CodeWalker

Deterministic, offline code exploration — ground truth for AI agents. CodeWalker carves any git repo into navigable units and extracts factual signals about what the code calls, imports and integrates — LLM call sites (provider + model), frameworks, integrations, prompts, git recency, Taxi schemas — each with file:line evidence, no LLM, no code execution.

It's the deterministic half of a two-step idea:

① CodeWalker explores the repo → facts, every one a file:line. ② your LLM reads those facts and maps the codebase → grounded, cited, not guessing.

It exposes its analysis as an MCP server (so Claude Code and any agent use it as ground truth) plus a local web explorer.

Philosophy

• Signals are ground truth, derived from AST / regex / filesystem / git — never an LLM guess. Classification is a navigational heuristic; the evidence is the truth.
• Every signal carries evidence: a file:line you can jump to. No claim without a location.
• Cheap deterministic substrate: any LLM "judgment" runs on top of this output, not inside it.
• Generalize: nothing is hardcoded to a repo, company or domain — every target is configuration.
• First-class Taxi: .taxi schemas are parsed into a semantic-type usage graph (which models/services consume which semantic types).

What it extracts (per unit)

Size & languages, models referenced (gpt-*, claude-*, …), LLM call sites (provider + model + via + file:line), regex parse sites, frameworks, integrations, an AST-derived call graph (imports / subprocess / http / aws / db / tools / a2a / mcp), prompts + a request table, output formats, key deps, git recency, a heuristic classification + status, a deterministic plain-English summary, an architecture graph, and a full Taxi schema where present.

Run it in 30 seconds

No install needed — point uv at the repo:

# explore any repo in your browser (clones it if you pass a git URL):
uvx --from "git+https://github.com/jhammant/codewalker" codewalker web /path/to/repo

# or run the MCP server so Claude Code can use it:
uvx --from "git+https://github.com/jhammant/codewalker" codewalker mcp /path/to/repo

Install

# from the repo (works today):
pip install "git+https://github.com/jhammant/codewalker"

# from PyPI (once published — see .github/workflows/release.yml):
pip install codewalker          # or:  uv tool install codewalker

# from source (for development):
git clone https://github.com/jhammant/codewalker && cd codewalker
uv venv --python 3.13 .venv && uv pip install -e ".[dev]" && pytest

Requires Python 3.10+. tree-sitter grammars (for deep AST analysis) install automatically via tree-sitter-language-pack.

Use

# Analyze one repo (caches OUTSIDE the repo, in ~/.cache/codewalker)
codewalker analyze /path/to/repo
codewalker analyze /path/to/repo --reindex     # force rebuild

# Estate report / map
codewalker report /path/to/repo
codewalker map /path/to/repo --group-by package

# Taxi schema (global merged "router" view, or a single unit's)
codewalker taxi /path/to/repo
codewalker taxi /path/to/repo <unit_id>

# Agent Pack — a deterministic markdown briefing for a coding agent
codewalker pack /path/to/repo                  # whole repo
codewalker pack /path/to/repo <unit_id>        # one unit
codewalker pack /path/to/repo --out PACK.md

# Run the MCP server over stdio (primary interface)
codewalker mcp /path/to/repo

# Web explorer (live API at http://localhost:8765) — single repo OR a workspace
codewalker web /path/to/repo
codewalker web ~/dev --max-repos 40            # a whole folder of repos

# Bake a shareable offline static bundle
codewalker bake /path/to/repo --out ./bundle

Explore a whole developer's work (a "workspace")

Point CodeWalker at a directory of repos (or a GitHub user) and get a portfolio — every child git repo analyzed, with cross-repo aggregates and a per-repo + portfolio Agent Pack:

codewalker explore ~/dev                       # local folder of repos (offline)
codewalker explore ~/dev --max-repos 20
codewalker explore ~/dev --pack                # portfolio Agent Pack (markdown)
codewalker explore gh:someuser                 # clone+analyze a user's repos (needs the gh CLI)

In the browser you can also point at a new repo on the fly — the topbar's Open box takes a local path or a git URL (cloned + cached), no restart needed.

Screenshots

A unit's deterministic Explained view (every signal a file:line) and the Taxi semantic-type graph:

MCP tools

analyze, explore (workspace portfolio), list_units, get_unit, search, read_file, report, map, taxi_schema, agent_pack, tech_profile — plus read-only resources unit://<id>, report://estate, taxi://<unit_id>. Tool descriptions state explicitly that results are deterministic static analysis with file:line evidence — facts to verify, not prose to trust.

The Agent Pack (agent_pack / codewalker pack) is the deep, deterministic counterpart to a GitHub-metadata "agent pack": a concise markdown briefing of what the code actually does (units, what each calls, the Taxi semantic-type web, frameworks), every claim backed by file:line.

The LLM-analysis layer (run it via Claude Code or any agent)

CodeWalker is the deterministic substrate; the judgment (audits, reviews, dossiers) is an LLM pass on top. Three ways to drive it:

• Claude Code plugin (plugin/): bundles the MCP server + slash-commands and a code-analysis skill. Install:

  /plugin marketplace add jhammant/codewalker
  /plugin install codewalker
  /cw-audit          # audit · /cw-explain <unit> · /cw-taxi-review · /cw-dossier <dir|gh:org>
  

• RECIPES.md: copy-paste analysis prompts for any MCP-capable agent (BYO).
• Agent SDK (examples/agent_sdk_audit.py): a headless Python harness that runs CodeWalker + an LLM audit pass — CI/cron friendly.

Each keeps the contract: facts come from CodeWalker (with file:line), the model adds judgment and verifies before concluding.

Coverage honesty

Languages with a tree-sitter grammar (Python, JS/TS, Go, Java, Rust, Ruby, C#, Kotlin, Scala, Shell, Rego, SQL, HCL) get the deep AST signals; other languages are line-counted and regex-scanned only, and each unit summary says so.

Contributing & license

Contributions welcome — see CONTRIBUTING.md (keep it deterministic, offline, evidence-backed). New detection patterns are usually just a data edit in config.py. Licensed MIT (see LICENSE).