Find test gaps, generate grounded tests, and dynamically prove behavior with mutation testing.
io.github.OrangeproAI/orangepro Model Context Protocol (MCP) Server
The io.github.OrangeproAI/orangepro MCP server helps identify test gaps, generate grounded tests, and dynamically prove behavior using mutation testing. Its focus is broader testing coverage via semantic and graph/ontology-based analysis, including support for static analysis and test automation.
π οΈ Key Features
Find test gaps
Generate grounded tests
Dynamically prove behavior with mutation testing
Uses/relates to semantic-analysis and graph-based-testing
Topics include ontology and knowledge-graph
π Use Cases
Integration and end-to-end testing coverage improvement
OrangePro maps every public function in your repository, links each one to the tests that actually call it, and ranks what's left by what it can break. For a test you care about, it breaks the function in an isolated copy and checks that the test fails. It runs on your machine, uses no AI model for scoring, and sends nothing to OrangePro. Mapping, gap ranking and proof work with no model key. Drafting a test with real assertions needs a model provider; with no key, provider: deterministic returns placeholders only (see Model setup under Reference).
Every run writes two reports and a spreadsheet of tests to .orangepro/:
File
For
What's in it
short_behavior-coverage.html
Leads, managers, anyone in a hurry
One page: what to do with the 20 highest-priority functions, the top five with the reason and the tests drafted for each, what changed since the last report, and how to reproduce the run.
behavior-coverage.html
Developers
Worklist: every ranked function with what to do, filters, and a side panel with the evidence, why it ranks, two drafted tests and the proof command. Map: where tests reach, by area. Run details: inputs, counts, proofs, fingerprints and settings.
tests.csv
QA, test managers
One row per test slot (preconditions, steps, test data, expected result, priority, folder), ready for Excel, Google Sheets, TestRail, Jira Xray or Zephyr Scale. Export tests (CSV) in the detailed report downloads it, and its Run details show how the columns map (Importing tests.csv).
Every function gets one of four things to do:
What it means
What to do
Done
A test fails when this code breaks.
Nothing. It moves back to Prove if the code changes.
Prove
A test runs it, or a coverage run executes it, but no test is proven to catch a break.
Run the proof command.
Check
A test file or test name points here, but no call was traced.
Read that test file first; prove it if it drives the function, otherwise write one.
Write
No test catches a break here yet.
Write a test (two are drafted for the top 20), then prove it.
The one-page summary
OrangePro one-page summary: what to do with the 20 highest-priority functions, and the top five with their tests
The detailed report
OrangePro detailed report: the worklist with filters and the side panel for one function
Both from OrangePro run on its own repository at commit 64a5213, without a model key, so no tests were drafted. With a key, each of the top 20 shows two drafted tests, and the side panel shows their steps and code.
Why developers use it
It tells you what a test actually checks, not just what it runs. A test that calls a function isn't necessarily checking it. opro prove replaces the function body with a fixed return value in an isolated copy and reruns your own test, unchanged. If the test still passes, it wasn't protecting that function.
It puts consequences first. One filter shows the code paths that reach a delete or purge (cache evictions don't count); others show code that changes often, has many callers or runs on a schedule.
It's honest about evidence. A function is "linked" only when a test calls that exact function. A test with a similar name is a lead, never coverage. Tests that only check mocks don't count.
It's repeatable. Same commit, full git history, same config and same version give the same ranking. Each report records fingerprints, so you can tell when a change in results came from the code and when it came from the tool.
Your coding agent can drive it. As an MCP server it gives Claude Code, Cursor, Copilot, Codex and others the ranked gaps, the suggested test location and the proof step in one loop.
It's free, local and open source (MIT). No account, no API key for analysis, and no upload.
Quick start
bash
cd /path/to/your/repo
# install the repo's own dependencies first (npm ci, uv sync, go mod download, ...)
npx -y @orangepro/mcp-server@latest start .
open .orangepro/short_behavior-coverage.html
Mapping, gap ranking and proof work with no model key. Drafting a test with real assertions needs a model provider (see Model setup); deterministic mode returns placeholders only.
