TokenSaver

A structural warm start for AI coding agents in .NET. Instead of loading whole files into your assistant, TokenSaver hands it a cheap map of your code — every type and member as a signature, each tagged with its line range — so the model reads only the slice it needs instead of slurping the file. Built on the Roslyn compiler platform.

Where it pays off: outlining a file costs 70–95% fewer tokens than reading it — up to 90% on a large file. The end-to-end win is biggest on smaller / cheaper models (which over-read the most) and on large codebases: on real tasks it trims a Haiku-class model's token use by ~8%, and the savings climb with file size. A top-tier model already reads tightly, so it sees less benefit — the leaner the model, the more a warm start helps.

Works with Visual Studio 2026 (GitHub Copilot Chat), Claude Code, VS Code Copilot, Claude Desktop, and any other MCP client that speaks stdio.

→ Full docs and setup guide: mcp/README.md
→ Changelog: CHANGELOG.md

---

Install

See tokensavermcp.com/install for one-click install buttons, per-client config snippets, upgrade/uninstall instructions, and troubleshooting.

---

What the tools do

<!-- BEGIN:generated:tools -->

Single-file tools

Tool	What it does	Reduction
OutlineCSharpFile	Signatures of every type and member — no bodies. Best for "what's in this file?"	70–95%
MinifyFile	Lossless minify of an entire file — strips comments and whitespace, logic unchanged. Auto-dispatches by extension (C#, Razor, JS/TS, Python, HTML, CSS, JSON, YAML, XML, C, C++, X++, VB.NET).	20–50%

* Reductions are measured against reading the whole file — the real alternative when you'd otherwise load it. The end-to-end saving on a task is smaller, because a capable model already reads somewhat selectively; it is largest on smaller/cheaper models and large files (see Token savings in practice below).

Cross-file traversal tools

These scan an entire project directory in one call — no need to know which file to look in first.

Tool	What it does
TraceDiRegistrations	Finds every Dependency-Injection registration referencing a type (interface or concrete) and returns a compact table of file:line, method, ServiceType -> ImplType, and key. Answers "where is IFoo wired, and to what?"

Both accept a directory path or .csproj file — obj/ and bin/ are excluded automatically. <!-- END:generated:tools -->

---

Token savings in practice

Measured against this project's own FocusedEmitter.cs (9,261 tokens raw):

Question type	Tool	Tokens sent	Reduction
"What's in this file?"	OutlineCSharpFile	1,039	89%
"Read one method body"	OutlineCSharpFile + narrow Read of the // L.. range	~300	~95%
"Audit the whole file"	MinifyFile	5,525	40%

Those are per-read figures versus loading the whole file. The end-to-end saving on a real task is smaller, because a capable model already reads somewhat selectively — and that's the honest part of the story:

Model running the agent	Real-task token saving vs no tool
Top-tier (reads tightly already)	~1% — negligible
Mid-tier (Sonnet-class)	~5–7%
Smaller / cheaper (Haiku-class)	~8%

The leaner the model, the more it over-reads on its own — and the more a warm start saves. TokenSaver is most worth it for agents running on small/cheap models, large files, or any workflow that would otherwise read whole files. It won't make a top model meaningfully cheaper, and it won't make a model write better code — it curbs wasteful reading, which is where cheap-model token budgets actually leak.

---

Collective impact

Every invocation is counted at tokensavermcp.com — a live dashboard showing total tokens saved by the community. Fewer tokens processed means less GPU compute and a smaller carbon footprint for AI-assisted development.

---

Language support

Tier	Languages
Primary (Roslyn, full support)	C# .cs, Razor .razor, VB.NET .vb, .NET project files .csproj .props .config .xml
Basic (comment-strip + whitespace collapse)	JavaScript, TypeScript, Python, HTML, CSS/SCSS/LESS, JSON/JSONC, YAML, C, C++, X++

OutlineCSharpFile and TraceDiRegistrations are C# only; MinifyFile works across every supported language.