CodeMap — Turn Your AI Agent Into a Semantic Dragon

Stop feeding your AI agent raw source files. Give it a semantic index instead.

CodeMap is a Roslyn-powered MCP server that lets AI agents navigate C#, VB.NET, and F# codebases by symbol, call graph, and architectural fact — instead of brute-reading thousands of lines of source code. One tool call. Precise answer. No context flood.

Average token savings: 90%+ versus reading files directly.

Install via Claude Code or manually

The fastest way to install is to paste the prompt below into a Claude Code shell. Claude will check your environment, install the tool, and register it as an MCP server — no manual steps needed.

Check whether .NET 10 SDK is installed by running dotnet --version. If the reported version is below 10.0, install it: on Windows run winget install Microsoft.DotNet.SDK.10, on macOS/Linux download from https://dotnet.microsoft.com/download/dotnet/10.0. Verify with dotnet --version once done. Once .NET 10 is confirmed, install CodeMap: if codemap-mcp is not yet installed run dotnet tool install --global codemap-mcp, otherwise run dotnet tool update --global codemap-mcp to get the latest version. Verify the binary is reachable with codemap-mcp --version. Finally, register it as a global MCP server in Claude Code by running claude mcp add codemap-mcp codemap-mcp --scope user and confirm it appears in the output of claude mcp list.

Or install manually:

# Install .NET 10 SDK if needed (Windows)
winget install Microsoft.DotNet.SDK.10

dotnet tool install --global codemap-mcp
codemap-mcp --version
claude mcp add codemap-mcp codemap-mcp --scope user

Requires .NET 10 (LTS). If you're working on a C# or VB.NET codebase you almost certainly have it already — check with dotnet --version.

Upgrading to v2.10.0

No re-index and no index format change. What you may notice:

• index_cleanup and index_remove_repo need repo_path. They delete data, so they no longer fall back to the only registered repo. Cleanup now protects the baseline of every workspace in every CodeMap process, and a baseline that is in use is never half-deleted (it's reported in skipped_in_use).
• References from Blazor components, MVC views and Razor Pages are back (missing since v2.5.2). graph_callers / refs_find on a service or model now include uses written in .razor / .cshtml files. They appear when a baseline is next built (any new commit).
• Idle memory is returned. After indexing, an idle server gives most of the build's memory back (nopCommerce: 2.9 GB → 1.1 GB after 2.5 minutes, 0.2 GB after 13 minutes).
• config.json has two keys, log_level and shared_cache_dir. shared_cache_dir now works (it was never read); budget_overrides is gone (it never had an effect). Unknown keys give one warning in the log. A relative shared_cache_dir stops startup, as a relative CODEMAP_CACHE_DIR does.
• Boolean parameters sent as strings ("include_code": "false") work instead of returning INTERNAL_ERROR.

Details: CHANGELOG [2.10.0].

Upgrading to v2.9.0: new tool names

Tool names use underscores now (symbols_search instead of the dotted form), because LLM APIs don't allow dots in tool names. Claude Code users see no change: Claude Code already showed mcp__codemap__symbols_search. The old dotted names keep working until v2.11.0 at the earliest, for clients and scripts that still send them. Each such call adds a note naming the new tool. If your project's CLAUDE.md contains the CodeMap block, refresh it from docs/CLAUDE-INSERT.MD.

Upgrading from v1.x

v2.0.0 uses a new binary storage engine (memory-mapped segments instead of SQLite). Your old .db baselines are not migrated — run index_ensure_baseline once per repo to rebuild. The old ~/.codemap/baselines/ folder is unused and can be deleted. The SQLite engine was removed in v2.1.0.

The Problem

An AI agent working on a C# codebase without CodeMap does this:

Agent: I need to find who calls OrderService.SubmitAsync.
→ Read OrderService.cs       (3,600 tokens)
→ Read Controllers/...       (3,600 tokens)
→ Grep across src/           (another 3,600 tokens)
→ Maybe find it. Maybe not.

With CodeMap:

refs_find { symbol_id: "M:MyApp.Services.OrderService.SubmitAsync", kind: "Call" }
→ 220 tokens. Exact file, line, and excerpt for every call site. Done.

