Ziplex

English | 한국어

Which tests does this change affect, and what depends on this file? Ziplex answers from a zero-config dependency graph — for your CI, your editor hook and your coding agent.

Ziplex parses a project with Tree-sitter and resolves imports the way the toolchain does — relative/index files, tsconfig paths aliases, pnpm/npm workspace packages and their exports, Python packages, re-exports, conftest.py and package __init__.py files, Go packages — into one graph across languages. It keeps that graph in an incremental local index (no LLM, no network, no API key) and answers from it in milliseconds.

⚠️ Under active development. Interfaces may still change.

Where it helps

• Running only the tests a change can affect — locally (pytest $(ziplex impact --format paths)) or on every pull request (CI template). Measured on 99 real commits of repos Ziplex was never tuned on: recall 0.976, or 0.988 without the imports pytest itself makes (eval). It falls back to the full suite whenever a targeted run isn't safe.
• Giving a coding agent the impact answers grep can't produce in one step — ziplex-mcp answers "which tests cover this change", "what is transitively affected", "how does A reach B". On held-out repos, Claude Code with it spent 57% less and finished 55% faster on those questions, at the same accuracy (EXPERIMENTS #15).
• Polyglot and monorepo codebases — a TypeScript frontend, a Python backend and shared workspace packages resolve into one graph, with no per-language tool to configure.
• An editor hook — after Claude Code edits a file, one line tells it how many files depend on it and which tests to run.

Honest scope: for a direct "who imports X" question a single grep is already enough (the MCP costs ~12% more there), and in small, local bug fixes agents gained nothing measurable from it (EXPERIMENTS #16). The value is in transitive impact and test selection.

Quick start

pip install ziplex                      # or: pipx install ziplex
cd your-project
ziplex impact                           # tests affected by your uncommitted changes
pytest $(ziplex impact --format paths)  # run just those (prints nothing -> full suite when unsure)
ziplex deps src/app.py --in --depth 2   # who depends on a file, by hop
ziplex install claude --hook            # register the MCP server + post-edit hook with Claude Code

The first command builds .ziplex/index.db (gitignored automatically; django's 5,700 files take ~10s, later updates under a second). More in Graph index, MCP server and the CI template.

Packing a project into aif.json (optional)

Ziplex can also pack a project into aif.json: per-file summaries (structural, or written by an LLM you choose), inferred coding rules, a project guide and the relationship graph, reviewable in a GUI before it's saved. Useful for pasting a project's shape into a chat with no tool access; for agents that have tools, the graph commands above are what measurably helps (EXPERIMENTS #1-#9).

Pack quick start

pip install ziplex        # or: pipx install ziplex (isolated env, PATH handled for you automatically)
# or: venv\Scripts\activate && pip install -e . from a clone
ziplex-gui

A native window opens: pick a project folder with the folder picker, check off which files to include (or just accept the safe default), and pack — no flags to remember. Want to try it before committing to an API key? Check "no LLM" and you get a real aif.json from structural summaries alone — no signup, no network call. Uncheck it (with a GEMINI_API_KEY in .env, or another provider picked on the Options page) for real AI-written summaries, inferred coding rules, and a project guide instead. Packing pauses for review before anything's saved, and browsing a project you've already packed works the same way — see GUI for the full walkthrough.

ziplex/ziplex-gui command not found after installing? pipx avoids this by design — it installs into its own isolated environment and puts that on your PATH automatically (pipx ensurepath, then restart your terminal). With plain pip install ziplex on Windows, if you don't have write access to your global Python's site-packages (common without admin rights), pip silently falls back to a per-user install whose Scripts folder often isn't on PATH — you'll usually see a WARNING: The script ziplex.exe is installed in '...' which is not on PATH line scroll by during install. Fix: either switch to pipx install ziplex, or add that printed folder (typically %APPDATA%\Python\Python3XX\Scripts) to your user PATH yourself and open a new terminal.

