# fittok **Retrieve only the relevant source code for a question — instead of the model reading whole files — so an LLM answers codebase questions on a small, focused slice of context.** Less input = fewer tokens, lower cost, faster answers. Works three ways from one install: an **MCP server**, a **CLI**, and a **Python library** — plus a **Claude Code plugin** that injects context automatically. 📖 **[Full command reference → docs/HANDBOOK.md](https://github.com/likhithreddy/fittok/blob/main/docs/HANDBOOK.md)** mcp-name: io.github.likhithreddy/fittok --- ## How it works ``` codebase ──▶ graphify ──▶ slurp ──▶ readable slice ──▶ LLM answers (parse) (select) (trim to budget) ``` 1. **graphify** — parses the repo with tree-sitter into a knowledge graph of functions / classes / methods (Python, JS, JSX, TS, TSX, Java, Go, Rust). Supports multi-language call/import/reference edges. 2. **slurp** — scores every node against the question using a **4-signal hybrid**: - **Semantic embeddings** (all-MiniLM-L6-v2) — meaning-based matching - **Content-BM25** (with camelCase/snake_case splitting) — keyword matching - **Summary-BM25** (node name + file + callers + callees) — structural matching - **PageRank** — graph centrality / hub detection Signals are fused via **Reciprocal Rank Fusion (RRF)** — rank-based, no score calibration issues. Nodes are selected via **round-robin directory interleaving** (guarantees facet coverage on multi-aspect queries — one node from each code area before any gets a second) with a **per-node token cap** (25% of budget, so large components don't crowd out smaller functions). A **relevance cliff** (semantic OR BM25 OR summary-BM25 threshold) excludes noise. 3. **readable output** — returns the **actual source code** of selected nodes, plus a **codebase map** (table of contents with docstrings, inspired by Karpathy's LLM Wiki / Google's OKF) so the model can route follow-up calls precisely. The model answers directly from it — no file reads needed. As you edit, a file watcher (auto-started on first query) updates the graph **incrementally** — only changed files are re-parsed and merged, and only changed functions re-embed. Graphs and embeddings are cached on disk (`~/.cache/fittok`). Set `FITTOK_AUTOWATCH=false` to disable the watcher, in which case an edit triggers a full re-parse on the next query. --- ## Getting the best results (and known limitations) ### Ask focused questions fittok ranks code against your question using a **4-signal hybrid** (semantic + BM25 + structural + PageRank, fused via RRF) with **round-robin directory diversity** — so multi-facet questions surface code from multiple areas (UI, server, database) instead of clustering in one dominant area. It's most accurate with **focused, specific questions** — ideally one concern each, and naming the function/component/route when you can. Multi-facet questions are supported via **decomposition** (the tool description tells the model to call once per aspect) and the **codebase map** (a table of contents prepended to every response). - ✅ *"How does `runSandboxQuery` execute and isolate a SQL query?"* → surfaces the exact function + its isolation code. - ✅ *"How does the querydle client submit a query and render results?"* → surfaces the UI component. - ✅ *"Trace the full lifecycle: UI submission, sandbox execution, data isolation"* → decomposition + round-robin diversity covers multiple facets; the codebase map routes the model to any missed files. **Rule of thumb:** one concern per question (or 2–3 facets max). For "explain the whole feature," split it into a few focused questions instead of one mega-query. ### Known limitations - **GitHub Copilot Chat truncates large MCP outputs (the big one).** Copilot caches MCP tool results above ~7 KB to a `content.json` file, where the entire markdown collapses to ONE physical JSON line (newlines escaped) — and its Read tool truncates any line at ~2,000 characters. So an output over ~7 KB is effectively chopped to ~2,000 chars *regardless of total size*; the model can't see most of the code and falls back to reading source files directly. This is Copilot's delivery layer, not fittok — **every MCP server hits this wall.** By default (0.10.0+) fittok returns *all* relevant code uncapped, which is correct for clients that deliver inline but will be truncated by Copilot. Two workarounds: - **Cap the output for Copilot** so it's delivered inline (under the ~7 KB threshold). Set `FITTOK_MAX_BUDGET=1200` in your MCP server's env: ```json { "servers": { "fittok": { "command": "uvx", "args": ["fittok"], "env": { "FITTOK_MAX_BUDGET": "1200" } } } } ``` - **Use Claude Code or the CLI** for multi-file questions. They deliver MCP output inline with no truncation — which is where fittok's complete (uncapped) results and anti-re-read design actually pay off. - **Vocabulary gap on abstract queries.