That's 93.9% fewer tokens for a task agents do dozens of times per session. On a real production codebase (100k+ lines), savings are 95–99%+.

What It Does

CodeMap builds a persistent semantic index from your solution file using Roslyn — the same compiler that powers Visual Studio. Supports both .sln (all Visual Studio versions) and .slnx (VS 2022 17.12+ / .NET SDK 9+) solution formats — auto-discovered when solution_path is omitted (prefers .slnx). Short commit SHAs are auto-expanded. The index captures:

• Every symbol (classes, methods, properties, interfaces, records)
• Every call relationship and reference (who calls what, where)
• Type hierarchy (inheritance chains, interface implementations)
• Architectural facts extracted from code: HTTP endpoints, config keys, DB tables, DI registrations, middleware pipeline, retry policies, exception throw points, structured log templates

All of this is exposed via 28 MCP tools that any MCP-compatible AI agent can call. Starting from v1.3, CodeMap also navigates DLL boundaries — lazily resolving NuGet and SDK symbols on first access, with optional ICSharpCode.Decompiler source reconstruction and cross-DLL call graphs.

Supported languages: C#, VB.NET, and F#. Mixed-language solutions (.sln / .slnx containing C#, VB.NET, and F# projects) are indexed in a single pass. All 28 MCP tools work identically for symbols from any language. C# and VB.NET use Roslyn's MSBuildWorkspace; F# uses FSharp.Compiler.Service (MSBuildWorkspace doesn't support .fsproj). F# architectural fact extractors (endpoints, DI, config) are not yet implemented — symbol search, call graphs, references, and type hierarchy all work.

Blazor / Razor (v2.5.0+): .razor components are indexed via the Razor source generator. ComponentBase-derived classes appear in symbols_search. @page routes surface in surfaces_list_endpoints with a PAGE HTTP method. [Inject] and [Parameter] properties emit dedicated RazorInject / RazorParameter facts. Since v2.10.0, code written in components, MVC views and Razor Pages is part of the reference graph: graph_callers on a service method shows the component or view that calls it. The generator's own scaffolding (tag-helper plumbing) is left out.

Multi-target projects (v2.5.1+): <TargetFrameworks>net8.0;net9.0;net10.0</TargetFrameworks> previously produced one extraction per TFM (3× duplication). CodeMap now collapses to a single extraction on the highest-ranked TFM, with ProjectDiagnostic.TargetFrameworks listing every TFM in the group. Symbol counts on heavily multi-targeted Blazor libraries drop 60–80%.

Interface-aware graph_callers (v2.6.0+): in DI-dispatched codebases (most production .NET) graph_callers on a concrete method silently under-reported because real call sites resolve through the registered interface. CodeMap now detects interface implementation at query time and surfaces an interface_implementation_hint listing the interface members and an estimated count of additional callers routed through them. Pass follow_interface: true to union those into the result (deduped by from_symbol). No baseline-format change, no re-index required. Handles both implicit and explicit interface implementations.

Indexing perf + correctness (v2.5.2): large reduction in indexing wall-clock by skipping auto-generated trees (*.g.cs, *.Designer.cs, files with <auto-generated>, paths under obj/), short-circuiting type-position identifier classification (typeof / generic args / base lists / attributes), and parallelizing Pass-2 reference & fact extraction across projects. Validated on a 9-repo Blazor corpus: Blazorise drops from 408 s → 95 s (−77 %), ant-design-blazor from 47 s → 25 s (−47 %), OrchardCore (single-target sentinel) from 131 s → 96 s (−27 %), and a 78-csproj distributed-database project (ByTech.Bedrock) indexes in 27 s with an 11.2× Pass-2 parallel speedup. Five query-correctness bugs also fixed: symbols_search browse-by-kinds now honours namespace / file_path / project_name filters; workspace-mode namespace filter is case-insensitive (matches committed mode); refs_find cache key includes resolution_state; workspace browse-by-kinds now includes overlay-new symbols; codemap_guide's decision table no longer advertises surfaces_list_di_registrations (which was never a registered tool).

The Transformation

Here's what changes when you give an agent CodeMap:

The agent stops reading your codebase and starts understanding it.

