Fuse

The MCP server that collapses an AI coding agent's explore phase on .NET into one scoped call.

Website . Documentation . Connect your agent . Benchmarks

Fuse is a Model Context Protocol server for AI-assisted development on .NET. It hands your coding agent (Claude Code, Cursor, GitHub Copilot) the right code, scoped and reduced, in a single call, instead of letting it burn its context window opening thousands of files during the explore phase. It cuts input tokens while keeping the public API intact, scopes to the files a task needs, and trims the round-trips an agent makes. The same engine is also a fuse CLI.

Unlike generic repo packers (Repomix, Code2Prompt, Gitingest), Fuse understands C# structure: dependency graphs, skeleton extraction, BM25 query scoping, git change detection, and convention patterns. An opt-in Roslyn precision tier and a hybrid-retrieval reranker raise accuracy further. The MCP server exposes nine tools; see Connect your agent.

Full documentation lives at fuse.codes.

Why Fuse

Measured over a pinned corpus of four real .NET libraries (MediatR, FluentValidation, AutoMapper, Newtonsoft.Json), counted with the o200k_base tokenizer. Reduction ratios transfer across models even though absolute token counts do not. Every figure is reproducible with one command and reported in full, including the arms where Fuse ties or loses, on the benchmarks page.

• One scoped call instead of an explore loop. Over 24 real merged pull requests, one fuse --query call delivers the files a change needs in about 40,000 tokens at 51 percent recall, against a generic-packer dump of about 512,000 tokens (about 13 times more) at the same single call, and about 494,000 tokens to read the repository blind. Fuse ties a packer at one call and wins on tokens; with a git base the change-scoping mode reaches 88 percent recall. Read the token number with its recall.
• Cuts tokens without dropping API. Default reduction removes 7 to 10 percent and --all removes 21 to 40 percent of tokens while keeping 99 to 100 percent of public types and methods. --skeleton removes 66 to 93 percent for an architecture map.
• Smaller than the generic packers. Repomix output runs 1.3 to 3.9 percent larger than raw concatenation on these repositories; Fuse is smaller than raw in every mode.
• Finds the files a change touches. Change scoping recalls 88 percent of the files in 24 real merged pull requests at 61 percent precision, and all three scoping modes beat an agent-style grep baseline.
• Trustworthy skeletons on hard code. The opt-in Roslyn tier keeps 100 percent of method signatures on all four libraries, including Newtonsoft.Json, where the regex skeleton kept 4 percent.
• Cheap repeated calls. The on-disk analysis index roughly halves warm-call wall-clock across a session, so a multi-call task pays the analysis cost once.
• Native AOT and no runtime reflection on the default path. The fast path ships as an ahead-of-time-compiled binary; Roslyn and the vector reranker are opt-in tiers isolated from it.

Reproduce every number with pwsh -File tests/benchmarks/harness/run-all.ps1.

Install

Fuse is a developer tool for .NET developers, so the recommended install is the .NET global tool (.NET SDK 10.0 or later):

dotnet tool install -g Fuse

Or run it on demand without installing: dnx Fuse -- serve.

If you do not have the .NET SDK, install a self-contained binary (Native AOT, no runtime required) with the script for your platform:

# Linux
curl -fsSL https://fuse.codes/install.sh | sh

# Windows
irm https://fuse.codes/install.ps1 | iex

On Windows you can also use winget install Litenova.Fuse, or download a binary from Releases. Verify with fuse --help. Full notes: fuse.codes/docs/start/install.

Connect your agent

Register Fuse once; your AI client launches fuse mcp serve automatically when MCP is enabled:

fuse mcp install --rules

That writes MCP config for Claude Code, Cursor, and GitHub Copilot in the current project, and --rules (recommended) writes a short rule into each client's instruction file (CLAUDE.md, .cursor/rules/fuse.mdc, .github/copilot-instructions.md) so the agent reaches for the fuse_* tools instead of grepping blindly. Use fuse mcp install --scope user to register for every project on this machine, or --client cursor to configure one client only.