Prefer the terminal — scripting, CI, or an environment with no GUI?

ziplex pack ./your-project/ --auto --no-llm

ziplex pack ./your-project/                        # full pipeline, interactive
ziplex pack ./your-project/ --auto --auto-correct  # fully non-interactive (CI, scripted runs)

Config file

ziplex init ./your-project/ scaffolds .ziplex.json in the target project (not in Ziplex's own repo) so include/ignore glob patterns don't need retyping on every pack:

// your-project/.ziplex.json
{
  "include": ["src/**/*.py", "*.md"],  // empty = everything not ignored (default)
  "ignore": ["**/*.generated.*"]        // extra patterns beyond DEFAULT_IGNORE/.gitignore
}

--include/--ignore CLI flags add to this file's patterns rather than replacing them. Every subcommand that previews what pack would collect (collect, tree, tokens, search, freshness) reads the same file, so none of them drift out of sync with what a real pack on that project would actually see. Worth committing alongside the project the same way aif.json/detail.json already can (see Team use below) — it documents how that project gets packed.

How it works

Ziplex collects a project's files (skipping build artifacts and anything .gitignored), security-scans them for secrets, and parses the safe ones with Tree-sitter to pull out signatures, imports, and API routes. Function bodies get compressed down to a single marker; an LLM then writes a one-line summary per file plus project-wide coding rules and an AI-facing guide. A human reviews and edits all of that — only low-confidence summaries get flagged, not every file — before the final aif.json (small, loaded up front) and its heavier sibling detail.json (full compressed body, fetched on demand) are saved.

Output format

A pack produces three files, each read differently:

// aif.json — small, loaded up front
{
  "project": {
    "name": "...", "prompt": "...",
    "tech_stack": [{ "manifest": "package.json", "language": "JavaScript/TypeScript", "package_manager": "npm", "dependencies": ["react", "..."], "dependencies_truncated": false }],
    "security_scan": { "flagged": 0, "included_anyway": 0, "excluded": 0 },
    "format_notes": "..."  // fixed legend explaining confidence/⋮---- to a reader with no other Ziplex context
  },
  "rules": ["..."],
  "folders": { "src": { "summary": "..." } },  // one sentence per folder describing its role
  "tokens": { "GPT-4o": { "original": 3100, "compressed": 749, "saved_pct": 75.8 } },
  "files": { "src/App.tsx": { "summary": "...", "confidence": 0.83 } },
  "relationships": { "src/App.tsx": { "internal": ["..."], "external": ["react"] } }
}

// out.detail.json — heavier, fetched only when a file actually needs a closer look
{
  "src/App.tsx": { "compressed": "import React ...\n    ⋮----\nexport default App" }
}

// out.cache.json — internal bookkeeping, not meant for an AI to read; a
// content-hash snapshot of what was packed, for check_freshness to diff
// against later
{
  "src/App.tsx": "3b1c2e...(sha256)"
}

Features

• Multi-language, structure-aware compression — Python, Java, TypeScript, JavaScript, Lua, GDScript, Go, C++, Rust, C#, PHP, Ruby, Bash via Tree-sitter, plus dedicated JSON/YAML/Markdown/plain-text/Dockerfile compressors, all preserving structure while cutting tokens. Works on any collection of local files with cross-file relationships, not just git repos — game mods and asset projects included.
• Security scanning built in — every file is checked for secrets (secretlint, regex fallback) before it ever enters the pipeline.
• Human-in-the-loop, but it scales — every LLM output (summaries, rules, project guide, dependency tree) is reviewable and editable, or skippable entirely with --auto-correct. Only low-confidence summaries get flagged for review — review time doesn't grow with project size.
• Three ways to consume the result — an MCP server for Claude Code/Cursor/etc., a local GUI for packing and browsing without a terminal, or a Claude Agent Skill export that needs no server at all.
• Cheap to keep current — incremental re-pack only re-summarizes files that actually changed (content hash), retries with backoff on LLM flakiness, and checkpoints a failed run instead of restarting from scratch.
• Bring your own LLM — Gemini by default, or switch to any OpenAI-compatible endpoint (real OpenAI, Ollama, LM Studio, vLLM, OpenRouter, ...) or Claude directly, via LLM_PROVIDER or the GUI's Options page — including a fully local setup with no API key or network call at all (see Other AI providers).
• No API key required, if you want — pack --no-llm skips every LLM call and still produces structural summaries, dependency graphs, and tech-stack detection. Scope any pack with .ziplex.json/--include/--ignore, and guard CI budgets with --max-tokens.