Showpiece: graph_trace_feature

The most powerful tool. Replaces 5–10 manual calls with one:

graph_trace_feature {
  "repo_path": "/path/to/repo",
  "entry_point": "M:MyApp.Controllers.OrdersController.Create",
  "depth": 3
}

Returns an annotated call tree with architectural facts at every node:

OrdersController.Create  [POST /api/orders]
  → OrderService.SubmitAsync
      → [Config: App:MaxRetries]
      → [DI: IOrderService → OrderService | Scoped]
      → Repository<Order>.SaveAsync
            → [DB: orders | DbSet<Order>]
            → [Retry: WaitAndRetryAsync(3) | Polly]

One query. Full feature flow. Every config key touched, every table written, every retry policy applied — surfaced automatically from the index.

Token Savings Benchmark

Measured across 24 canonical agent tasks on a real .NET solution:

Raw tokens = reading all source files. On production codebases (100k+ lines), savings reach 95–99%+.

Run it yourself:

dotnet test --filter "Category=Benchmark" -v normal

28 Tools Across Six Categories

Discover

Navigate

Architecture

Workspace

Index Management

Repo

Workspace Mode — See Your Own Edits

CodeMap tracks uncommitted changes via an overlay index. Every agent session gets its own isolated workspace:

1. index_ensure_baseline   → index HEAD once
2. workspace_create        → agent gets isolated overlay
3. Edit files on disk
4. index_refresh_overlay   → re-indexes only changed files (~63ms)
5. Query with workspace_id → results include your in-progress code

Three consistency modes:

• Committed — baseline index only (default, no workspace needed)
• Workspace — baseline + your uncommitted edits merged
• Ephemeral — workspace + virtual file contents (unsaved buffer content)

Multi-Agent Supervisor Support

Running multiple agents in parallel? What works today:

• Each workspace is isolated: agents that use different workspace_ids never see each other's edits
• workspace_list shows every workspace: IsStale, SemanticLevel, fact count
• Stale detection fires when a workspace's base commit diverges from HEAD
• Supervisor can inspect, clean up, or re-provision any agent's workspace

What to know (measured in the multi-agent benchmark, docs/CODEMAP-MULTI-AGENT-SCALING-FINDINGS.md):

• One CodeMap process per agent. Each MCP client or sub-agent starts its own server, so memory grows with the number of agents: roughly 0.3–0.5 GB per process on small and medium solutions, 2.5–4 GB on large ones. Every process builds the same baseline itself the first time, and concurrent builds slow each other down. A shared host is the planned fix (ADR-048).
• A workspace can't be shared across processes. If another CodeMap process holds that workspace_id, the call returns WORKSPACE_IN_USE (retryable): retry, or give each agent its own ID.
• Agents in fresh worktrees must restore first. A new clone or git worktree add has no restore output; run dotnet restore before index_ensure_baseline, or the index is incomplete (reported as semantic_level: partial, L-12).
• Cleanup is safe across agents (v2.10.0). index_cleanup never removes a baseline that any agent's workspace uses, in any process, and never deletes a baseline another process has open. index_remove_repo refuses (WORKSPACE_IN_USE) while another process has a workspace open in the repo.
• An idle process gives memory back (v2.10.0). Two minutes after indexing, the server compacts its heap; after 10 minutes without an overlay refresh it also drops the cached Roslyn solution (the next refresh reopens it). This lowers each process's idle cost; the per-process multiplication stays until the shared host.

Self-Healing Under Broken Builds

When a file doesn't compile, CodeMap doesn't drop references. It stores unresolved edges with syntactic hints. When compilation succeeds again (after a fix), a resolution worker automatically upgrades them to fully-resolved semantic edges.

refs_find returns both. Filter with resolution_state: "resolved" if you need certainty.

DLL Boundary Navigation

CodeMap resolves DLL symbols lazily on first agent access — NOT_FOUND at a DLL boundary triggers automatic extraction rather than a dead end.

Two levels, both permanent (cached in baseline DB):

After Level 2, cross-DLL call graph edges are extracted so graph_callees and graph_trace_feature traverse INTO and THROUGH DLL code seamlessly.

source discriminator in symbols_get_card response:

• "source_code" — symbol is from your own source
• "metadata_stub" — Level 1 only (decompilation unavailable)
• "decompiled" — Level 2 source reconstructed and ready

graph_trace_feature applies a max_lazy_resolutions_per_query budget (default 20) when encountering previously-unseen DLL types to bound decompilation latency.

