ContextManager

[ContextManager logo]

ContextManager

Structural C# context for AI coding agents, powered by Roslyn.
Navigate before you read.

[NuGet version]

Install • Tools • How to use it • FAQ • Full install guide

ContextManager is an MCP server that saves an agent from reading whole C# files. It does two things:

• Inspects files. It turns a .cs file into a JSON contract with its types, public members, constructor dependencies, base types, interfaces, and events. Method bodies, private members, and XML docs are left out; each method carries its line range so the agent can read only the body it needs.
• Maps the solution. It builds a knowledge graph of a .sln so the agent can find which files matter, and how risky a change is, before opening any of them.

Install

You need the .NET 10 SDK or later.

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/nushey/ContextManager/main/scripts/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/nushey/ContextManager/main/scripts/install.ps1 | iex

The installer installs or updates the ContextManager .NET global tool and runs context-manager setup, a wizard that adds the server to the MCP clients you pick (Claude Code, Claude Desktop, Codex, Antigravity, Kilo Code, Cursor, Devin, Opencode).

To install by hand instead:

dotnet tool install -g ContextManagerclaude mcp add --scope user context-manager -- context-manager   # Claude Code

For other clients, flags, config-file paths, and troubleshooting, see INSTALL.md.

project_scan on .NET Framework 4.8 solutions only works on Windows and needs Visual Studio Build Tools with the .NET desktop build tools workload. Everything else works on any platform.

Tools

Graph tools take the path of the .sln and keep the graph up to date on their own: before answering, they re-extract only what changed since the last scan. The graph is saved in .context-manager/<solution name>.graph.json next to the .sln, so every session on that solution shares it.

How to use it

Start from a file

When the user mentions a file, start there:

1. inspect_file(path): read the type and build its node ID as Namespace.TypeName.
2. graph_get_dependencies(solutionPath, nodeId): find the neighbors worth reading.
3. inspect_file on those neighbors only.
4. graph_impact_analysis(solutionPath, nodeId): check the blast radius before changing anything.

A short impact list means the change is contained. A long one means the type's public contract is load-bearing and should be preserved; it does not mean the agent has to read every file in it.

Node IDs use Roslyn's fully qualified format, for example MyApp.Orders.OrderService or MyApp.Orders.OrderService.GetOrderAsync(System.Guid). Pass the IDs that the graph tools return as they are.

// graph_impact_analysis("/abs/path/MyApp.sln", "MyApp.Orders.IOrderRepository"){  "affectedIds": ["MyApp.Orders.OrderService", "MyApp.Api.OrdersController"],  "affectedFiles": {    "MyApp.Orders.OrderService": "/abs/path/Orders/OrderService.cs",    "MyApp.Api.OrdersController": "/abs/path/Api/OrdersController.cs"  }}

Find code by name

When you know roughly what something is called but not which file it is in, and you have no node ID, call graph_search with several candidate names:

{"solutionPath":"/abs/path/MyApp.sln","queries":["Login","Authentication","SignIn"]}

What it searches. A term matches when it appears anywhere in a file name (without .cs) or a method name (without its signature). Matching is ordinal and ignores case, and any one term is enough. It searches every method already in the graph, private ones included, and finds constructors under their type name. It does not search directory paths, parameter types, method bodies, or namespaces.

Valid terms. Terms are trimmed and must have at least 3 characters. An empty list, or a single term that is too short, rejects the whole query before the graph is synchronized, with code invalid_query and message Provide at least one query term; each term must be at least 3 characters long after trimming whitespace.

Order and limit. Exact matches come first, then partial ones; within each group, results keep the order in which they appear in the graph. A name that matches several terms does not rank higher. The search keeps at most 50 matches, and the limit counts matches, not files: each matching file counts once and each matching method ID counts once, even when several terms match it. There is no pagination. When truncated is true, at least one match was left out; use more specific names to narrow the search.

{"results":[{"file":"/abs/path/Auth/Session.cs","namespaces":["MyApp.Auth"],"fileMatched":false,"methods":[{"name":"Login","nodeId":"MyApp.Auth.Session.Login()"}]}],"truncated":false}

Response. The response is one line of compact JSON, grouped by file. Every file entry has file, namespaces, fileMatched, and methods, and empty arrays stay as []. namespaces lists every namespace in the file. fileMatched is true when the file's own name is one of the matches kept within the limit. A file can also appear only because one of its methods was kept, even if its name matched too but fell outside the limit. With no matches, the response is {"results":[],"truncated":false}.

From there, call inspect_file on a returned file, or pass a method's nodeId unchanged to the other graph tools. A graph saved by an older version is re-extracted once to add names and namespaces; later queries reuse the updated graph while nothing changes.

To make your agent follow these workflows, copy docs/AGENTS-template.md into the AGENTS.md of your project.

FAQ

Is it a background service? No. It is an MCP stdio server: the client starts it when a session opens and stops it when the session ends.

What state does it keep? Only the graph files written next to each .sln. The inspection tools read the files you pass and keep nothing.

Does it read my whole solution? Only the graph tools do. inspect_file and inspect_context read just the files you give them.

Build and test

dotnet builddotnet test

License

MIT © nushey