Testing

pip install -e ".[dev]"   # adds pytest on top of the base install
pytest

Covers the deterministic core — compressors, the Tree-sitter extractor, the collector's ignore/binary-file filtering, the dependency-graph operations (build_tree/has_cycle/move_file), and the pure aif-editing API — plus a full pack() run against a network-free MockProvider instead of Gemini, exercising checkpointing, parallel summaries, and token counting end to end without the cost or latency of a real LLM call.

Want to smoke-test pack against a real project without waiting on Gemini? LLM_PROVIDER=mock ziplex pack <project> --auto --auto-correct runs the whole pipeline network-free in under a second.

Graph index: deps, impacted tests, editor hook

No pack and no LLM needed: these commands keep an incremental dependency-graph index in <project>/.ziplex/index.db (a disposable cache, gitignored automatically) and answer from it. The project defaults to the current directory.

ziplex index [--check]                        # build/refresh the index; --check exits 1 if anything changed
ziplex deps src/app.py                        # what it imports (with line numbers)
ziplex deps src/app.py --in --depth 2         # who depends on it, by hop
ziplex impact                                 # tests affected by your uncommitted changes (= --diff worktree)
pytest $(ziplex impact --format paths)        # run just those tests
ziplex path src/app.py src/db.py              # shortest import chain
ziplex install claude --hook                  # register the MCP server + a post-edit hook with Claude Code

impact --format paths prints nothing when a targeted run isn't safe (a build/config/lockfile changed, or a file was deleted), so the pytest $(...) line above falls back to the whole suite; the reason goes to stderr. --format commands prints ready-to-run pytest/jest/go test commands instead.

With --hook, ziplex install claude adds a PostToolUse entry (ziplex-hook) to <project>/.claude/settings.json. After Claude edits a file it tells Claude, in one short line, how many files depend on it and which tests to run. It never blocks, stays silent in projects without an index, and repeats at most once per file per 10 minutes. --dry-run shows the command and the settings diff without changing anything.

MCP server

ziplex-mcp gives Claude Code, Cursor or any MCP client the same live graph as the commands above -- no pack needed, refreshed from disk before every call.

claude mcp add ziplex -- ziplex-mcp --project /abs/path/to/project   # or: ziplex install claude

Answers are compact text (tens to a few hundred tokens), and the five tool definitions cost ~720 tokens per session. Paths can be repo-relative, absolute or a unique suffix (query_service.py).

Legacy mode. ziplex-mcp --aif <name>.json [--project .] (or --legacy-tools) serves the previous ten tools over a packed aif.json instead (get_overview, list_files, get_folders, get_relationships, get_graph_summary, get_dependents, get_blast_radius, get_detail, check_freshness, search_project), so existing registrations keep working.

GUI

A local, single-user GUI for two situations: packing a project without touching a terminal, and browsing an already-packed project somewhere Claude Code (or MCP generally) isn't available but a browser-based AI chat is.

ziplex-gui                                            # native window (pywebview)
ziplex-gui --aif out.json --project ./your-project/   # prefill the landing page
ziplex-gui --no-window                                # plain browser tab instead

Packing a project

Pick a project folder with the native folder picker, check off which files to include (the same safe/dangerous split collect's security scan produces), pick the packing language (pack --lang's GUI equivalent — what language summaries/rules/the AI guide are written in, separate from the topbar's own display-language toggle), optionally check "no LLM" (pack --no-llm's GUI equivalent), and watch the pack run in the background.

Reviewing before anything is saved

Analysis pauses for review before anything is saved: edit the project name, guide, rules, and folder/per-file summaries (only the low-confidence ones are flagged, same triage the CLI uses).

