TestAtlas

The problem

Large test-automation solutions are hard to navigate — for humans and for AI agents. Asked to automate a new story, an agent can't see which steps already exist, where similar code lives, or what conventions the solution follows — so it duplicates steps and misplaces code. TestAtlas indexes the solution once into a single SQLite map and answers those questions precisely: deterministically, offline, without a model or a network call.

See it in action

Index the bundled 8-project sample once — testatlas index samples/SampleShop/SampleShop.sln — then ask it questions from the terminal. Every number below is real output from that run:

$ testatlas stats sampleshop.db
TestAtlas map: sampleshop.db (schema v5)

totals: 8 project(s), 15 class(es), 36 method(s), 14 step definition(s)
class kinds:
  api_client     7
  page_object    4
  step_class     4

gherkin: 4 feature(s), 5 scenario(s), 16 step(s)
bound steps: 16 · unbound: 0 · ambiguous: 0
endpoints: 3 (3 call site(s))

…or let your agent ask them over MCP. Here an agent checks whether a step it is about to write already exists — resolve_step answers with the definitions that would bind it, or the closest near-misses to reuse instead:

// agent → testatlas: resolve_step { "text": "I add the product to my cart" }
{
  "status": "none",            // nothing binds this exact text — don't invent it from scratch:
  "suggestions": [             // these existing steps are the closest, ranked by shared terms
    { "expression": "product (.*) is added to the cart with quantity (.*)",
      "keyword": "When", "location": "SampleShop.Tests.Api/Steps/CatalogApiSteps.cs:34" },
    { "expression": "the cart is not empty",
      "keyword": "Then", "location": "SampleShop.Tests.Api/Steps/CatalogApiSteps.cs:43" }
    // …8 more
  ]
}

Live sample outputs, committed from that same solution: HTML report (features, scenarios, bindings, class kinds, endpoints) · dependency map (the eight projects in layers — click one to see what it uses, what uses it, and which tests a change can reach). (GitHub serves .html as source, so the links route through htmlpreview.github.io — or download from docs/ and open locally.)

What you get

TestAtlas statically analyses the solution and emits codemap.db — projects and their dependency edges, Gherkin features/scenarios/steps, step definitions and their bindings (bound / unbound / ambiguous), page objects, API clients, helpers, test classes, and the call/usage edges connecting them — then turns that map into answers:

All of it offline, deterministic, and reproducible — same input, same map, every time.

🚀 Quick start

Requires the .NET SDK 8.0+. On a corporate machine where dotnet tool install fails with 401, see docs/troubleshooting.md.

1 — Install the CLI (and the MCP server, if you'll connect an agent):

dotnet tool install --global TestAtlas.Cli
dotnet tool install --global TestAtlas.Mcp

2 — Index your solution. This produces the map (./codemap.db) that every query, report, and MCP answer reads — nothing works without it:

testatlas index path/to/YourSolution.sln

No need to build or restore the solution first — indexing is a syntax-only pass, so an unrestored checkout maps fine. Point index at a folder (or nothing) and it auto-discovers a single .sln/.csproj there.

3 — Query it:

testatlas stats
testatlas search "login"
testatlas report        # writes codemap.html
testatlas map           # writes codemap-map.html

…and to serve it to your AI agent, continue to MCP setup.

git clone https://github.com/Karzone/TestAtlas.git
cd TestAtlas
dotnet build TestAtlas.sln
dotnet run --project src/CodeMap.Cli -- index path/to/YourSolution.sln

🔌 Use it from an AI agent (MCP)

TestAtlas ships an MCP server — testatlas-mcp — that serves the map to any MCP-aware client (Visual Studio / VS Code Copilot, Claude Code, and others) over stdio JSON-RPC. The agent asks a precise question and gets an exact, structured answer straight from the .db — instead of stuffing source files into its context window.

