Frequently Asked Questions

Fuzit is a local-first repository intelligence engine for AI-native development. It parses your source code into a typed graph, understands relationships, and generates bounded, task-specific context that you can feed into any LLM or AI coding tool.

Most coding agents rely on dumb text search or embedding similarity to find files, often missing crucial structural dependencies or overflowing your token budget. Fuzit maintains a persistent, typed graph of your repository, allowing it to retrieve context deterministically based on actual code semantics, not just keywords. See our architectural argument for details.

No. Repository packers simply concatenate all files into a single payload, which destroys context limits and dilutes LLM attention. Fuzit is a context selector; it builds a highly specific subset of your repository tailored to the exact task. Compare the approaches on our Comparison page.

Fuzit's core indexing, graph traversal, and context selection are entirely deterministic and run locally without an LLM. It prepares the context *for* an AI, ensuring the preparation phase is fast, private, and mathematically verifiable.

No. Fuzit operates entirely local-first. Your repository source code is analyzed, indexed, and stored on your local machine. It does not require a cloud backend to function.

Fuzit collects zero telemetry. Your repository graph, task queries, and environment details remain strictly on your machine. Learn more on our Security page.

Yes. Fuzit applies security filtering before disclosure. Candidate evidence is scanned against Secretlint rules and sensitive path policies (like `.env` files). Omitted files are securely blocked and logged in the explainability report.

No. Fuzit performs static analysis to build its typed graph. It does not execute the source code it is analyzing.

Fuzit currently provides deep typed graph analysis for TypeScript, JavaScript, Python, Java, and Go. It also recognizes strong frameworks within those languages (like React, Next.js, and FastAPI).

After the initial cold index, Fuzit runs a local watcher. When a file changes, it only updates the affected nodes in the graph rather than rebuilding the entire repository model. This makes repeated queries near-instant.

Instead of treating code as plain text, Fuzit builds a relationship map: it knows that a specific Symbol is exported by a File, consumed by a Service, and verified by a Test. This allows Fuzit to traverse edges to find related context that text search would miss.

Yes. You can specify a strict token budget. Fuzit will rank candidate evidence and truncate the output to fit precisely within your limit, prioritizing the most structurally relevant information first.

Yes. Fuzit produces provenance data for its context bundles. It will explicitly tell you if a file was selected due to a graph edge, semantic relevance, or if it was truncated due to budget constraints.

Yes, but only when explicitly authorized. Fuzit uses explicit network intent. It will not contact GitHub for a local directory, but if you run a command against a GitHub URL or a PR, it will fetch the required remote state.

Yes. Fuzit's flagship remote workflow (`fuzit review`) fetches the PR title, body, diffs, changed files, review comments, and check statuses. It engineers this evidence into a single context bundle ready for architectural AI review. See Examples.

Yes. You can use an issue URL as the task input. Fuzit will fetch the issue discussion and use it to select the relevant implementation context from the repository.

Fuzit uses anonymous public access where possible. For private repositories or elevated rate limits, it resolves `FUZIT_GITHUB_TOKEN` or `GH_TOKEN`. It does not accept tokens via CLI arguments to prevent secret leakage in shell history.

Yes. Fuzit provides a read-only local stdio Model Context Protocol (MCP) server. Compatible AI clients (like Claude Desktop) can connect to it to query your local repository intelligence.

Yes. The MCP surface is strictly read-only by design. It can query the repository graph and fetch context, but it cannot mutate your filesystem.

Yes. Fuzit is model-neutral. Because it outputs standardized markdown or JSON, and supports the MCP standard, its context bundles can be injected into any major AI tool or API.

Fuzit offers a restricted, out-of-process plugin SDK for teams who need to add custom static analysis rules or proprietary framework detection without modifying the core engine.

Repomix is a highly optimized repository packer designed to concatenate code into an AI-friendly prompt. Fuzit is a persistent, incremental intelligence engine designed to select *subsets* of code based on a typed graph and task relevance. We respect both approaches. Read our detailed architectural comparison.

Fuzit benchmarks focus on reproducible workloads like incremental update speed and context selection determinism. See our Benchmarks registry for the measurement architecture.

Yes. Fuzit is open source software released under the permissive MIT License. The full source code and official license terms are available on GitHub.