Manual registration is also supported. For Claude Code, add .mcp.json to your project root:

{
  "mcpServers": {
    "fuse": {
      "type": "stdio",
      "command": "fuse",
      "args": ["mcp", "serve"]
    }
  }
}

Or register with the Claude CLI: claude mcp add fuse --scope project -- fuse mcp serve (use --scope user for all projects). Cursor uses .cursor/mcp.json and GitHub Copilot uses .vscode/mcp.json; see Connect to your AI for both.

A recommended agent flow on a large codebase: survey with fuse_toc or fuse_skeleton, drill in with fuse_focus or fuse_search, then review a branch with fuse_changes. Or call fuse_ask with a task and a token budget and let Fuse pick the strategy. See Context for an agent.

Tool catalog and parameters: MCP Tools and MCP Resources. MCP Registry manifest: mcp-registry/server.json.

Command-line quickstart

The same engine runs as a CLI:

# fuse a .NET project with the DotNet template
fuse dotnet --directory ./src

# maximum C# reduction, public API intact
fuse dotnet --directory ./src --all

# architecture overview, signatures only
fuse dotnet --directory ./src --all --skeleton

# cheap survey before fetching files (tree, symbol outline, token costs)
fuse dotnet --directory ./src --toc

# accurate skeletons and dependency edges with the Roslyn precision tier
fuse dotnet --directory ./src --skeleton --semantic

# PR-scoped fusion with diff hunks and the callers of each changed file
fuse dotnet --directory ./src --changed-since main --review

# query-scoped fusion
fuse dotnet --directory ./src --query "payment gateway" --query-top 10

Output defaults to Documents/Fuse; use --output and --name to control the destination. Walkthrough: fuse.codes/docs/start/quickstart.

Commands

Command	Purpose
fuse	Generic fusion. All extensions unless you set --only-extensions.
fuse dotnet	.NET projects: C# reduction, structural maps, dependency-aware scoping.
fuse wiki	Azure DevOps wikis: Markdown only.
fuse init	Create fuse.json in the current directory.
fuse mcp install	Register Fuse with MCP clients (Claude Code, Cursor, Copilot).
fuse mcp serve	MCP server entry point on stdio (launched by your client, not run manually).

Full option lists: Commands and Options.

How it works

A fusion is a four-stage pipeline:

• Collection - scan the source directory and apply filters (extensions, .gitignore, binary detection, test projects, globs).
• Filtering - optional scoping: focus, git changes, or BM25 query with dependency expansion.
• Reduction - normalize whitespace, run language and format reducers, apply skeleton, markers, and secret redaction.
• Emission - count tokens, build the manifest, apply the output format, and write within a token budget.

The concept in plain terms is at How Fuse works; the internals are at The pipeline.

Repository layout

src/
  Core/                               Pipeline libraries
    Fuse.Collection/                  File discovery, filters, templates
    Fuse.Reduction/                   Content pipeline, caching, redaction
    Fuse.Emission/                    Output writers, token budget, manifest
    Fuse.Fusion/                      Orchestration, scoping, analysis, enrichment, DI
  Host/
    Fuse.Cli/                         CLI and MCP server
  Plugins/                            Extension-keyed capability providers
    Fuse.Plugins.Abstractions/        Capability interfaces (shared contract)
    Fuse.Plugins.Languages.CSharp/    C# language plugin (regex, AOT-clean default)
    Fuse.Plugins.Languages.CSharp.Roslyn/  Opt-in Roslyn precision tier (excluded from the AOT build)
    Fuse.Plugins.Formats.Web/         Format reducers (HTML, JSON, YAML, SQL, TS/JS, etc.)
tests/                                Unit, golden-output, and integration tests; benchmarks
site/                                 The fuse.codes website and documentation (Next.js + Fumadocs)
assets/                               Benchmark figure and the chart-generating script
mcp-registry/                         MCP Registry server manifest

Development

dotnet build Fuse.slnx --configuration Release
dotnet test Fuse.slnx --configuration Release --no-build
dotnet format Fuse.slnx --verify-no-changes

Contribution workflow: Contributing. Agent instructions: AGENTS.md. The documentation site is in site/; see site/README.md to run it locally.

License

MIT. Copyright (c) 2026 Litenova Solutions. See LICENSE.