** When the query uses words that don't appear in the code (e.g. "isolation" → `REVOKE`/`DENY`), neither semantic nor BM25 can bridge it. The codebase map (file names + docstrings) and round-robin diversity help; naming the function/file routes the model to it. - **Incremental edge-loss:** editing a file can drop call/import edges *into* it from unchanged files until a full re-parse. fittok auto-recovers on restart or `reset_graph`. - **Token counts are approximate:** counts use `cl100k_base`, so real usage drifts ~10–20% vs Claude's tokenizer (only matters when you opt into a `FITTOK_MAX_BUDGET` cap). - **Very large repos:** PageRank is not yet vectorized — fine through low-thousands of nodes, slower beyond that. --- ## Installation fittok ships as an **MCP server**, a **CLI**, and a **Python library**. It uses `torch` for embeddings, so a Python runtime must be present. **Pick one runtime** below, then follow the section for your client. > Every config below launches fittok as `uvx fittok`. If you chose Python or > `pipx`, swap that for `python -m fittok` or `pipx run fittok` respectively. ### Prerequisites — choose a runtime (one of) **A. `uv` — recommended (no Python needed on the machine)** ```bash curl -LsSf https://astral.sh/uv/install.sh | sh # Linux / macOS winget install astral-sh.uv # Windows brew install uv # macOS (Homebrew) ``` Launch command: `uvx fittok` — `uv` provisions its own Python + all deps in isolation. One static binary, so it's deployable org-wide via MDM/Intune/winget. **B. Python 3.10+ (already on the machine)** ```bash python -m pip install fittok # Linux / macOS py -m pip install fittok # Windows ``` Launch command: `python -m fittok` (Windows: `py -m fittok`). > Managed Linux may reject `pip install` with PEP 668 > ("externally-managed-environment") — use option A to avoid it. **C. `pipx` — isolated, no global install** ```bash brew install pipx # macOS pip install --user pipx && pipx ensurepath # Linux / Windows ``` Launch command: `pipx run fittok`. ### MCP server — Claude Code ```bash claude mcp add fittok -s user -- uvx fittok ``` Restart Claude Code → `/mcp` → confirm `fittok` is **connected**, then ask codebase questions normally. ### MCP server — VS Code / GitHub Copilot Chat ```bash code --add-mcp '{"name":"fittok","command":"uvx","args":["fittok"]}' ``` Or paste into `.vscode/mcp.json` (workspace) or your user `mcp.json`: ```json { "servers": { "fittok": { "type": "stdio", "command": "uvx", "args": ["fittok"] } } } ``` Then in Copilot Chat: **Agent** mode → enable fittok's tools (*Configure Tools*). ### MCP server — GitHub Copilot CLI ```bash copilot mcp add fittok -- uvx fittok copilot mcp get fittok # verify status + tools ``` ### MCP server — Cursor / Windsurf / any MCP client ```json { "mcpServers": { "fittok": { "command": "uvx", "args": ["fittok"] } } } ``` ### Auto-trigger (optional, every MCP client) To make fittok fire on **every** codebase question — without naming it — **and** stop your client from re-reading files fittok already returned (which would discard the savings), add this one line to your client's instructions file: > *"For any codebase question, call fittok first and answer from its output — > don't re-read files it already returned code from."* The first half triggers fittok; the second keeps the client from opening the same files afterward. They reinforce each other — one shapes *strategy* (use fittok), the other stops the *double-read*. For a stronger, more explicit block: > For any codebase question ("how does X work", "where is Y"): > 1. Call the fittok MCP tool first, once. > 2. Answer directly from its `optimized_context` — it is the real, authoritative > source for that question. > 3. Do NOT read or grep the files fittok already returned code from. That > discards the token savings fittok exists to provide. For the strongest effect, put it in your **user-global** instructions so it applies to every repo, not just one: | Client | Instructions file | |---|---| | Claude Code | `CLAUDE.md` (repo) or `~/.claude/CLAUDE.md` (user-global) | | GitHub Copilot | `.github/copilot-instructions.md` or Copilot user instructions | | Cursor | `.cursor/rules/*.mdc` (or `.cursorrules`) | | Windsurf | `.windsurfrules` | > fittok also bakes this rule into every response (an "answer from this, don't > re-read" line above the code), so it works even without the snippet above — > the snippet just makes it the client's default across all questions. ### CLI ```bash cd /path/to/your/repo uvx fittok index # optional pre-warm (~15s, cached) uvx fittok query "how does auth work" # LLM answers from relevant code uvx fittok query "how does auth work" --budget 1500 # cap the slice at 1500 tokens uvx fittok query "how does auth work" --code # raw relevant code, no LLM uvx fittok graph # interactive browser graph uvx fittok graph --query "auth" # graph with relevant nodes highlighted ``` `query` sends the relevant code slice to an LLM and streams the answer. Set one key in your shell and it just works: ```bash export ANTHROPIC_API_KEY="sk-ant-..." # → claude-haiku-4-5 (recommended) export OPENAI_API_KEY="sk-..." # → gpt-4o-mini (fallback) ``` Users of Claude Code already have `ANTHROPIC_API_KEY` set — no extra step needed. If neither key is set, fittok falls back to `--code` and prints a setup hint. `graph` requires `pyvis`: `uv pip install "fittok[ui]"`. ### Python library ```bash uv add fittok # in a uv project (or: uv pip install fittok in a venv) ``` ```python from fittok import optimize result = optimize("/path/to/repo", "how does authentication work", token_budget=1500) print(result["optimized_context"]) # the relevant code slice print(result["savings"]) # token reduction stats ``` ### Upgrading `uvx` caches the environment, so a new `fittok` release isn't picked up automatically — server restarts reuse the cached version. Upgrade with one command (no need to re-register the MCP server): ```bash uvx --refresh fittok # re-resolve from PyPI → latest version ``` Then restart the MCP server (reload the window in VS Code, or restart the Copilot CLI) so it boots the new version. For other runtimes: - **pip:** `python -m pip install --upgrade fittok` - **pipx:** `pipx upgrade fittok` --- ## Why tree-sitter, not LSP? fittok uses **tree-sitter** (fast, syntactic AST parsing) instead of **LSP** (Language Server Protocol — semantic analysis with types, cross-file references, go-to-definition). This is a deliberate tradeoff: | | tree-sitter (fittok) | LSP (e.g. Serena MCP) | |---|---|---| | **What it returns** | Actual source code — the model answers directly | Symbol metadata (names, refs, types) — the model must still Read files | | **Setup** | Zero config, works on any directory | Needs language servers installed + project config (tsconfig, pyproject, etc.) | | **Languages** | 8 out of the box (Python, JS/TS/TSX, Java, Go, Rust) | Only as many as LSP servers you've installed | | **Startup** | ~15s (parse + embed) | Minutes (full project indexing per language) | | **Memory** | ~100 MB (graph + embeddings) | 500 MB+ per language server | | **Model calls per question** | 1–5 (one-shot retrieval) | 5–20+ (iterative symbol navigation) | | **Token cost** | ~2,500 tokens (code delivered directly) | ~15,000+ tokens (metadata + file reads) | **fittok's USP is token savings.** It returns the actual code in one call so the model doesn't need to read files. LSP-based tools return metadata (symbol names, reference lists) — precise, but the model still has to open the files to see the implementation. More round-trips, more tokens. **The tradeoff:** tree-sitter can't resolve cross-file references as accurately as LSP (a `fetch("/api/run")` call in a .tsx file won't perfectly link to the route handler). fittok compensates with 4-signal retrieval (semantic + content-BM25 + structural summary-BM25 + PageRank, fused via RRF) and round-robin directory diversity — which cover the gap in practice. **Complementary, not competitive:** LSP-based tools like [Serena](https://github.com/oraios/serena) excel at symbol-level navigation ("find all callers of `runSandboxQuery`"). fittok excels at semantic retrieval ("how does SQL execution work?"). Install both — the model picks the right tool per task. --- ## Token savings — honest numbers On a real Next.js/TS repo (~5k functions), fittok returns a **~1.5–3.5k-token slice** instead of the model reading **15–20k+ tokens** of files — an **~80–90% reduction on input**, deterministic and reported in the `savings` footer. On Opus 4.8, a broad question cost **~84k total tokens without fittok vs ~27k with it** — because fittok replaced a 58k-token Explore subagent with one tool call. **How to measure it honestly:** - Use the **`🪙 saved X%` footer** or your **API bill** (total tokens). - Do *not* judge by Claude Code's `/context` Messages number — it excludes subagent tokens and is dominated by model reasoning, which fittok doesn't touch. --- ## Configuration | Variable | Default | Description | |---|---|---| | `ANTHROPIC_API_KEY` | — | Enables LLM answers via `claude-haiku-4-5` | | `OPENAI_API_KEY` | — | Fallback LLM via `gpt-4o-mini` | | `FITTOK_SHOW_SAVINGS` | `true` | `🪙 saved X%` footer on MCP answers; set `false` to disable | | `FITTOK_MAX_BUDGET` | `0` (unlimited) | Code-token cap. `0` = return ALL relevant code in full (default — complete results in Claude Code / CLI). Set `1200` for **GitHub Copilot**, which truncates MCP outputs >~7 KB (see Known limitations). | | `FITTOK_AUTOWATCH` | `true` | Auto-start the file watcher so graph updates are incremental (only changed files re-parse); set `false` to fall back to full re-parse on edits | | `FITTOK_EMBED_MODEL` | `all-MiniLM-L6-v2` | Embedding model | | `FITTOK_DEVICE` | `auto` | `auto` / `cuda` / `mps` / `cpu` | | `FITTOK_CACHE_DIR` | `~/.cache/fittok` | Cache location | Full reference: **[docs/HANDBOOK.md](https://github.com/likhithreddy/fittok/blob/main/docs/HANDBOOK.md)** --- ## Requirements Python ≥ 3.10. First run downloads a ~90 MB embedding model. Graph visualization (`fittok graph`) is included by default. Optional extras: - `uv pip install "fittok[ui]"` — Gradio web dashboard (`launch_ui` tool) - `uv pip install "fittok[gpu]"` — torch/CUDA for GPU-accelerated embeddings ## License MIT