Karate Graph

Karate Graph is an MCP-powered analyzer for Karate Framework projects. It helps AI agents scan .feature files, Java helpers, JavaScript helpers, database usage, execution reports, and dependency graphs so the AI can answer impact, reuse, failure, and visual-report questions with project context.

What It Helps With

• Scan Karate projects and build dependency graphs.

• Generate interactive visual HTML reports.

• Search API, workflow, feature, Java, JavaScript, DB, and reusable helper usage.

• Analyze change impact and suggest test cases to rerun.

• Inspect failure hotspots, flaky risk, execution history, and debug context.

• Help AI understand feature intent, variables, assertions, duplicate steps, and call/read chains.

Related MCP server: codescan

Visual Report Preview

Real project scan:

Clean overview mockup:

Install

Install from PyPI:

pip install -U karate-graph

Install a specific version:

pip install -U karate-graph==0.1.1

Verify the package:

python -m pip show karate-graph
python -m karate_graph_analyzer.mcp_server --help
python -m karate_graph_analyzer.mcp_server --version

If karate-graph-mcp is on your PATH, this also works:

karate-graph-mcp --help
karate-graph-mcp --version

On Windows, if where karate-graph-mcp cannot find the command, use the python -m MCP config below. That avoids PATH issues.

Configure MCP

Karate Graph is mainly used through an MCP client such as Codex, Claude Desktop, Cursor, or another AI tool that supports MCP.

Recommended Windows Config

Use the full Python path from:

where python

Example Codex config.toml:

[mcp_servers.karate-graph]
command = "C:\\Program Files\\Python311\\python.exe"
args = ["-m", "karate_graph_analyzer.mcp_server"]
env = { PYTHONIOENCODING = "utf-8" }

If your Python is installed somewhere else, replace command with that path.

Generic MCP JSON Config

{
  "mcpServers": {
    "karate-graph": {
      "command": "python",
      "args": ["-m", "karate_graph_analyzer.mcp_server"],
      "env": {
        "PYTHONIOENCODING": "utf-8"
      }
    }
  }
}

Local Source Development Config

Only use PYTHONPATH when you want the MCP server to run directly from your local repo source instead of the package installed by pip.

{
  "mcpServers": {
    "karate-graph": {
      "command": "python",
      "args": ["-m", "karate_graph_analyzer.mcp_server"],
      "env": {
        "PYTHONPATH": "E:/Project/auto/karate_graph/src",
        "PYTHONIOENCODING": "utf-8"
      }
    }
  }
}

After editing MCP config, restart your MCP client.

End-To-End Usage

The normal flow is:

1. Check MCP server health.

2. Register the Karate project.

3. Analyze the project.

4. Generate the visual report.

5. Use search, impact, DB, reuse, and failure tools.

1. Health Check

Ask your AI to call:

mcp_health()
mcp_version()
list_projects()

Expected result:

• MCP server is reachable.

• Package version is returned.

• Existing registered projects are listed.

2. Register Project

Register a Karate project once:

register_project(
  name="karate-core",
  root_path="E:/Project/auto/karate-fw/karate-core",
  feature_file_patterns=["**/*.feature"]
)

Notes:

• name is the project alias used by all other tools.

• root_path should be the absolute path to the Karate project.

• feature_file_patterns can be omitted; default is usually ["**/*.feature"].

For very large projects where real test cases live only under feature(s)/**testcase, enable strict test-case directory mode. This lets reusable feature packages such as reuse, shared, helpers, or services stay outside the test-case count without renaming the project structure:

register_project(
  name="karate-core",
  root_path="E:/Project/auto/karate-fw/karate-core",
  feature_file_patterns=["**/*.feature"],
  parser_config={
    "strict_test_case_directory_mode": true,
    "test_case_directories": ["testcase", "testcases"],
    "feature_root_directories": ["feature", "features"],
    "default_non_test_feature_type": "COMMON",
    "common_directories": ["common", "services", "reuse", "reusable", "shared", "helpers"],
    "scan_exclude_directories": [
      ".git", ".gradle", ".idea", ".karate_cache", "build", "dist",
      "node_modules", "out", "output", "reports", "target"
    ],
    "javascript_file_patterns": ["**/*.js"],
    "large_project_streaming_scan": true,
    "scan_parallel_enabled": true,
    "scan_parallel_workers": 8,
    "scan_parallel_threshold": 2000,
    "incremental_scan_enabled": true,
    "scan_store_enabled": true,
    "scan_log_every": 2000,
    "visualization_large_graph_threshold": 1500,
    "visualization_physics_enabled": null,
    "visualization_node_limit": 5000,
    "visualization_edge_limit": 12000,
    "visualization_progressive_enabled": true,
    "visualization_chunk_size": 1000,
    "visualization_auto_load_chunks": 0,
    "cycle_detection_enabled": true,
    "cycle_detection_node_limit": 20000,
    "ai_context_cache_feature_threshold": 5000,
    "ai_context_similarity_pool_limit": 1000,
    "ai_context_duplicate_location_limit": 25
  }
)

When parser_config is provided, Karate Graph still keeps values auto-detected from karate-config*.js and only applies the explicit overrides above.