Shared Baseline Cache

Index once, reuse everywhere — across machines, CI, Docker containers:

export CODEMAP_CACHE_DIR=/shared/codemap-cache

or, in config.json (see Observability): { "shared_cache_dir": "/shared/codemap-cache" }.

• index_ensure_baseline pulls from cache first (~2ms vs ~30s Roslyn build)
• Auto-push after every new baseline build, except baselines this machine degraded (an unrestored checkout or a source generator that failed to load), since another machine would get an incomplete index
• Self-healing: corrupt cache entries are detected and overwritten
• Precedence: CODEMAP_CACHE_DIR (whenever set; blank = disabled) > config.json shared_cache_dir > disabled. With neither, all cache ops are no-ops
• Must be an absolute path or start with ~/; a relative value stops the server at startup (exit code 2), since it would resolve against the client's working directory

v2 Storage Engine — 10x Faster Queries

v2.0.0 replaces SQLite with a custom binary storage engine using memory-mapped segment files. The Roslyn extraction pipeline is unchanged — only the on-disk format is new.

Query speedup (measured across 15 query types on real repos):

Indexing speedup (Roslyn compilation dominates, but I/O is faster):

What changed:

• Baselines stored as contiguous packed binary segments (symbols, edges, files, facts) with mmap reads — no SQL parsing overhead
• Custom search index with tokenized FTS (CamelCase splitting, signature/documentation indexing)
• WAL-backed overlay for workspace mutations (same isolation model)
• Zero native DLL dependencies (no e_sqlite3.dll)

Validated on 9+ repos including dotnet/roslyn (174K symbols, 768K references), dotnet/fsharp (157K symbols via FCS), and Bitwarden. Zero functional bugs. See docs/ENGINE-COMPARISON-RESULTS.MD for full data.

Self-Hosting Validated

CodeMap indexes its own 18-project solution (5,576 symbols, 20,960 references). All 28 tools verified against real-world architectural complexity. Self-hosting exposed and fixed cross-project reference bugs, CamelCase FTS edge cases, overlay StringId resolution issues, and multi-line SQL extraction gaps. Every tool in this README was tested against the codebase that implements it.

Installation

.NET Global Tool — NuGet (recommended)

See the Install via Claude Code or manually section at the top for the one-paste Claude Code prompt and manual steps.

NuGet package: nuget.org/packages/codemap-mcp

Docker

docker build -t codemap-mcp .
docker run -i \
  -v /path/to/your/repo:/repo:ro \
  -v /path/to/cache:/cache \
  codemap-mcp

-i is required — MCP uses stdio transport. Without it the container gets immediate EOF.

Uses the .NET SDK base image (~800MB) because MSBuildWorkspace needs MSBuild at runtime for index_ensure_baseline. Mount a cache volume (-v /path/to/cache:/cache) to avoid rebuilding the index on every container start.

Connect to Your AI Agent

Claude Code (Claude Desktop / claude.ai)

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "codemap": {
      "command": "codemap-mcp"
    }
  }
}

Any MCP-Compatible Client

CodeMap speaks standard MCP over stdin/stdout (JSON-RPC 2.0). Any MCP client works.

CLAUDE.md Integration

Drop the instruction block from docs/CLAUDE-INSERT.MD into your project's CLAUDE.md to wire up automatic CodeMap usage for any Claude agent working on that project. The block includes the session startup sequence, a tool substitution decision table, and the "refresh before grep" rule that keeps agents in semantic mode.

Tip: Write XML Docs — CodeMap Uses Them

CodeMap indexes /// <summary> XML doc comments on all classes, methods, and interfaces. They appear in symbols_get_card, symbols_get_context, and symbols_search results — giving agents intent and context without reading implementations.

When writing C# code with CodeMap enabled, always add XML doc comments. This isn't just style — it directly improves every downstream query. Agents using graph_trace_feature see annotated call trees that read like specs. codemap_export includes docs in the portable context for other LLMs.