Tips for the most accurate run:
Use a full clone, not a shallow one. Change history drives part of the ranking, and the report says when history was partial.
Exclude what isn't product code with rank_exclude_paths (see Configuration).
Rerun after a change. Each report says what changed since the last one. When nothing changed (same commit, files, settings, coverage and proofs), start reuses the last reports at once and says so; --fresh runs in full anyway.
What's written:
code
.orangepro/
βββ short_behavior-coverage.html β one-page summary
βββ behavior-coverage.html β detailed report: worklist, map, run details
βββ tests.csv β drafted tests, one row per test
βββ report-data.json β the data both reports show (used to compare and reuse runs)
βββ run-key.json, last-start.json β what the last run saw, so an unchanged repo is not analysed again
βββ graph.json β the evidence graph (deterministic)
βββ ledger.json β proof certificates
βββ rtm.md β traceability matrix
βββ config.json β optional per-repo settings
orangepro_generated/ β generated tests (real assertions need a model provider; deterministic mode returns placeholders only); your files are never edited
Evidence tiers
Every function gets exactly one evidence label. The four things to do above are read from it.
Evidence
What it means
To do
Proven
A test passed on the original code and failed at its own assertion when this function was broken in an isolated copy.
Done
Runs under tests
A coverage tool you ran executed this code.
Prove
Linked
A test calls this exact function. That's a structural link, not proof.
Prove
Unconfirmed
A test file or a test with a similar name points here, but no call was traced. That's a lead, not evidence.
Check
No test found
Nothing links a test to it.
Write
A proof that ran and did not catch the break turns the function into Write; a proof made before the code changed turns it back into Prove.
After opro export-tests, a function can also show a patch status: Verified test patch (not merged) means a new test, in orangepro-tests.patch, catches a fault the existing tests missed. It is not an evidence tier. The row keeps its label and what to do until the merged test is proven.
A successful orangepro_prove sets the oracle field proven to true and writes the ledger status reproven. That status is used the first time too. The report's Proven count is how many behaviors still match one of those certificates. The traceability row says Reproven (this run). That label does not mean the function was proved earlier.
What OrangePro doesn't count, so linked numbers are a floor, not a coverage percentage:
tests that only assert on mocks;
calls made over HTTP or from end-to-end suites;
in Python, objects that reach a test only through a fixture;
in JavaScript and TypeScript, calls on an object a test gets from a fixture (for example Playwright's test.extend), and calls to an inherited method, which are not linked to the class that defines it.
Check the tests before writing new ones for a flagged path.
"Proven 0" is normal on a first run. Proof runs your tests, so it happens only for the functions you choose, or within the attempt budget of opro start.
How the ranking works
Each unproven function gets an OrangePro Risk Score, ORS = P Γ I Γ D:
P: how likely it is to change. Change history, fan-out, new code and size.
I: what it can break. Incoming references, entry-point position, data sensitivity, and a floor for paths that reach a delete or purge within two calls.
D: how hard a break would be to notice. Evidence tier (a linked test or a coverage run makes a break easier to notice than a name match or nothing), and whether it runs unattended (jobs and schedulers).
The score sets the order of work. It is not a defect probability. The weights are fixed and no AI model is involved. The report shows the inputs behind every row.
Prove a test
bash
# Python: the mutation value is derived from the function (return annotation or its observed result)
opro prove-loop --target-symbol 'sym:app/billing/invoices.py#void_invoice' \
--test'tests/test_invoices.py::test_void_marks_invoice_void' --replacement sentinel
# TypeScript / JavaScript: give the inert body to substitute
opro prove-loop --target-symbol 'sym:src/orders.service.ts#OrdersService.cancel' \
--test src/orders.service.spec.ts --replacement 'return null;'
Proven: the test passes on the original code and fails at its own assertion on the broken copy. The oracle then says proven: true. The ledger stores that attempt as reproven, including the first time, and the traceability row says Reproven (this run).
