Ship better Java with your coding agent.
JAIPilot helps Codex and Claude Code clean, test, review, modernize, and optimize real Java repositories without drifting into unnecessary code or unproved changes.
By default, your agent runs Java builds, tests, analysis, profiling, and benchmarks on a ready remote Java machine whenever the task does not require laptop-only access or state.
JAIPilot does not replace your coding agent or add another AI. It gives your agent focused Java workflows, remote compute, and one rule: show evidence, not confidence.
codex plugin marketplace add JAIPilot/jaipilot
codex plugin add jaipilot@jaipilot/plugin marketplace add JAIPilot/jaipilot
/plugin install jaipilot@jaipilot
Open a Java repository and ask:
Make my current Java changes production-ready without changing their behavior. Add missing tests,
remove unnecessary code, improve only performance you can measure, and run the full verification.
Use JAIPilot Remote for substantial Java commands unless this work needs my laptop.
The Java skills work locally as soon as the plugin is installed. When remote execution is useful,
your agent asks before uploading the current tracked and unignored Git files. Approve the upload and
sign in when prompted; the agent handles packaging, integrity checking, upload, execution, logs, and
workspace deletion. You do not create an archive, configure a VM, provide an API key, or copy files
manually. .git, ignored files, and remote edits are never transferred back automatically.
If packaging or upload cannot be verified, JAIPilot does not create the workspace. Your agent must show the failing step instead of silently uploading a different source tree.
| 12.2–80.3% faster | 61.3–62.5% faster | 87.5–92.4% faster | 8 → 2 SQL statements |
|---|---|---|---|
| OTel lookup medians | Micrometer merges | Calcite JMH medians | N+1 removed |
The original Petclinic PR already had a green build. JAIPilot reviewed that exact head and produced this companion change:
| Metric | Without JAIPilot | With JAIPilot | Outcome |
|---|---|---|---|
| Tests | 75 | 85 | +10 tests (+13.3%) |
| Changed-method line coverage | 0/12 (0%) | 11/11 (100%) | +100 percentage points |
| Changed-method branch coverage | 0/8 (0%) | 8/8 (100%) | +100 percentage points |
Owner class line coverage | 22/53 (41.5%) | 33/51 (64.7%) | +23.2 points with less code |
| Production change | Unused helper remained | +2/-9 lines | 7 net lines removed |
Owner methods | 16 | 15 | 1 unused method removed |
Owner complexity | 25 | 24 | 4% lower |
| Clean Maven verification | 75/75 passed | 85/85 passed | Both stayed green |
The important result is not simply “more tests.” Without JAIPilot, the build passed while the new behavior had zero coverage and unused code remained. With JAIPilot, the same behavior stayed green, the edge cases became executable tests, and production code became smaller.
The comparison uses the original PR head and JAIPilot's direct child commit, clean worktrees, the
same ./mvnw -q clean verify command, and fresh JaCoCo 0.8.14 reports.
On skrcode/calcite at exact commit
d3a5d8d,
JAIPilot profiled Calcite's existing
DefaultDirectedGraphBenchmark.removeAllVertices{10,50,90}Benchmark. The 50% workload attributed
31.1% of runnable samples to Collection.removeIf: the implementation scanned the complete global
edge set once for every removed vertex.
The candidate changed two files (+28/-5), removed the repeated scans, and added behavior tests for the majority-removal and self-loop paths. Lower JMH scores are better:
| Removed vertices | Baseline median (µs/op) | JAIPilot median (µs/op) | Improvement | Baseline p95 (µs/op) | JAIPilot p95 (µs/op) | Improvement |
|---|---|---|---|---|---|---|
| 10% | 26.710 | 2.029 | 92.4% | 27.142 | 2.439 | 91.0% |
| 50% | 74.619 | 9.140 | 87.8% | 87.245 | 14.993 | 82.8% |
| 90% | 77.423 | 9.677 | 87.5% | 89.514 | 10.052 | 88.8% |
Baseline and candidate ran on the same 4 CPU/8 GiB remote workspace with the same Temurin JDK 17,
built JMH jar, command, and workload. Each row contains 21 measured observations: seven forks with
three measured iterations per fork after warm-up. The identical focused command passed 15/15 tests
before and after the production edit. A fresh exact-SHA :core:clean :core:check then completed
16,644 tests with 0 failures and 155 skips, and the tested remote diff matched the local candidate
digest.