Prerequisites: both tools installed and a map built — steps 1–2 of the Quick start. Then register the server:

Visual Studio / VS Code (GitHub Copilot agent mode) — add to your .mcp.json (%USERPROFILE%\.mcp.json or <SolutionDir>\.mcp.json):

{
  "servers": {
    "testatlas": {
      "type": "stdio",
      "command": "testatlas-mcp",
      "args": ["C:\\path\\to\\codemap.db"]
    }
  }
}

Pass the map path explicitly (as above, or via a TESTATLAS_DB env var) — most agents launch the server from their own working directory, not your solution folder, so relying on auto-discovery leaves the server with no map: it starts and lists its tools, but every call answers with an error saying how to supply one (releases before 0.1.11 exited with code 2 instead). In Visual Studio you can also use Tools picker → + → Add custom MCP server to write this entry for you. On the .NET 10 SDK you can skip the install and use "command": "dnx", "args": ["TestAtlas.Mcp", "--yes", "C:\\path\\to\\codemap.db"] — dnx fetches and runs the server on demand.

Claude Code — as a plugin (server, a skill that says when to use which tool, /testatlas:index, /testatlas:report, and a session-start check that tells Claude when the map is missing or stale):

claude plugin marketplace add Karzone/TestAtlas
claude plugin install testatlas@testatlas

The plugin starts testatlas-mcp with no arguments, so it serves the codemap.db at the root of whatever project the session is opened in — build it with /testatlas:index; from 0.1.12 the server picks the map up on the next call, with nothing to reconnect. Both tools must be on the PATH (dotnet tool install --global TestAtlas.Cli, then the same for TestAtlas.Mcp). The plugin lives in claude-plugin/.

Claude Code — one server, no plugin:

claude mcp add testatlas -- testatlas-mcp path/to/codemap.db
claude mcp list        # the testatlas row should read: ✔ Connected

By default the server is registered for the current project (--scope local). Add --scope user to make it available in every project on your machine, or --scope project to share the registration with your team via a committed .mcp.json.

[!IMPORTANT] MCP clients load servers at session start — if you register mid-session, restart your agent session before the testatlas tools appear. To confirm it's actually being used (and not silently ignored), run the checks in docs/troubleshooting.md.

Tools exposed:

• resolve_step — resolve a Gherkin phrase to the existing step definition(s) that would bind it (regex/cucumber, keyword-agnostic). exact / ambiguous / none (+ near-match suggestions ranked by shared terms). Reuse-first authoring: don't write a step that already exists.
• step_catalog — the reusable step vocabulary with extracted placeholders and allowed values (cucumber {type}, regex (a|b) enums). Compose scenarios from what exists.
• impact — blast radius of a change: the scenarios affected by a given class, method, step definition, or endpoint.
• search_steps — full-text search over step definitions (expression text + method + class name).
• search_scenarios — full-text search over scenarios (feature + scenario name + step text + tags).
• get_scenario — full detail of scenario(s) by name: feature, tags, kind, example-row count, and the ordered steps.
• get_step_definition — full detail of step definition(s) by expression: keyword, params, C# class/method/signature, and the scenarios that use it.
• list_tags — the tag taxonomy with per-tag scenario counts, most-used first — tag new scenarios consistently.
• list_endpoints — the HTTP endpoints the suite calls, each with verb, route, and scenario blast radius (highest-reach first).
• project_dependencies — the implied project dependency graph (depends-on / depended-on-by), e.g. "what depends on the Party project?".
• stats — summary counts: projects, classes, methods, class-kind breakdown, endpoints, and edge tallies.

Retrieval runs locally against the SQLite file — deterministic, offline, and a few hundred tokens per answer. Protocol details in specs/codemap-mcp.md; registration from source and every failure mode in docs/troubleshooting.md.

📖 Commands

index  --output <file> · --config <file> · --include <glob> (repeatable) · --exclude <glob> (repeatable) · --verbose · --quiet

search  --steps (step definitions only) · --scenarios (scenarios only)

Exit codes  0 ok · 1 completed with warnings · 2 fatal · 3 bad arguments

Run testatlas --help for the full usage text.

