncertainties and Missing Information. Guides are pinned to the commit they analyzed, so different team members can regenerate the same document and check it line by line.

Three ways to run it

The announcement describes three interfaces:

• a web app at lemonbeam.app, where you paste a public GitHub URL and supply your own OpenRouter API key
• a CLI, which works in a local folder or against a public GitHub URL and saves the guide as Markdown
• an MCP server that can be added to tools such as Claude Code or Codex so an assistant can produce onboarding guides on request

Privacy and a costly bug

LemonBeam keeps no database, according to the dev.to post. Downloaded repositories live in a temporary folder that is deleted when the scan ends, even when the scan fails, and API keys exist only in memory for a single request. The only code that leaves the machine is the evidence sent to the model through OpenRouter.

The team also recounts an engineering lesson from before launch: an early version of the pipeline was silently sending every code chunk to the model, which made guides expensive and capped the supported repository size. Fixing that bug reportedly cut the cost per guide from roughly 70 cents to about 2 cents and raised the maximum supported repository size from 2 MB to 50 MB.

What's next

Planned work includes splitting generation into separate, parallel AI calls with tuned prompts for each section, tightening citation checking so it verifies that cited lines actually support the claim, and adding support for larger repositories, monorepos, private repositories, and languages beyond JavaScript and TypeScript. The team also wants to compare embeddings-based retrieval against its current rule-based approach.

Contributions are welcome, and the suggested entry point is to run LemonBeam against its own repository under the oslabs-beta GitHub organization, then consult the contributing guide in the repo.

Why it matters

AI-generated documentation is only useful if you can trust it, and trust depends on being able to check the answer. By constraining the model to rule-selected evidence, validating citations mechanically, and exposing gaps instead of hiding them, LemonBeam turns "the AI said so" into something closer to a reviewable artifact. Pinning each guide to a specific commit also makes the output reproducible, which matters for teams whose onboarding material needs to stay accurate as a codebase evolves. More broadly, the project illustrates a useful pattern in applied AI: rather than waiting for smarter models, engineering the retrieval and verification layer around a model can deliver the reliability that raw chatbot answers still lack.