This is a controlled result for Calcite's existing graph-removal workloads, not a claim that every Java workload becomes faster.
On skrcode/opentelemetry-java at exact commit
35636ae,
JAIPilot found that immutable attribute sets were sorted by key name during construction but still
used a full linear scan for every lookup. This matters at the default span limit of 128 attributes.
The three-file draft change preserves the small-set and first-four-entry fast path, then uses binary search for the rest. It also adds a large-set behavior test and a repository-native JMH benchmark. Lower values are better:
| Lookup | Baseline median (ns/op) | JAIPilot median (ns/op) | Improvement | Baseline p95 | JAIPilot p95 | Improvement |
|---|---|---|---|---|---|---|
| First | 2.483 | 2.179 | 12.2% | 2.637 | 2.272 | 13.8% |
| Middle | 169.124 | 85.905 | 49.2% | 178.250 | 88.537 | 50.3% |
| Last | 346.195 | 87.323 | 74.8% | 358.321 | 90.084 | 74.9% |
| Missing | 141.560 | 69.552 | 50.9% | 150.321 | 74.261 | 50.6% |
Last as Value | 368.629 | 72.684 | 80.3% | 387.467 | 78.087 | 79.8% |
Baseline and candidate ran in the same 4 CPU/8 GiB remote workspace with Temurin JDK 21, the same
JMH jar, command, warm-up, and workload. Each row has 21 observations. The new focused behavior test
passed before and after the production edit; a clean :api:all:check passed all 147 tasks including
Animal Sniffer, Checkstyle, Spotless, tests, and japicmp. The tested remote Git delta matched the
local candidate digest. The Value workload still allocates about 16 B/op; JAIPilot reports the
lookup-time win without claiming that allocation disappeared.
On skrcode/micrometer at exact commit
22207bf,
JAIPilot found that adding or replacing one Tag or KeyValue went through temporary varargs and
iterable merge machinery even though the backing arrays were already sorted.
The six-file draft change adds a bounded binary-search merge for the single-value overloads, behavior tests, and four workloads in Micrometer's existing JMH module:
| Replacement workload | Baseline median (ns/op) | JAIPilot median (ns/op) | Improvement | Baseline p95 | JAIPilot p95 | Improvement | Allocation |
|---|---|---|---|---|---|---|---|
KeyValues.and | 57.941 | 22.449 | 61.3% | 63.554 | 23.487 | 63.0% | 136 → 104 B/op |
Tags.and | 58.968 | 22.101 | 62.5% | 63.388 | 24.425 | 61.5% | 136 → 104 B/op |
Single-value insertion reduced median allocation by 17.6% and p95 allocation by 46.2%. Its median latency improved by only 6.7–8.9%, below JAIPilot's 10% shared-hardware threshold, so it is not presented as a speed win. The same-workspace experiment used 21 observations per workload; 92 focused tests and the final clean scoped build passed, with 1,132 tests, zero failures, and the remote production diff matching the local digest.
On the current Petclinic JDBC vet listing, JAIPilot found a real N+1 query path: six vets required
eight SQL statements (2 + N). It kept the ordered vet query and replaced the per-vet specialty
lookups with one joined association query.
| Evidence | Baseline | JAIPilot candidate |
|---|---|---|
| SQL statements, six vets | 8 | 2 |
| Growth with vet count | 2 + N | constant 2 |
| Focused JDBC tests | 11/11 | 15/15 |
| Clean repository build | 75/75 | 79/79 |
The new tests lock vet ordering, specialty ordering, vets without specialties, duplicate links, shared specialty identity, empty data, and the exact two-statement ceiling. The final clean Maven build and JaCoCo report passed on a 4 CPU/8 GiB remote workspace, and the local and remote binary diff digests matched. This is deterministic query-count evidence; JAIPilot does not turn it into an invented latency claim.
