TerseSharp 0.84.1

dotnet tool install --global TerseSharp --version 0.84.1
                    
This package contains a .NET tool you can call from the shell/command line.
dotnet new tool-manifest
                    
if you are setting up this repo
dotnet tool install --local TerseSharp --version 0.84.1
                    
This package contains a .NET tool you can call from the shell/command line.
#tool dotnet:?package=TerseSharp&version=0.84.1
                    
nuke :add-package TerseSharp --version 0.84.1
                    

TerseSharp

Roslyn for your coding agent — 86 tools over MCP. Fewer tokens. Fewer turns of the loop. Only code you would merge.

Coding agents read C# the hard way. They open whole files, grep for names, and guess at line numbers. TerseSharp gives them the compiler's view instead. It loads your solution into Roslyn once and answers by meaning, not by text.

Without it: grep Submit → 40 hits → read a 2,000-line file → read three callers → edit by line number. With it: find_usages → the 6 real references, ~200 tokens, one call.

💸 Save tokens ⚡ Go faster 🧪 Ship better code
4.56M tokens not spent in one measured week, 2.9× the whole bill every round trip it removes saves 6.1 s of model wait every edit is compile-gated and checked against your code health bar

Install

dotnet tool install -g TerseSharp
terse install --skill --guard

Restart your agent and ask it about your code. TerseSharp finds your .sln, .slnx or .csproj on its own. It works with Claude Code, Cursor, VS Code, Windsurf and any MCP client over stdio. No IDE, no licence, no API key. It runs on the .NET 11 runtime; a solution pinned to the .NET 10 SDK still loads. terse doctor checks the setup.

💸 Save tokens

We replayed 508 real Claude Code sessions on a 31,000-file solution. Each of the 6,045 TerseSharp calls was re-priced against the built-in tool it replaced, run for real.

Tokens
What TerseSharp actually cost 2.41M
What Read, Grep and the shell would have cost 6.97M
Saved 4.56M (2.9× the bill)
You ask Built-in tools TerseSharp Saving
What is on this 2,000-line type? read it all, ~6,000 tokens get_type_outline, ~450 13×
Who calls this method? grep and follow-ups, ~4,000 find_usages, ~200 20×
Rename it everywhere ~5,000, and it misses the interface rename_symbol, ~150, correct 30×
Why is the build red? ~8,000 tokens of MSBuild log build, ~600 13×
What did I just change? the whole patch diff_symbols, the changed declarations 10×

With every assumption pushed against TerseSharp, the saving is still 2.09M tokens, 1.9×. A successful edit answers in one line; failures, rollbacks and warnings always print in full.

⚡ Go faster

Measured over 36,075 real tool calls, every turn of the agent's loop costs 6.1 seconds before the tool even runs. The fastest call is the one the agent never needs to make.

  • One call instead of a chain. get_symbol usages=true answers what a symbol is, who uses it and what implements it.
  • Batches everywhere. Read 10 files, apply 25 edits, or search 10 terms in one call.
  • Edits that land together. Change a signature and its callers in one compile-gated step.
  • Fast tests. run_tests builds once and runs test projects in parallel, 38% faster. A repeat with nothing changed answers instantly.
  • Parallel-safe reads. Read tools are marked read-only, so clients run them side by side.

🧪 Ship better code

Every edit is compile-gated. An edit that adds a compile error is rolled back, in the changed project and in every project that depends on it. Results are tagged EXACT or HEURISTIC, and the server never guesses.

Code health, scored on every edit. Every edited file gets a score from 1 to 10, in the style of CodeScene's Code Health. A file starts at 10 and each smell takes points off:

Smell Triggers when a method… Penalty
Brain Method is complex and large (on top of the two below, so 2.5 in all) 1.5
Deep Nested Complexity nests 4 or more levels of if, loops or switch 0.75
Complex Method has a cyclomatic complexity of 10 or more 0.5
Large Method has more than 70 lines 0.5
Bumpy Road has 2 or more separate blocks of nested logic 0.5
Excess Function Arguments takes more than 4 parameters 0.25
Complex Conditional joins 3 or more conditions with && / \|\| 0.25
Code Duplication repeats another method of the file with only names and literals changed 0.5
Low Cohesion (a type) splits into 3 or more groups of members that share no state and never call each other 1.0
Large File (the file) is over 1,000 lines 1.0

