Cranot/roam-code

What it solves

Roam-code provides a local codebase intelligence layer for AI coding agents to reduce the number of turns, tokens, and costs associated with codebase navigation and comprehension. It prevents agents from performing expensive, repetitive search operations (like multiple greps) by providing them with pre-resolved code-graph facts before they begin editing.

How it works

Roam-code indexes a local repository into a SQLite-backed graph of symbols, calls, imports, and git history across 28 languages. It uses a "task compiler" that recognizes the type of question an agent is asking and injects the answers (e.g., caller lists, filtered git history) directly into the prompt.

Additionally, it implements a verification loop that runs after every edit to check for naming convention violations, import hallucinations, syntax errors, complexity, secrets leaks, and algorithmic patterns. It can be integrated as an MCP server for tools like Claude Code, Cursor, and Continue.

Who it’s for

Developers using AI coding agents who want to reduce LLM costs and latency while ensuring that changes are safe and mergeable through local, credential-free analysis.

Highlights

• Task Compiler: Pre-resolves repo facts to reduce agent turns (up to 83% reduction in some benchmarks) and input tokens.
• Local-First: No API keys or cloud logins required; source parsing and indexing stay on the machine.
• Tamper-Evident Evidence: Generates signed ChangeEvidence packets that document the context, risk, and verification of a change.
• Verification Gate: Automatically checks for hallucinations, secrets, and structural issues after an edit.
• Broad Language Support: Supports 28 languages using a SQLite-backed code graph.