See docs/CODEMAP-AGENT-GUIDE.MD for the full agent workflow guide.

Architecture

Your Git repo              CodeMap Server
     │                          │
     │  repo_path               │
     ├─────────────────────────►│ GitService         (repo identity, HEAD SHA)
     │                          │      │
     │  solution.sln/.slnx      │      ▼
     ├─────────────────────────►│ RoslynCompiler     (MSBuildWorkspace for C#/VB, FCS for F#)
     │                          │      │
     │                          │      ▼
     │                          │ Extractors         (Symbols + Refs + TypeRelations + Facts)
     │                          │      │
     │                          │      ▼
     │                          │ CustomSymbolStore   (v2 binary segments, mmap'd)
     │                          │      │        ↕
     │                          │      │   SharedCache  (file-based, optional)
     │                          │      ▼
     │  your uncommitted edits  │      ▼
     ├─────────────────────────►│ OverlayStore        (WAL-backed incremental overlay)
     │                          │      │
     │                          │      ▼
     │                          │ MergedQueryEngine  (baseline + overlay, transparent merge)
     │                          │      │
     │  MCP tool call           │      ▼
     ├─────────────────────────►│ McpServer          (stdio JSON-RPC 2.0, 28 tools)
     │                          │      │
     │  JSON response           │      ▼
     │◄─────────────────────────│ ResponseEnvelope   (answer + evidence + timing + token savings)

Layer dependencies (enforced at build time — violations are build errors):

CodeMap.Core      ← zero dependencies (domain types + interfaces)
CodeMap.Git       ← Core (LibGit2Sharp)
CodeMap.Roslyn    ← Core (Roslyn 5.x + MSBuildWorkspace)
CodeMap.Storage.Engine ← Core (v2 binary segments, sole engine since v2.1.0)
CodeMap.Query     ← Core + Storage.Engine (query engine + cache + overlay merge)
CodeMap.Mcp       ← Core + Query (MCP tool handlers)
CodeMap.Daemon    ← ALL (DI composition root, the executable)

Observability

Every response includes:

• Per-phase timing — cache_lookup_ms, db_query_ms, ranking_ms (sub-millisecond on v2)
• Token savings — tokens saved and cost avoided vs raw file reading
• Semantic level — Full / Partial / SyntaxOnly (index quality signal, see below)
• Overlay revision — which workspace revision answered the query
• Workspace ID — which workspace context answered (null for committed mode)

When the level isn't Full, index_ensure_baseline's project_diagnostics says why, per project:

• compiled: false: the project couldn't be compiled, so it was indexed syntactically.
• missing_restore_output: the checkout was never restored (fresh clone or git worktree) and packages, sometimes the framework too, are unresolved. Run dotnet restore, then rebuild the baseline (L-12).
• generator_load_failures: a source generator from a newer .NET SDK couldn't load, so its code (e.g. Razor components) is missing (L-11).

The same level is reported on every later query against that baseline (for baselines built by v2.8.2 or later; older ones don't record these fields).

Structured logs to ~/.codemap/logs/codemap-{date}.log (daily rotation, JSON lines). Cumulative savings to ~/.codemap/_savings.json (persists across restarts). Config at ~/.codemap/config.json, read once at startup. It has two keys:

{ "log_level": "Information", "shared_cache_dir": "/shared/codemap-cache" }

• log_level: Trace, Debug, Information (default), Warning or Error.
• shared_cache_dir: the shared baseline cache, used only when CODEMAP_CACHE_DIR is unset.

Any other key (a typo, or budget_overrides, which never had an effect and was removed in v2.10.0) is ignored with one warning in the log.

Data Directory (CODEMAP_HOME)

Everything above lives under one data root, ~/.codemap by default. Set CODEMAP_HOME to move it (config, logs, store, _savings.json), for example per test run, CI job or sandbox. It must be an absolute path or start with ~/; a relative value stops the server at startup (exit code 2).

Baselines are stored in <data root>/store/<repoId>/baselines/<commitSha>/ as binary segment files. Overlays are in <data root>/store/<repoId>/overlays/<workspaceId>/. Use index_list_baselines to inspect and index_cleanup to reclaim space.