Dependency graph & relationship editing

The dependency graph opens as a collapsible whole-tree overview first — click a file's name once you've spotted one worth fixing to drop into an edit view for just that file, linking or unlinking individual edges (a file with more than one real parent keeps its other references intact). The Relationships page (for a project already packed) reuses this same whole-tree-then-edit-one-file flow, so a relationship noticed after the fact can be fixed without re-running the pipeline.

Browsing an already-packed project

The same overview/files/relationships/detail/search views the MCP server exposes, as web pages instead of MCP tool calls. Each page has a Copy button; the intended flow is looking around here and pasting what's useful into a separate AI chat by hand, not the GUI talking to that chat itself.

Debugging in a plain web chat (no file access, no MCP): paste aif.json's content once at the start of the conversation — it's small on purpose, just summaries and relationships, meant to be pasted whole. Once the AI points at a suspect file from those summaries/relationships, don't paste the rest of detail.json too — open that one file's page here and hit its own Copy button (or run ziplex detail <name>.detail.json <file-key> for just a line range of a large file) and paste only that into the same chat. The AI now has that file's full compressed source without you ever handing over anything for the files it didn't ask about.

Options page

The default output folder new packs save to, and which AI provider every pack uses (Gemini / OpenAI-compatible / Claude — see Other AI providers), with one-click presets for Ollama's and LM Studio's default local ports so a fully local setup needs no typed URL. Display language and light/dark theme live in the topbar itself instead (two small toggles next to the nav links) — a display-only chrome preference, not something worth a trip to a separate page for.

Binds to 127.0.0.1 only — no --host flag, no way to expose it to a network.

Claude Agent Skill export

A third way to reach a packed project, alongside the MCP server and the GUI — for when Claude Code is available but registering an MCP server isn't (or is more setup than the moment calls for):

ziplex skill .ziplex/my-project.json               # writes .claude/skills/my-project/
ziplex skill .ziplex/my-project.json -o some/dir    # custom output directory

Generates a Claude Agent Skill — SKILL.md plus references/overview.md/files.md/relationships.md/detail.json — that Claude Code discovers and progressively loads on its own once it sits under .claude/skills/, no server process required. Unlike repomix's equivalent --skill-generate feature, references/files.md never embeds full raw source — it stays to summaries and confidence scores, exactly what aif.json itself already restricts to; the compressed body only ships as references/detail.json. Committing the generated directory works the same way committing aif.json/detail.json does (see Team use below) — it's just files.

Team use

Ziplex packs one person's local snapshot — there's no shared server or live sync. That doesn't rule out team use, though: aif.json/detail.json/cache.json are just files, so a team can commit them to the project's own repo the same way any other generated-but-versioned artifact works.

• Whoever runs pack after a meaningful change commits the refreshed output alongside their code change.
• Everyone else's copy is only as current as the last commit that touched it — check_freshness (or get_overview/list_files with project_path passed, see above) tells them whether their working copy has drifted from what's committed, without re-packing just to find out.
• This is a convention, not a feature Ziplex enforces or coordinates: no merge/conflict resolution, and no defined "who wins" if two people repack independently before either commits. If it does conflict, don't hand-resolve the JSON — see merge-conflicts.md for the recommended fix (clear the conflict, re-pack, review); gitattributes keeps these files collapsed in GitHub's PR diff view.

CI gate (optional): ziplex freshness <project> <name>.cache.json exits non-zero when the committed output has drifted — no LLM calls, no GEMINI_API_KEY, just a hash comparison. Wire it into a PR check (template) to catch a stale, un-re-reviewed aif.json before it merges, without automating pack itself (which costs real LLM calls and skips human review — left as a manual step on purpose).

Tech stack

Python 3.10+ · Tree-sitter (Python/Java/TypeScript/JavaScript/Lua/GDScript/Go/C++/Rust/C#/PHP/Ruby/Bash grammars, GDScript via tree-sitter-language-pack) · tiktoken · ruamel.yaml · plain REST via requests against Gemini (gemini-flash-latest by default, overridable via GEMINI_MODEL), any OpenAI-compatible endpoint (real OpenAI, Ollama, LM Studio, vLLM, OpenRouter, ...), or Claude — see Other AI providers · MCP · Flask · pywebview · secretlint · pathspec

License

MIT