The default bar is 9. An edit that creates a file below it, takes a file below it, or makes an unhealthy file worse, is reported as TERSE114, naming each smell and the method that carries it. analyze alone adds developer congestion and change coupling from the last 200 commits as HEURISTIC TERSE114 lines; an edit never runs git. Like every policy rule it warns by default. Make it a hard stop, move the bar or switch it off in .terse.json:

{ "policy": { "rules": { "codeHealth": { "limit": 9, "action": "reject" } } } }

Fifteen policy rules, your standards. A policy section in .terse.json checks every edit: cognitive complexity (TERSE100), statements per method, methods per type, constructor dependencies, parameters, method name length, meaningless suffixes, naming, async void, condition operands, chained references, nesting depth (default 3), comments (TERSE112), XML doc comments (TERSE113) and code health (TERSE114). Each is reject, warn or off, every rule that is on by default warns, and only what an edit introduces counts. The file cascades like .editorconfig, and terse install writes your home copy.

One call for the whole quality gate. gate runs analyze → format → cleanup → analyze on the files you changed and answers one line. cleanup verify=true fix=ci checks both CI format commands in one call.

🛡️ Stays on the rails

  • The guard. terse install --guard blocks Read, Grep, cat and dotnet build on C#, and hands the agent the exact TerseSharp call to make instead, with the ToolSearch select: that loads its schema.
  • Read-only mode. terse serve --read-only makes every write refuse.
  • A smaller surface. A solution with no XAML, Razor or .resx advertises 55 tools, ≤25,700 tokens. The full list stays under a 31,100-token ceiling, checked on every push.

🧩 Your stack

What you get
Blazor / Razor razor_validate catches a parameter that compiles but throws at render.
XAML — MAUI, WPF, WinUI, Avalonia xaml_bindings validate=true checks every {Binding} path, which WPF never does.
ASP.NET Core list_endpoints lists every route; find_registrations finds DI registrations.
Localization .resx / .resw families, missing translations, placeholder mismatches.

🧰 All 86 tools

Group Tools
Workspace load_workspace · workspace_status · list_workspaces · unload_workspace · list_projects
Navigate search_symbols · get_symbol · get_file_outline · get_type_outline · get_symbol_source · find_usages · find_implementations
Beyond grep find_registrations · list_endpoints
Analyze & clean analyze · format · cleanup · gate · clean · get_diagnostics
Edit replace_symbol_body · replace_symbol · add_member · delete_symbol · rename_symbol
Refactor extract_interface · move_type_to_file · move_type_to_namespace · change_signature · undo_last_change
Projects & packages solution_projects · solution_add_project · solution_remove_project · project_create · project_properties · project_set_property · project_add_reference · project_remove_reference · package_list · package_add · package_remove
XAML xaml_outline · xaml_names · xaml_resources · xaml_resolve · xaml_styles · xaml_bindings · xaml_validate · xaml_find · xaml_codebehind · xaml_localization · xaml_set_property · xaml_add_element · xaml_remove_element
Razor / Blazor razor_outline · razor_component · razor_find · razor_bindings · razor_codebehind · razor_validate · razor_set_attribute · razor_add_element · razor_remove_element · razor_set_directive
Localization resx_files · resx_get · resx_find · resx_usages · resx_set · resx_remove · resx_rename · resx_validate
Files read_text · write_text · edit_text · find_files · search_text · search_regex
Git changed_files · diff_symbols · diff_text · history
Build & test build · run_tests · rerun_failed · list_tests

MIT licensed. Built on Roslyn and the MCP C# SDK.

There are no supported framework assets in this package.

Learn more about Target Frameworks and .NET Standard.

This package has no dependencies.

Version Downloads Last Updated
0.84.1 83 11/13/2026
0.84.0 69 11/13/2026
0.83.0 118 11/8/2026
0.82.1 103 11/7/2026
0.81.0 133 10/31/2026
0.80.0 146 10/23/2026
0.79.0 148 10/17/2026
0.78.0 144 10/17/2026
0.77.0 153 10/15/2026
0.76.0 151 10/10/2026
0.75.1 156 10/7/2026
0.74.0 147 10/7/2026
0.73.0 163 10/2/2026
0.72.0 116 10/1/2026
0.71.1 126 10/1/2026
0.71.0 121 9/29/2026
0.70.0 119 9/26/2026
0.69.1 124 9/26/2026
0.68.0 127 9/23/2026
0.67.0 121 9/21/2026
Loading failed