Architectural Context for LLMs, Minus the Token Burn

What it does

Codemap is a Go CLI that scans a codebase and synthesizes architectural context—file trees, dependency graphs, diffs, and change summaries—into compact, LLM-ready payloads. It is designed to give coding agents like Claude or Codex immediate project awareness without stuffing their context windows with irrelevant source code. You can point it at a local repo or a public GitHub/GitLab URL and get structured output in text, JSON, or an HTTP API response.

The interesting bit

The standout idea is the multi-agent handoff artifact: a deterministic, layered snapshot split into stable hub summaries (prefix) and dynamic recent changes (delta), complete with content hashes and cache-reuse metrics. This lets one agent pick up exactly where another left off without re-ingesting the entire repo. The tool also treats context limits as a first-class constraint, offering per-project byte budgets and subsystem routing rules in .codemap/config.json.

Key highlights

• Layered handoffs with deterministic hashes (prefix_hash, delta_hash) and an append-only metrics log to track cache reuse across agent sessions.
• Dependency flow across 18 languages via ast-grep, surfacing import hubs and blast radius for changed files.
• Built-in budgets and routing in .codemap/config.json, letting projects cap bytes per hook and define keyword-driven subsystems.
• Remote repo analysis via shallow clones of public GitHub/GitLab repos without manual cloning.
• MCP server, HTTP API, and Claude hooks for integration beyond the terminal.

Caveats

• The README is truncated mid-sentence during the “Skills” section, so the full scope of that feature is unclear.
• Dependency analysis requires installing ast-grep separately unless you use the codemap-full release artifact; the base tarball is not self-contained.
• Several features (drift detection, subsystems, skills) are mentioned but not fully documented in the provided excerpt, leaving some behavior ambiguous.

Verdict

Worth a look if you routinely hit context limits with Claude Code or Codex on large or complex repos. Probably overkill if you just want a pretty tree replacement for human reading.