Fuzit
Features

Repository intelligence
built for context engineering.

Fuzit maintains structured repository knowledge, keeps it fresh, and turns it into bounded, explainable context for development tasks.
$npm install -g @fuzit/cli
Docs

Repository Discovery

Fuzit identifies repository roots, workspaces, manifests, and Git state. It applies strict ignore behaviors and sensitive-path boundary filtering to determine the exact set of allowed repository evidence before analysis begins.
Repository
+ Detect Workspaces / Manifests
- Apply .gitignore & .fuzitignore
- Drop sensitive paths (keys, .env)
Canonical Source Set

Canonical Incremental Index

Repository understanding does not need to start from zero every invocation. Fuzit maintains a persistent local atomic index. When files change, it updates only the semantically affected state, ensuring the index stays fresh without rebuilding everything.
State Invalidation Model
Changed File
src/auth/session.ts
Behavior
Invalidates bounded dependents only
Unchanged
98% of repository remains cached
Guarantee
Incremental result mathematically identical to full rebuild

Zero-Execution Static Analysis

Source code is analyzed to extract normalized symbols, imports, exports, references, and frameworks without ever executing repository code. Fuzit does not run build scripts, compiler plugins, or arbitrary project hooks.
Normalized Evidence Extraction
TypeScript / JS
ESM/CommonJS, React, Next.js, Exports
Python
Modules, Packages, FastAPI, Flask, Django
Java
Classes, Maven/Gradle metadata, Spring Boot
Go
go.mod, Packages, Replacements, Tests

Typed Context Graph

Code is a system, not a folder. Fuzit constructs typed nodes and relationships representing your architecture. The graph connects files to symbols, symbols to tests, and routes to handlers, preventing related context from being missed simply because it lives in a different directory.
Example Traversal Path
app/api/route.ts (Route)
handleRequest()
src/services/Auth.ts (Class)
verifySession()
src/services/auth.test.ts (Related Test)

Task-Aware Selection

Selection changes according to the developer's actual task. Fuzit combines lexical relevance, exact symbols, graph distance, tests, Git recency, and lifecycle signals to rank the best candidates rather than producing the same flat repository dump for every question.
CandidateSignalsDecision
session.tsExact symbol + Task relevanceSelected
session.test.tsGraph distance + Related testSelected
avatar.tsWeak task relationOmitted

Bounded Context

Fuzit does not try to fill the entire context window just because it can. It produces bounded context strictly according to configured token/byte limits and selection rules, maximizing the signal-to-noise ratio for the downstream agent.
BUDGET ENFORCEMENT
Candidate Evidence Pool140,000 tokens
↓ Budget Boundary ↓
Selected Context Delivered~12,000 tokens

Fully Explainable Selection

Retrieval is not magic. Fuzit exposes deterministic reasons for why evidence was included, omitted, expanded, truncated, or redacted. You can inspect ranking components, graph paths, and provenance directly.
Selection Audit
Target File
src/auth/token.ts
Status
Included
Reason: Provenance
Local repository index
Reason: Graph Path
Imported by session.ts (Depth 1)
Reason: Ranking
Task lexical match + Explicit Anchor

Security Filtering

Context engineering requires security before output. Fuzit filters sensitive content, keys, and environments by policy before they can gain relevance to a task. It operates local-first, preventing accidental source code leaks to third-party endpoints.
Blocked Candidatesrc/config/.env.production
Reason: Sensitive path policy matched
Explore the security model

A pull request is more than its diff.

GitHub is a read-only explicit remote context source. Remote repositories, PRs, and issues can enter the exact same Fuzit repository-intelligence pipeline used for local sources.

PR & Issue Intelligence

Fuzit acquires exact base/head state, bounded patches, PR reviews, comment threads, checks, and statuses to build a complete review bundle. Issues can also act as the task specification used to select relevant repository evidence.
$ fuzit review https://github.com/OWNER/REPO/pull/123
✔ Acquired PR metadata and patches
✔ Resolved base repository graph
✔ Processed 14 review comments
→ Generated bounded review context
Supported Evidence
PR State
Exact base & head revisions
Discussions
Reviews & thread comments
Automation
Checks & statuses
Context
Issue bodies as task intent

Framework & Ecosystem Intelligence

Fuzit performs zero-execution static analysis. We use authoritative evidence-based detection rather than depending on dependency presence alone to declare framework support.

LanguageEcosystem Signals & Extraction
TypeScript / JSnpm/pnpm/Yarn workspaces, package exports, React, Next.js (App/Pages), Express, Fastify, NestJS, Vitest, Jest.
PythonModules, packages, FastAPI, Flask, Django, test framework evidence without executing repository modules.
JavaMaven/Gradle metadata, classes, interfaces, Spring Boot, JUnit. Extracted without invoking build tools.
Gogo.mod, replacements, packages, functions, tests, HTTP registrations, build constraints.

One Engine, Five Surfaces

A major strategic advantage of Fuzit is infrastructure reuse. The canonical intelligence model serves multiple consumption surfaces without divergent implementations.

SurfacePurposeState ModelTrust Characteristics
CLIDirect user-facing control surface for packing, generation, inspection, and verification.Stateless (reads index)Direct execution
WatcherFilesystem change observation keeping canonical repository state fresh incrementally.Stateful daemonDirect execution
Stdio MCPAllows Claude, Cursor, and other clients to query Fuzit via standard Model Context Protocol.Stateless serverRead-only, explicit roots
VS CodeNative editor integration consuming the canonical engine for IDE workflows.Managed lifecycleRespects Workspace Trust
Plugin HostControlled extensibility for custom parsers, retrievers, or policies.Out-of-processDeny-by-default bounds

The Canonical Pipeline

Every request passes through the same deterministic, 10-stage processing pipeline.

01
DiscoverIdentify roots, workspaces, configurations, Git state, and security boundaries.
02
ScanStream files through ignore rules, path policies, and content hashing.
03
IndexStore canonical file records and versioned identities in an atomic incremental index.
04
AnalyzeExtract symbols, imports, tests, and schemas without executing repository code.
05
GraphConstruct typed nodes and edges linking dependencies and structural paths.
06
EnrichAdd framework intelligence, lifecycle facts, Git evidence, and detector provenance.
07
RankScore context with graph distance, relevance, budgets, and explicit anchors.
08
ExplainPreserve score components, graph paths, and omission/security decisions.
09
PackageProduce deterministic context in structured formats (Markdown, JSON, XML).
10
ServeExpose identical services through CLI, watcher, MCP, VS Code, and plugins.

Put repository intelligence to work.

$npm install -g @fuzit/cli