- Less agent drift — changes stay bounded, lean, and aligned with the repository.
- Better Java code — remove proven waste, reduce complexity, improve tests, review risky diffs, modernize safely, and optimize measured bottlenecks.
- Real verification — use the repository's Maven or Gradle build, tests, coverage, architecture rules, and performance measurements.
- Remote-first execution — use disposable hardware with JDK 17, 21, and 25, Maven, and Gradle ready unless private networks, local services, secrets, hardware, or state require the laptop.
- Remote performance lab — profile and compare a bounded optimization on one 4 CPU/8 GiB workspace, with matching patch identity, raw observations, median, p95, and correctness evidence.
- Works on your current change — staged, unstaged, and untracked files can be tested without committing or pushing first.
- Your agent stays in control — Codex or Claude Code chooses every edit and command and reports exactly what JAIPilot achieved.
Additional acceptance runs used repository-native verification:
| Use case | Result |
|---|---|
| Optimize Apache Calcite graph removal | Existing JMH workload medians improved by 87.5–92.4% across 10%, 50%, and 90% removal cases; matching behavior tests and a 16,644-test clean check passed. |
| Optimize OpenTelemetry attribute lookup | Five 128-attribute lookup workloads improved by 12.2–80.3% at median and 13.8–79.8% at p95; clean API compatibility and verification passed. |
| Optimize Micrometer single-value merges | Replacement merges improved by 61.3–62.5% at median and allocated 23.5% less; sub-threshold insertion latency was not marketed as a speed win. |
| Cover previously untested behavior | 7 focused tests added with no production or dependency change; target coverage moved from 0% to 100% for lines and branches; 82/82 tests passed independently on Java 17 and 21. |
| Run the current change remotely | An uncommitted file reached the workspace; ./mvnw clean test passed 75/75 tests in 44.6 seconds; job recovery, cancellation, and workspace deletion were verified. |
| Remove a JDBC N+1 query | Vet listing SQL statements fell from 8 to 2 on the six-vet fixture; 15/15 focused and 79/79 clean-build tests passed, with identical local/remote diff digests. |
These are reproducible acceptance results, not claims that every repository will see the same coverage, code reduction, or speed.
Clients that consume the official MCP Registry can discover the hosted remote tools as
io.github.JAIPilot/jaipilot. This installs remote execution only; install the Codex or Claude Code
plugin above for the six Java engineering skills as well.
| Skill | Outcome |
|---|---|
jaipilot-optimize-java | Make one bounded Java change leaner, safer, and faster. |
jaipilot-generate-tests | Add meaningful tests and fresh coverage evidence. |
jaipilot-clean-java | Remove waste, enforce architecture, and simplify code. |
jaipilot-review-diff | Find regressions, unnecessary code, and missing proof. |
jaipilot-fast-execution | Run substantial Java verification efficiently. |
jaipilot-remote-java | Default applicable Java execution to remote hardware. |
JAIPilot can work with repository-configured tools such as JaCoCo, PIT, ArchUnit, OpenRewrite, Checkstyle, PMD, SpotBugs, SonarQube, JMH, and JFR. It never weakens a quality gate merely to get a green result.
The beta permits one active remote workspace. Remote work is disposable and never commits, pushes, or publishes code. JAIPilot defaults applicable Java execution to remote hardware; it stays local when a corporate VPN, private artifact service, internal database, unavailable secret, machine-specific state, or another laptop-only resource is required.
See Security, Privacy, Support, Terms, and the Changelog.
Licensed under the MIT License.