Not proven: the test still passes on the broken copy, so it doesn't protect that function. That's a finding too.
Unrunnable: setup failed. This is never counted either way.
Arguments for orangepro_prove
For a JavaScript or TypeScript function the call needs:
target_symbol: sym:<file>#<Symbol>. A bare name is refused.
test_path: the test file, relative to the repo.
replacement: empty, or a single return of a literal, such as return null;. A throw, a call, or any other statement is refused.
replacement_mode: return-json for a normal function (the default). promise-json when the function is async, so the literal is returned as a resolved promise.
runner: jest, vitest, mocha, or auto. Use the runner the repo uses.
Worked example for an async run in index.js, tested with Jest:
npx -y @orangepro/mcp-server@latest agent --client codex prints the setup
A prompt that runs the whole loop:
"Use orangepro_start, then orangepro_find_test_gaps. For the top gap, write a test at the suggested path, run it, then call orangepro_prove_loop and tell me whether it was proven."
Configuration
Optional. Put it in .orangepro/config.json in the repository, or in ~/.orangepro/config.json for defaults across repositories. Every setting that changes the ranking is shown in the report.
json
{"classification":{"rank_exclude_paths":["ui/**","docs/**","scripts/**"],"test_support_paths":["internal/testutil/**"],"destructive_sinks":["archive*"]},"tuning":{"churn_window_days":180},"proof":{"python_runner":"auto","attempt_limit":20},"overrides":[{"symbol":"sym:src/legacy/shim.ts#shim","action":"suppress","reason":"generated shim, not product code"}]}
rank_exclude_paths removes paths from the ranking. They are still counted as behaviors.
destructive_sinks adds delete-like calls for your codebase.
Overrides (suppress, pin, reclassify) each require a reason, and the report lists them.
Weights and tiers can't be configured.
Language support
Language
Map functions
Link tests that call them
Draft tests
Mutation proof
TypeScript / JavaScript
β
β
β Jest, Vitest, Mocha
β Jest, Vitest, Mocha
Python
β
β
β pytest
β pytest
Go
β
β
β *_test.go
β go test
Java
β
β
β JUnit 4 and 5
β Maven with JUnit 5 only
Kotlin, Rust, PHP, C#, Ruby, Swift, C, C++
β
name match only
not yet
not yet
"Name match only" means a test file or test name that looks like the function makes it Unconfirmed (Check). No call is traced, so these languages never reach Linked or Proven today.
TypeScript and JavaScript are parsed with the TypeScript compiler; the other languages with tree-sitter.
Python proof runs pytest through the repository's own environment: an existing .venv or venv, then uv or poetry when their lock files are present, else python -m pytest. Set proof.python_runner to use your own command.
Java proof needs Maven (Surefire) and JUnit 5. Gradle and JUnit 4 projects are mapped and linked, but not proven yet.
Not traced yet, so tested code can show as Unconfirmed: calls through Go test-suite methods and helper constructors, dependency injection that is wired at runtime, reflection, and calls made over HTTP or RPC.
Requirements and limits
A patched Node.js: 22.23.4 or newer 22.x, 24.22.1 or newer 24.x, or 26.12.1 or newer, on macOS or Linux. Windows isn't tested: proofs use POSIX paths and symlinks.
Use a patched Node.js. The October 21, 2026 Node.js security releases (22.23.4, 24.22.1, 26.12.1) fix several high-severity issues, and Node.js 20 is end-of-life and gets no fix. Odd-numbered lines (23, 25) have no patched release. The engines field in package.json records this, but npm only warns about it unless you set engine-strict=true. Check your version with node -v. See the Node.js security release notice.
Git, for change history. Without it, or in a shallow clone, the ranking uses structure only and the report says so. A folder that isn't a git checkout always runs in full.