For projects with tens of thousands of feature files, prefer narrow feature patterns when possible:

feature_file_patterns=[
  "src/test/java/features/**/*testcase/**/*.feature",
  "src/test/java/feature/**/*testcase/**/*.feature",
  "src/test/java/features/**/reuse/**/*.feature",
  "src/test/java/feature/**/reuse/**/*.feature"
]

Performance tips:

• Use large_project_streaming_scan=true for very large repositories. It avoids keeping every parsed feature AST in memory.

• If not set manually, streaming scan is automatically used when the matched feature count reaches large_project_streaming_threshold (default 20000).

• Keep include_structural_nodes=false for the first scan. Turn it on only when you need folder/file nodes in the visual graph.

• Keep scan_exclude_directories updated with generated folders such as target, reports, output, and node_modules.

• Narrow javascript_file_patterns if only some JavaScript folders matter, for example ["src/test/java/**/*.js", "src/test/resources/**/*.js"].

• For huge graphs, leave visualization_physics_enabled=null. Karate Graph will disable physics automatically after visualization_large_graph_threshold nodes.

• Full 100k-node interactive canvases are not practical in a browser. The report opens with a capped working view (visualization_node_limit) and, when visualization_progressive_enabled=true, writes extra .js chunks next to the HTML so you can load more nodes from the report without freezing the first page load.

• Keep visualization_auto_load_chunks=0 for very large reports. Use the report's Load more button to add chunks on demand.

• Keep cycle_detection_node_limit enabled. Full cycle enumeration is skipped above the limit because it can dominate scan time on 100k-node graphs.

• AI feature/DB context tools stream large feature sets instead of caching every parsed AST after ai_context_cache_feature_threshold.

• scenario_similarity_map uses ai_context_similarity_pool_limit to avoid O(N²) comparisons across the whole repository.

• Duplicate step/flow tools count all matches but only keep ai_context_duplicate_location_limit example locations per group.

• Turn on scan_parallel_enabled=true for faster parsing on large repos; tune workers with scan_parallel_workers.

• Keep incremental_scan_enabled=true to re-parse only changed files and skip unchanged files when scanning again.

• Use get_scan_output(project_name) after analyze to inspect changed_files, unchanged_files, deleted_files, node/edge counts, and cache usage.

3. Scan / Analyze Project

Build the dependency graph:

analyze_project(
  project_name="karate-core",
  include_structural_nodes=true
)

Use include_structural_nodes=true when you want the visual report to include folder/file structure. Use false for a cleaner graph focused on test and dependency nodes.

4. Generate Visual Report

Create the interactive HTML graph:

visualize_project(
  project_name="karate-core",
  output_path="E:/Project/auto/karate-fw/karate-core/output/karate_graph.html"
)

Open the generated file in your browser:

E:/Project/auto/karate-fw/karate-core/output/karate_graph.html

The dashboard supports search by:

• test case ID or Jira tag

• scenario name

• feature path

• API path

• Java class or method

• JavaScript file or function

• database table, host, provider, or query keyword

5. Render Execution Report

If you have Karate JSON execution reports, process a report folder:

process_reports_folder(
  project_name="karate-core",
  directory_path="E:/Project/auto/karate-fw/karate-core/target/karate-reports",
  output_path="E:/Project/auto/karate-fw/karate-core/output/execution_graph.html"
)

Or render one report file:

render_execution_report(
  project_name="karate-core",
  report_path="E:/Project/auto/karate-fw/karate-core/target/karate-reports/karate-summary-json.txt",
  output_path="E:/Project/auto/karate-fw/karate-core/output/execution_graph.html"
)

Execution reports add pass/fail/not-run status, failure hotspots, flaky risk, failure history, and debug context.

Common AI Prompts

After MCP is configured, you can ask your AI:

Register project karate-core at E:/Project/auto/karate-fw/karate-core and analyze it.

Generate visual report for karate-core and tell me the output file path.

Search all test cases related to order creation.

Find reusable random string or random number helpers before adding a new one.

If locators/checkout.json changes, which test cases are impacted?

