OrangePro

What the OrangePro MCP server does

The OrangePro MCP server examines a local repository and builds a map of its observable behaviors, code relationships, flows, and test evidence. It assigns each behavior an evidence tier, ranging from dynamically proven to having no signal, so a passing-looking test suite is not treated as proof by default.

The result is written to the repository’s .orangepro directory. The generated HTML report includes a system map, priority gaps, evidence tiers, and changes since the previous run. It also writes a deterministic graph, a Markdown coverage report, and generated-test output when applicable. Source files are not modified by generated tests; drafts are placed under orangepro_generated.

How it works

Analysis walks the source tree and creates or refreshes the evidence graph. This phase does not require a model key. OrangePro then calculates a graph-readiness score from 0 to 100 and compares behaviors with available test evidence.

Dynamic proof uses a mutation-kill oracle: a targeted behavior change is introduced, and a real test must fail for the behavior to receive the Dynamically Proven tier. Runtime coverage and structural links are recorded separately. AI-generated output does not itself upgrade evidence; only the mutation result can do that.

Static mapping works across TypeScript, JavaScript, Python, Go, Java, Kotlin, Rust, PHP, C#, Ruby, Swift, C, and C++. Generated tests and dynamic proof are documented for TypeScript/JavaScript, Python, Go, and Java, using Jest, Vitest, Mocha, pytest, Go test files, or JUnit as applicable.

Setup and configuration

Install the repository’s own dependencies first, then run the package through npx:

npx -y @orangepro/mcp-server@latest start .

Open .orangepro/behavior-coverage.html to inspect the report. No API key is required for analysis, scoring, or proof. Test generation can use a provider key supplied through ANTHROPIC_API_KEY, OPENAI_API_KEY, or an Ollama endpoint specified with OLLAMA_BASE_URL.

To expose the OrangePro MCP server to an MCP client, configure a server entry that runs npx -y @orangepro/mcp-server@latest mcp. Claude Code, Cursor, VS Code, Codex, Windsurf, and other MCP-compatible agents can then call the tools. A typical workflow starts analysis, generates tests for selected gaps, runs those tests, and invokes proof.

Tools and capabilities

The tool set supports several stages of test-evidence work:

• Start or inspect a workspace with orangepro_start, orangepro_status, and orangepro_doctor.
• Build, refresh, or incrementally update evidence with orangepro_analyze_sources and orangepro_update_graph.
• Find and prioritize weak behaviors with orangepro_find_test_gaps, orangepro_graph_score, orangepro_rtm, and orangepro_stats.
• Generate, explain, and validate tests with orangepro_generate_tests, orangepro_explain_test, orangepro_prove, and orangepro_prove_loop.
• Inspect repository changes with orangepro_changed_impact, record test results with orangepro_record_run, and export metadata with orangepro_export_evidence_pack.
• Request optional AI suggestions through orangepro_ai_links and orangepro_ai_flows.

Limitations and notes

Dynamic proof has narrower language and runner support than static mapping. Kotlin, Rust, PHP, C#, Ruby, Swift, C, and C++ can be mapped statically, but generated tests and dynamic proof are listed as planned for those languages.

The OrangePro MCP server does not treat generated tests as evidence by themselves. Proof requires targeted mutation testing, and an initial run may show zero dynamically proven behaviors. Coverage artifacts can be recorded with suite metadata; without that metadata, the tool conservatively labels unclear coverage as unclassified rather than assuming it is unit or integration coverage.