Install the repository's dependencies first (npm ci, uv sync, go mod download, mvn -q install, β¦). Proof runs your own tests in an isolated copy and needs them to pass there.
What's read: source and test files, skipping anything in .gitignore or .orangeproignore (negations with ! aren't supported), the usual build and dependency folders (node_modules, vendor, dist, build, target, coverage, .venv, β¦), files over 1 MB, and generated code (files marked Code generated β¦ DO NOT EDIT). Up to 100,000 files; ORANGEPRO_MAX_FILES lowers that and ORANGEPRO_MAX_ANALYZE_MS sets a time limit for the scan.
Test drafting covers the 20 highest-priority functions, two tests each, in the order the worklist shows, and the reports show those 20. Drafts are kept until the function's code changes.
What counts as "nothing changed" for reusing a run: the commit, every modified or untracked file that git doesn't ignore, OrangePro's settings and ignore files, coverage files, the proof ledger, the start options, whether a model key is set, and the analysis version. A run whose test drafting failed on a provider error (out of credits, rate limit) runs again in full.
The CSV column mapping follows each tool's CSV importer (see Importing tests.csv). Check the import preview the first time.
Privacy and network use
Nothing about your code or your runs is sent anywhere. There is no usage telemetry.
Source is read in-process and never stored or uploaded. Reports and the evidence graph contain metadata, not your source.
Your source files are never edited. Proofs run in an isolated copy.
Model calls happen only if you configure a key, and they go directly from your machine to the provider you chose. Keys, the provider, its endpoint and the model are read only from your shell environment, or from the env block of your MCP client's config. OrangePro reads no .env file, so a repository you scan cannot choose where your key is sent, and OrangePro never writes keys anywhere. opro setup saves only the provider and model name.
With a key, the reports and tests.csv include the tests your model drafted, and the prompts include the function names, files and evidence they're based on.
Reports load nothing from the network. The only outbound links are ones you click.
Feedback
Reports include a Give feedback link and a This looks wrong link on each finding. Once in a while, after the findings, they also ask whether the report helped.
The links carry nothing about your project. The form sends only what you review and submit.
You stay anonymous unless you leave an email.
bash
opro feedback # print the link and your settings
opro feedback off # stop the "Did this help?" question (links stay)
ORANGEPRO_FEEDBACK_URL=off # hide every feedback link
The question appears at most once every 14 days (ORANGEPRO_FEEDBACK_COOLDOWN_DAYS). That preference is stored only on your machine. MCP clients get the link as optional result metadata (_meta["ai.orangepro/feedback_url"]), and your agent is never asked to prompt you for feedback.
Reference
CLI
bash
opro # same as opro start .
opro start . --no-ai --no-auto # analyze + reports, no model calls, no proof attempts
opro start . --fresh # run in full even when nothing changed since the last run
opro start --base main # scope to a branch diff
opro analyze # build the evidence graph and reports
opro gaps --limit 10 # top unproven behaviors, ranked
opro prove-loop ... # mutation proof for one function (see above)
opro roast . # find tests whose mutant survives (no key needed)
opro doctor --proof # why top targets aren't proven yet
opro generate --base main # tests for this branch (needs a model provider; --provider deterministic returns placeholders only)
opro rtm # traceability matrix
opro export# metadata-only evidence pack
opro coverage # find or generate runtime coverage artifacts
opro feedback [on|off] # feedback link and invitation setting
opro mcp # run as an MCP server (stdio)
Add --json to any read command for machine output. Run opro help for every flag.