Which feature steps are duplicated and safe to refactor?

Which test cases touch the orders table?

Show the top failure hotspots and prioritize what to fix first.

Build an AI debug context pack for this failed node and error message.

Tool Groups

Project And Graph

• mcp_health() checks MCP server connectivity.

• mcp_version() returns package version.

• register_project(name, root_path, feature_file_patterns, parser_config) registers a project.

• list_projects() lists registered projects.

• delete_project(name) removes one project from registry.

• clear_all_projects() resets all registered projects.

• analyze_project(project_name, include_structural_nodes) scans a project and builds graph data.

• bulk_analyze() analyzes all registered projects.

• query_dependencies(component_id, transitive) finds dependencies of a component.

• impact_analysis(component_id) finds impacted tests and workflows from a changed component.

• get_subgraph(node_id, radius) returns local graph context around a node.

• find_path(source_id, target_id) finds dependency paths between nodes.

• global_search(query) searches all graph nodes.

• query_node_by_metadata(key, value) searches graph nodes by metadata.

Visual And Export

• visualize_project(project_name, output_path) generates an interactive graph HTML file.

• render_execution_report(project_name, report_path, output_path) generates a graph from one execution report.

• process_reports_folder(project_name, directory_path, output_path) scans a report folder and generates execution visualization.

• compare_projects(base_project_name, new_project_name, output_path) creates a diff visualization.

• merge_projects(project_names, new_project_name) merges multiple project graphs.

• export_graph(project_name, format) exports graph data as json or graphml.

API, Workflow, Test Case, Page

• search_api(project_name, method, path, domain) searches API endpoints.

• get_api_stats(project_name, keyword) summarizes API usage.

• search_workflow(project_name, path, scenario_tag, keyword) searches workflow/scenario usage.

• search_test_case(project_name, jira_tag, name_pattern) searches test cases by ID/tag/name.

• get_page_stats(project_name, domain) summarizes page/action usage.

• common_usage_map(project_name, limit) shows commonly reused components.

• unused_components(project_name, limit) lists unused components.

Java And JavaScript Reuse

• search_java_usage(project_name, query, include_methods) finds Java class/method usage and linked tests.

• search_js_usage(project_name, query, include_functions) finds JavaScript file/function usage and linked tests.

• javascript_structure_map(project_name, limit) shows JS files, functions, dependencies, and test usage.

• search_reusable_function(project_name, query, language, limit) searches Java/JS helpers for reuse candidates.

search_reusable_function returns tags, aliases, usage examples, stability score, source location, and related test usage where available.

Smart Retest And Change Impact

• change_impact_preview(project_name, changed_paths, limit) previews tests impacted by changed files/components.

• test_selection_suggestion(project_name, changed_paths, limit) suggests a compact rerun set.

• get_impact_radius(node_id, depth) shows impact within a graph radius.

• get_component_importance(project_name) ranks important graph components.

Example:

change_impact_preview(
  project_name="karate-core",
  changed_paths=["src/test/java/helpers/RandomUtil.java", "locators/checkout.json"],
  limit=30
)

Feature Understanding

• feature_intent_index(project_name, query, limit) summarizes scenario intent and major signals.

• feature_behavior_map(project_name, feature_path, scenario_tag, scenario_name, node_id, limit) groups preconditions, actions, and expectations.

• variable_data_flow_trace(project_name, feature_path, scenario_tag, scenario_name, node_id, limit) traces variables from definition to usage.

• assertion_map(project_name, query, limit) indexes status/match/assert checks.

• call_read_deep_context(project_name, feature_path, scenario_tag, scenario_name, node_id, max_depth, limit) expands nested call/read chains.

• ai_feature_context_pack(project_name, feature_path, scenario_tag, scenario_name, node_id, max_call_depth, limit) returns an AI-ready context pack.

• scenario_similarity_map(project_name, query, limit, top_k) finds similar scenarios.

• feature_reuse_advisor(project_name, min_group_size, min_flow_length, limit, include_low_signal) finds duplicate steps and repeated flows.

• similar_common_components(project_name, limit) finds common components with similar dependency shape.

Use include_low_signal=false in feature_reuse_advisor to ignore generic Karate template steps such as When method POST or Then status 201.

Database Understanding

• db_query_index(project_name, query, limit, include_components, link_status) indexes DB queries and DB components.

• search_db_usage(project_name, query, limit, link_status) searches by table, operation, host, dialect, provider, path, or query keyword.

• db_data_flow_trace(project_name, feature_path, scenario_tag, scenario_name, node_id, limit) traces DB variables, calls, and assertions.

• db_assertion_map(project_name, query, limit) indexes DB-related assertions.

