TerseSharp 0.2.2

TerseSharp

Your agent stops reading whole C# files.

A Roslyn-powered MCP server that lets a coding agent navigate, read, edit and refactor a .NET solution semantically — no Read, no Grep, no line-number Edit, no shelling out.

v0.2.2 — 51 tools working end to end. Verified by 117 tests (29 unit + 88 E2E), where every E2E test drives a real server process over the real stdio transport against a real solution and asserts response values. Not yet built: the content-addressed index, trigram search and file watcher.

Why

An agent working a C# solution spends most of its context on the wrong shape of data:

Roslyn already knows all four answers semantically. TerseSharp hands them over in the shape the agent needs: a signature list instead of a file, real call sites instead of string matches, a solution-wide rename instead of a regex sweep, deduplicated diagnostics instead of build logs.

Install

No IDE, no licence, no Node, no Python, no language server, no API key, no network.

dotnet tool install -g TerseSharp

Register it with your agent — TerseSharp writes the config itself, you don't hand-edit JSON:

terse install                       # detect installed clients and register with all of them
terse install --client claude-code  # or pick one: claude-code | cursor | vscode | windsurf
terse install --skill               # also install the agent skill
terse doctor                        # verify SDK, MSBuild, workspace load, client registration

With no arguments the server walks up from the current directory, finds your .sln / .slnx / .slnf / .csproj, and loads it.

Prefer to configure it by hand:

{
  "mcpServers": {
    "terse-sharp": {
      "command": "terse",
      "args": ["serve", "--workspace", "C:/path/to/YourApp.slnx"]
    }
  }
}

What each tool replaces

The 51 tools

• Workspace — load_workspace, workspace_status, list_workspaces, unload_workspace, list_projects
• Navigation — search_symbols, get_symbol, get_file_outline, get_type_outline, get_symbol_source, find_usages, find_implementations
• Analyze & clean — analyze, format, cleanup, 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 & solutions — 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_bindings, xaml_validate, xaml_find
• Files — read_text, write_text, edit_text, find_files, search_text, search_regex
• Build — build, run_tests

Analysis without a licence

analyze runs the compiler plus every analyzer your projects already reference - CA rules, StyleCop, SonarAnalyzer, Roslynator, anything in your PackageReference list - down to info and hidden severity, which a normal build hides. It also reports dead code in the same list - unreferenced private members as TERSE001, plus the compiler's unused-field and unreachable-code hints - so one call covers everything. cleanup removes unused using directives, sorts what remains System-first and reformats to your .editorconfig. All Roslyn: no IDE, no external tool, no licence, no network.

Every response is one record per line, with an explicit truncated/total and an EXACT (Roslyn-resolved) or HEURISTIC (text/index) tag.

Safety

• Symbol-addressed edits — no old_string echo, no line numbers to drift.
• dryRun on every mutation returns the unified diff and writes nothing.
• Compile-gated — an edit that introduces a new compile error is rolled back and the error returned. Pre-existing errors never block an edit. allowErrors: true opts out.
• Diff-only responses — mutations return the diff and a changed-line count, never the file.
• Workspace containment — paths compare by whole segment, so root C:\repo does not contain C:\repoEvil.
• --read-only makes every mutating tool refuse and touch nothing.

Parallel worktrees

Run several agents at once across several git worktrees of one repo, and across unrelated repos — one server holding many workspaces (LRU, default 4), or many processes, or both. Every answer names its worktree and branch, and an ambiguous request returns AMBIGUOUS_WORKSPACE listing the candidates instead of guessing — answering from the wrong checkout is the one failure an agent cannot detect.

Links

• Source, full documentation and issues
• Changelog
• Contributing
• Security policy

License

MIT Licensed. See LICENSE.

Built on Roslyn and the MCP C# SDK.