Give your agent the right code, not the whole repo.
Fuse collects a .NET codebase, reduces it for token efficiency, and emits one structured payload an agent reads in a single call. Fewer tokens, the files a task actually needs, and the public API kept intact.
dotnet tool install -g Fuse
Measured: Newtonsoft.Json, --all mode. See the benchmarks.
Measured over a commit-pinned OSS corpus, counted with o200k_base. Reproduce the benchmarks.
Agents waste their context window finding the code
Before an AI coding tool changes a line, it explores: lists directories, greps, opens file after file to learn which ones matter. On a solution with hundreds of C# files that burns most of the context window and many slow round-trips on discovery, not on the task. Generic packers fix this by dumping the whole repo as text, which is rarely smaller and loses the structure. Fuse takes a different path.
Why Fuse, and how it comparesFewer tokens
Structural C# reduction removes comments, usings, namespaces, and whitespace, and can drop bodies to signatures. The default and --all keep the public API while cutting 7-40% of tokens.
Finds the right files
Scope a fusion to a type and its dependencies, the files a git diff touched, or the files a query ranks highest. Fuse expands through a dependency graph instead of dumping everything.
Keeps the API intact
Reduction is not deletion. An independent Roslyn oracle confirms default and --all keep 99-100% of public types and methods. fuse verify reports the preserved surface.
MCP-native
fuse serve is a Model Context Protocol server with eight tools for Claude Code, Cursor, and Copilot, so an agent fetches scoped context directly instead of reading files one by one.
Built for review
Scope to a branch with change recall of 88% on real merged PRs, prepend a review map of diff hunks and callers, and emit only what changed since the last turn.
Deterministic and AOT-clean
The default path uses no runtime reflection and ships as a Native AOT binary. Same input, same output. An opt-in Roslyn tier raises precision when you ask for it.
Reduction without losing the surface
Token reduction and public-API fidelity across the benchmark corpus. Every number comes from a harness anyone can rerun against the same pinned commits.
From install to output in 30 seconds
Install the global tool, point it at a .NET source tree, and read one file instead of thousands. The output opens with a manifest that lists every included file and its token cost.
Full quickstart# install the .NET global tool dotnet tool install -g Fuse # fuse a project at full API fidelity fuse dotnet --directory ./src --all
Fused 511 files Estimated tokens: 366,121 (-21.5%) cache: 0 hit / 511 miss Output: AutoMapper_2026-06-20_366k.txt
// .mcp.json
{
"mcpServers": {
"fuse": {
"type": "stdio",
"command": "fuse",
"args": ["serve"]
}
}
}# or register it with Claude Code in one line claude mcp add fuse --scope project -- fuse serve
Connect it to your AI in one line
Run fuse serve and your agent gets eight tools: survey a codebase, drill into a type, scope to a query or a branch, or ask one question and let Fuse pick the strategy. It works with Claude Code, Cursor, and GitHub Copilot.
Not a packer. Not an embedding index.
Generic packers concatenate files as text. RAG indexers return fuzzy chunks. Fuse understands C# structure and preserves it.
| Capability | Generic packers | RAG / embeddings | Fuse |
|---|---|---|---|
| Understands C# structure | No, plain text | No, opaque chunks | Yes, types and signatures |
| Cuts tokens | Often larger than raw | Per chunk only | 7-40% at full API fidelity |
| Keeps whole API surface | Only if you include it all | No, partial recall | 99-100% verified |
| Deterministic output | Yes | No, similarity-ranked | Yes |
| MCP server for agents | Rarely | Varies | Eight tools, built in |
Stop paying for the explore phase.
Install Fuse, connect it to your agent, and hand it scoped context that fits the window.