• db_impact_preview(project_name, changed_entities, limit) previews impacted tests from changed DB entities.

Supported DB detection includes PostgreSQL, MySQL, MariaDB, Oracle, SQL Server, SQLite, DB2, H2, Redshift, Snowflake, ClickHouse, MongoDB, Redis, DynamoDB, Cassandra/CQL, Elasticsearch/OpenSearch, Neo4j/Cypher, and generic SQL.

Example:

db_impact_preview(
  project_name="karate-core",
  changed_entities=["orders", "postgres", "jdbc:postgresql"],
  limit=30
)

Failure Debugging

• get_failure_hotspots(project_name) finds components contributing most to failures.

• top_hotspots(project_name, limit) returns top failure hotspots.

• prioritize_fix_queue(project_name, limit) ranks fixes by failure impact and risk.

• flaky_risk(project_name, limit) finds test cases with mixed pass/fail history.

• search_error_pattern(project_name, pattern, limit) searches failed nodes by error text or fingerprint.

• get_failure_history(project_name, node_id) returns execution trend and flaky score.

• get_failure_debug_context(project_name, node_id, error_message, radius, max_historical) builds an AI debug pack.

• record_fix(project_name, node_id, error_message, solution, description) records a known fix.

• get_fix_suggestions(project_name, node_id, error_message) finds historical fix suggestions.

• auto_fix_hint_pack(project_name, node_id, error_message, max_historical) builds a step-by-step fix checklist.

Typical Workflows

Find Existing Helper Before Coding

search_reusable_function(
  project_name="karate-core",
  query="random string random number uuid",
  language="all",
  limit=20
)

Use the highest-score helper when the tags, aliases, examples, and stability score match your need.

Check What A File Change Impacts

test_selection_suggestion(
  project_name="karate-core",
  changed_paths=["src/test/java/helpers/RandomUtil.java"],
  limit=20
)

Use this before running regression to choose a smaller high-signal test set.

Understand A Feature Before Editing

ai_feature_context_pack(
  project_name="karate-core",
  feature_path="checkout.feature",
  max_call_depth=3,
  limit=10
)

This gives the AI intent, variable flow, assertions, call/read chain, and graph context.

Find Duplicate Feature Steps

feature_reuse_advisor(
  project_name="karate-core",
  min_group_size=2,
  min_flow_length=3,
  limit=50,
  include_low_signal=false
)

Use this to find duplicate steps or repeated flows that can be extracted into reusable feature components.

Debug Failed Regression

process_reports_folder(
  project_name="karate-core",
  directory_path="E:/Project/auto/karate-fw/karate-core/target/karate-reports"
)

Then ask:

top_hotspots(project_name="karate-core", limit=10)
prioritize_fix_queue(project_name="karate-core", limit=10)
get_failure_debug_context(project_name="karate-core", node_id="<failed-node-id>")

Troubleshooting

where karate-graph-mcp Cannot Find The Command

The package may be installed correctly, but the Python Scripts folder is missing from PATH.

Use this instead:

python -m karate_graph_analyzer.mcp_server --help

And configure MCP with:

[mcp_servers.karate-graph]
command = "C:\\Program Files\\Python311\\python.exe"
args = ["-m", "karate_graph_analyzer.mcp_server"]
env = { PYTHONIOENCODING = "utf-8" }

MCP Command Looks Like It Is Hanging

That is normal when running the MCP server without a client:

python -m karate_graph_analyzer.mcp_server

The server is waiting for MCP stdio messages from the AI client.

Do Not Patch site-packages Manually

Avoid editing installed files under paths like:

C:/Users/<user>/AppData/Roaming/Python/Python311/site-packages/karate_graph_analyzer

If there is a real package bug, fix the source repo, bump the package version, publish a new version, and upgrade with:

pip install -U karate-graph==<new-version>

Manual patches are overwritten by pip upgrades and do not help other machines.

Development

Install local development dependencies:

git clone <repository-url>
cd karate-graph
pip install -e ".[dev]"

Run tests:

pytest

Build and validate package:

python -m pip install build twine
python -m build
python -m twine check dist/*

Publish to TestPyPI before PyPI:

python -m twine upload --repository testpypi dist/*
python -m twine upload dist/*

Use a PyPI API token with TWINE_USERNAME=__token__.

Project Structure

karate-graph/
|-- src/karate_graph_analyzer/
|   |-- mcp_server.py
|   |-- mcp_interface/
|   |-- parser/
|   |-- graph/
|   |-- services/
|   |-- utils/
|   `-- visualization/
|-- tests/
`-- pyproject.toml

License

MIT