# Index the solution (no build needed — the pass is syntax-only)
testatlas index YourSolution.sln --output atlas.db

# Before writing a new step — does one already exist?
testatlas search atlas.db "add a product to the cart" --steps

# About to change a shared client — what will it hit?
testatlas impact atlas.db --class ProductsApiClient

# Share human-readable snapshots
testatlas report atlas.db --html atlas.html
testatlas map    atlas.db --html atlas-map.html

samples/SampleShop is a self-contained 8-project solution mixing API tests and UI tests, so the map has plenty of connected nodes — real HttpClient API clients, real Selenium IWebDriver page objects, and Reqnroll suites driving both:

                       ┌─▶ Api.Catalog  ──┐
Tests.Api  ────────────┼─▶ Api.Cart     ──┤
Tests.E2E  ──┬─────────┴─▶ Api.Identity ──┼─▶ Core   (ApiClientBase : HttpClient)
             └─▶ Ui.Pages ─────────────────┘          (PageBase      : IWebDriver)
Tests.Ui   ────▶ Ui.Pages ──────────────────▶ Core

Reproduce the committed sample outputs yourself:

testatlas index  samples/SampleShop/SampleShop.sln --output sampleshop.db
testatlas report sampleshop.db --html docs/sample-report.html
testatlas map    sampleshop.db --html docs/sample-map.html

🔄 Keeping the map fresh

Answers are deterministic — but only as fresh as the map, so re-index on change, not on a timer. A full re-index is a single static pass (seconds), and its cost scales with solution size, not with how much changed:

• Locally — python scripts/check-map-age.py tells you when your map drifted; a version-controlled post-merge git hook can warn automatically after every pull.
• In CI (team model) — re-index on every merge to main and publish the .db to a shared feed (never commit it — it's a build artifact). Then teammates and agents only need TestAtlas.Mcp locally: download the shared map and point TESTATLAS_DB at it — no local TestAtlas.Cli or indexing required. (Index locally only to include your own uncommitted branch work.)

Details — the staleness checker, the git hook, and copy-paste CI recipes (GitHub Actions + Azure DevOps with a Universal feed) — in docs/keeping-the-map-fresh.md.

🧹 Uninstall

Two separate things — remove them in this order, so the editor isn't launching a server whose binary just vanished:

1. Remove the MCP registration — delete the testatlas block from your .mcp.json (<SolutionDir>\.mcp.json or %USERPROFILE%\.mcp.json), then restart Visual Studio / VS Code. (Or just toggle it off in the agent's tools/wrench picker to keep it for later.)
2. Uninstall the tools — they're .NET global tools, not editor add-ins:

   dotnet tool uninstall --global TestAtlas.Mcp
   dotnet tool uninstall --global TestAtlas.Cli
   dotnet tool list --global            # confirm they're gone
   

3. (Optional) delete the codemap.db map file — it's just data, nothing else references it.

🎯 Design tenets

• Zero config — a useful map on an unseen solution, no config file required.
• Solution agnostic — heuristic, overridable detection; no company-specific assumptions.
• Deterministic & offline — same input ⇒ byte-equivalent logical content; no network, no AI.
• Graceful degradation — solutions without Gherkin still yield a useful map.
• Public schema as contract — a versioned SQLite schema, so downstream consumers keep working even if a third party swaps in their own indexer.

The project folders use the indexer's working name (CodeMap); the shipped tools and packages are TestAtlas. Full specs: specs/codemap-indexer.md · specs/codemap-mcp.md. Repo layout: CONTRIBUTING.md.

🗺 Roadmap

• Indexer CLI — C# indexer + documented, versioned SQLite schema
• HTML visualization — self-contained report + project map generated from the db
• MCP server — testatlas-mcp exposes the map to AI agents over stdio JSON-RPC (11 tools)
• Second-language indexer — same schema, contract-tested

Deliberately not planned (see design tenets): LLM-assisted analysis inside the indexer, network calls at index/query time, running or generating tests, and semantic (compilation-based) analysis that would require a restored build. Releases and per-version notes live on the releases page; current distribution channels in docs/DISTRIBUTION.md.

📄 License

MIT © 2026 Karthik Kalaiyarasu