Known Limitations & Coverage Gaps

CodeMap won't surface a hit in every situation a grep would. The most common reasons are documented in docs/KNOWN-LIMITATIONS.md. Top items to be aware of:

• Multi-target conditional symbols. #if NET8_0-only types are invisible — extraction runs on the highest TFM only (L-01).
• Legacy MVC MapControllerRoute — convention-routed actions don't surface in surfaces_list_endpoints. Only attribute routing, minimal API, and Blazor @page are extracted (L-02).
• F# fact extractors not yet wired — F# gets symbols/refs/hierarchy only; endpoints / DI / config / DB tables don't extract from .fsproj yet (L-05).
• Unrestored checkout — a fresh clone or git worktree has no obj/project.assets.json, and CodeMap doesn't restore, so package references (sometimes the framework itself) are unresolved. It's reported (missing_restore_output, semantic_level: partial); run dotnet restore before index_ensure_baseline (L-12).
• .NET SDK newer than CodeMap's Roslyn — that SDK's source generators (e.g. Razor) may not load, and their output is missing. Reported as generator_load_failures; update CodeMap (L-11).
• Fresh clone with no build — Razor source-generator output may be invisible until you dotnet build once (L-08).
• Razor markup-only uses — @onclick="Save" and a component tag like <Greeting /> produce no reference yet, and Razor references point at the generated obj/…_razor.g.cs lines, not the .razor file (L-15).

When symbols_search returns nothing for code you can see in the editor, scan KNOWN-LIMITATIONS first before falling back to grep.

Documentation

Build & Test

# Build (zero warnings enforced)
dotnet build -warnaserror

# Test per project: `dotnet test` at the solution root starts every test runner at once and can run out of memory
dotnet run --project tests/CodeMap.Core.Tests          # likewise Roslyn, Query, Mcp, Storage.Engine, Daemon, Git

# Integration tests: restore the test fixtures once per checkout first (MSBuildWorkspace doesn't restore)
dotnet restore testdata/SampleSolution/SampleSolution.sln   # likewise SampleVbSolution, SampleBlazorSolution, SampleFSharpSolution
dotnet run --project tests/CodeMap.Integration.Tests   # includes Category=Concurrency (spawns real servers, ~1 min)

# Linux test rig (Docker; see docs/DEVELOPER-GUIDE.MD)
.\tests\docker\run-linux-tests.ps1 integration

# Token savings benchmark
dotnet test --filter "Category=Benchmark" -v normal

# Performance microbenchmarks (BenchmarkDotNet)
cd tests/CodeMap.Benchmarks && dotnet run -c Release

Performance Reference

What to expect when running CodeMap on your codebase. All v2 engine numbers (default since v2.0.0).

Indexing time by repo size

Subsequent runs on the same commit return immediately (already_existed: true). Incremental overlay refresh (after editing files) takes ~63ms.

Query response time (v2 engine)

Cold times scale with repo size (more symbols = more BFS/join work). Warm times are nearly flat across all repo sizes — L1 cache caps at 10,000 entries with LRU eviction.

Memory footprint (v2 engine)

mmap pages are demand-loaded by the OS — resident memory stays proportional to queries made, not total index size.

Process memory while idle (v2.10.0+)

The server process also holds what Roslyn used to build the index. Since v2.10.0 it gives that memory back when idle: one compacting GC two minutes after indexing, and the cached solution for overlay refreshes is dropped after 10 minutes without a refresh. Private memory of one idle server, measured (PHASE-21-12, docs/benchmarks/phase-21-12/):

Each server logs MEMORY_SNAPSHOT lines (working set, private bytes, GC heap) to ~/.codemap/logs.

28 MCP tools. 90%+ token savings. Roslyn-grade semantics. C#, VB.NET, F#, Blazor/Razor. DLL boundary navigation. .sln + .slnx auto-discovery. v2.10.0 — safe destructive tools across agents, idle memory returned, Razor component and view references, honest config.json. Validated on dotnet/roslyn (174K symbols), dotnet/fsharp (157K symbols), and a 9-repo Blazor corpus including Blazorise, MudBlazor, ant-design-blazor, OrchardCore. Your agent deserves better than grep.