MCP tools (19)
Tool
What it does
orangepro_start
Analyze, write both reports and tests.csv, return next actions; returns at once with unchanged_since when nothing changed (fresh forces a full run)
orangepro_analyze_sources
Build or refresh the evidence graph
orangepro_find_test_gaps
Unproven behaviors ranked by ORS
orangepro_prove_loop
Setup + mutation proof + report refresh for one behavior
orangepro_prove
Mutation proof only
orangepro_generate_tests
Grounded tests. A test with real assertions needs a model provider. provider: deterministic returns placeholders only, with no key.
orangepro_changed_impact
What a diff touches
orangepro_status
Workspace state without running anything
orangepro_doctor
What evidence to add next
orangepro_graph_score
Graph readiness (0β100)
orangepro_rtm
Traceability matrix
orangepro_stats
Aggregate statistics
orangepro_record_run
Record a test run result
orangepro_explain_test
Why a test was generated
orangepro_export_evidence_pack
Metadata-only evidence pack
orangepro_update_graph
Incremental graph update
orangepro_ai_links
Weak behaviorβcode suggestions (optional AI, never evidence)
orangepro_ai_flows
Candidate flows (optional AI, never evidence)
orangepro_install_github_action
Preview, or with confirmation write, an opt-in post-merge workflow and its profile; never commits or pushes (tested locally)
Coverage artifacts
Run your own unit and integration coverage first, then opro start. OrangePro ingests Go coverprofiles, lcov, coverage.py XML and JaCoCo XML, and keeps unit, integration and unclassified coverage separate. To label artifacts, add .orangepro/coverage-suites.json:
Auto-detect order: OpenAI β Ollama β Anthropic. Override with --provider and --model, or run opro setup.
Model output never changes an evidence tier. Only the mutation proof can mark a behavior Proven.
Generated tests land in orangepro_generated/ with the files and existing tests they're grounded on, and where and how to run them.
Drafts whose code fails the compile check are kept as plans ("Plan only"), with the reason named. Without a key, each of the top 20 shows two empty slots with a prompt you can copy into your coding agent.
Without a key, orangepro_generate_tests with provider: deterministic returns placeholders only. That mode smoke-tests the pipeline.
A placeholder checks nothing (expect(true).toBe(true), assert True, a t.Fatalf that never runs), so it can never prove anything. Under Mocha its toBe line fails outright, because Chai's expect has no toBe.
It is still marked runnable: true, with a run hint (suggested_path, run_command, and a prove_run when it targets a counted JavaScript, TypeScript or Python function), whenever its checks pass:
Python and Java: its syntax check passes (Python needs python3).
Go: gofmt passes and, next to Go source, a compile check passes (never under --static-only).
JavaScript and TypeScript: OrangePro can reuse a related test file's imports (even one matched only by name, whose imports may not include the code under test) or derive an import from a source file, and every import resolves.
Otherwise it is runnable: false, with no run hint.
Importing tests.csv
One row per test, UTF-8. The first 12 columns are for import; the rest keep the priority, evidence and proof command with each test. Priority is High for the five highest-ranked functions and Medium for the rest.
Column in tests.csv
TestRail
Jira Xray
Zephyr Scale
ID
Test Case Identifier
(leave Key empty)
Title
Title
Summary
Name
Section (a > b)
Sections Hierarchy
Folder (a/b)
Test Repository Path
Folder
Priority
Priority
Priority
Priority (map Medium to Normal)
Test type
Type
Labels
Labels
Preconditions
Preconditions
(separate issues in Xray; add to Description if you want them)
Precondition
Steps
Manual Test Step: Action
Test data
Manual Test Step: Data
Steps with data
Steps
Test Script (Steps) - Step
Expected result
Expected Result
Manual Test Step: Expected Result
Test Script (Steps) - Expected Result
References
References
Description
Objective
TestRail: Import from CSV with the Test Case (Text) template, one row per test case.
Jira Xray: Test Case Importer; each test imports as one manual step.
Zephyr Scale: it has no test-data field, so Steps with data carries it.
Check the importer's preview before you confirm.
PR workflow
bash
opro generate --base main # tests for what this branch changed (read-only git diff)
opro generate --changed # current branch vs its base
opro generate --pr 1234 # checks out PR #1234 (asks first; refuses on a dirty tree)