接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令？ 兩種安裝方式的差異
來自發布者
MCP server that gives AI agents full debugger control over running Java applications — inspect state, set breakpoints, evaluate expressions, and mutate values at runtime via JDWP/JDI.
Release notes — see . 2.0.0 is a breaking release: several tools were unified, and field watchpoints are new.
MCP server giving AI agents full debugger control over running Java applications — inspect state, set breakpoints, evaluate expressions, and mutate values at runtime via JDWP/JDI
CHANGELOG.mdjdwp_clear_*Built on the foundations of mcp-jdwp-java by Nicolas Vautrin — the original project that provided core JDI connectivity, thread/stack/variable inspection, stepping, and basic breakpoint management. Everything described below as "beyond standard JDWP" was built on top of that base.
An MCP server that speaks JDWP — the same protocol IntelliJ / Eclipse / VS Code use to attach — and exposes the debugger to Claude Code as 47 tools + 2 MCP resources over STDIO. Once the agent attaches to your JVM, it can:
Field.set that a line BP can never seeThe agent gets the same power as IntelliJ's debugger, with extras that exist specifically because an agent isn't a human staring at a UI (filtering, blocking resume, one-shot context dumps, recursion guards, reconnect after a wedge).
This MCP server runs entirely on your local machine:
./mvnw clean package). No pre-built binaries are downloaded or distributed.jdk.jdi).No separate Maven install is required — the repository ships with the Maven Wrapper (./mvnw), which downloads a pinned Maven 3.9.x into ~/.m2/wrapper/ on first use. The SessionStart hook and every build command in this README use the wrapper.
Option A: Plugin marketplace (recommended)
Installs the MCP server, the java-debug skill (debugging workflows, recipes, gotchas), and the .mcp.json configuration in one step:
/plugin marketplace add https://github.com/FgForrest/mcp-jdwp-java.git
/plugin install jdwp-debugging@mcp-jdwp-java
The server JAR is built automatically on first session start via the bundled ./mvnw wrapper (requires JDK 17+ on PATH — no Maven install needed). The hook is content-aware: it rebuilds when git HEAD moves, not just when the JAR is missing, so git pull + restart picks up updates without manual intervention. Restart Claude Code to pick up the plugin.
If you prefer to register the MCP server directly without the plugin (no skill, no auto-build):
1. Build the JAR:
git clone https://github.com/FgForrest/mcp-jdwp-java.git
cd mcp-jdwp-java
./mvnw clean package -DskipTests # use mvnw.cmd on Windows
2. Register with Claude Code:
claude mcp add jdwp-inspector -s user \
-e MCP_TIMEOUT=30000 \
-e MCP_TOOL_TIMEOUT=120000 \
-- java --add-modules jdk.jdi,jdk.attach -jar /path/to/mcp-jdwp-java.jar
To change the JDWP port (default 5005), add -DJVM_JDWP_PORT=12345 before -jar.
The MCP_TIMEOUT and MCP_TOOL_TIMEOUT environment variables are important — JVM startup is not instant (class loading, Spring context initialization), so the default MCP timeouts will cause Claude Code to give up before the server is ready. MCP_TIMEOUT=30000 gives the server 30 seconds to start, and MCP_TOOL_TIMEOUT=120000 allows up to 2 minutes for long-running tools like first-time expression evaluation (which discovers the target's classpath and compiles bytecode).
Re-installing requires removing first: claude mcp remove jdwp-inspector -s user
Drop -s user to scope to the current project only.
.mcp.json:
{
"mcpServers": {
"jdwp-inspector": {
"command": "java",
"args": [
"--add-modules", "jdk.jdi,jdk.attach",
"-jar", "/path/to/mcp-jdwp-java.jar"
],
"env": {
"MCP_TIMEOUT": "30000",
"MCP_TOOL_TIMEOUT": "120000"
}
}
}
}
The plugin-managed install also passes -DLOG_PATH=${CLAUDE_PLUGIN_ROOT}/logs/mcp-jdwp-inspector.log so the server's SLF4J log lands inside the plugin directory; add the same flag if you want a fixed log path here.
Maven Surefire (test debugging):
mvn test -Dmaven.surefire.debug
Starts the JVM with JDWP on port 5005, suspended until a debugger connects.
Any Java application:
-agentlib:jdwp=transport=dt_socket,server=y,suspend=n,address=*:5005
From there, just tell the agent what to do — "attach to JDWP, set a breakpoint in OrderService.createOrder line 42, run the failing test, dump the context when it hits". The bundled java-debug skill teaches the agent the right tool sequence; the next section gives you a smoke test that exercises the surface.
The jdwp-sandbox module ships 9 deliberately broken Java classes. Each one compiles fine, looks reasonable at first glance, and fails its test with a confusing message. Your job: attach with the JDWP MCP server and find the root cause.
This doubles as a setup verification — if you can solve these, everything works.
Each flight is built so that one tool group is the path of least resistance, and lists a par — the minimum tool calls that cleanly reveal the root cause. Hitting par is the elegant solve; the suite as a whole exercises the full tool surface (expression eval, exception breakpoints, field watchpoints, event history, marks, logpoints, runtime mutation, and multi-thread inspection).
The fastest path for fresh plugin users is a self-contained zip — no clone, no reactor build:
curl -L -o jdwp-sandbox.zip \
https://github.com/FgForrest/mcp-jdwp-java/releases/latest/download/jdwp-sandbox.zip
unzip jdwp-sandbox.zip && cd jdwp-sandbox
Inside, you get a parentless Maven project with src/, a pom.xml, a flight-game README.md, and a CLAUDE.md that briefs the agent on the game's house rules. Open the folder in Claude Code and ask it to play flight #1 — the bundled CLAUDE.md keeps it honest (no peeking at source, no spoiler-fetching). Continue with the workflow below.
Start Claude Code from the sandbox folder (or this repo's root, if you cloned) and type:
Use JDWP to debug <TestClass> in the jdwp-sandbox module — the test is failing, find the root cause.
Difficulty: Moderate | Test: SessionStoreTest | Package: session | Par: 4 | Exercises: expression eval
Symptom: retrieve() returned null — a session was stored, upgraded, and then... vanished from the map.
Hint: The session is still in the HashMap. The HashMap just can't find it anymore.
UserSession is used as a HashMap key, with both userId and role in hashCode(). upgradeUserRole() calls session.upgradeRole("PREMIUM"), which mutates role — changing the hashCode while the key is still in the map. The entry sits in the old hash bucket; lookups compute the new hash and search the wrong bucket.
Debug path: Breakpoint before and after upgradeRole(). Use jdwp_assert_expression("session.hashCode()", "<value-before>") after the upgrade — MISMATCH confirms the hash drifted.
Difficulty: Hard | Test: EventBusTest | Package: events | Par: 4 | Exercises: exception breakpoint + trigger gate
Symptom: expected stock < 100 but was 100 and an empty error summary — the order was supposed to reserve inventory, but nothing happened and nobody complained.
Hint: The handler runs on a background thread. Nothing in your code catches its failure — so there's no catch block to break on. Catch the throw itself.
OrderEvent narrows the raw quantity through byte (200 → -56), so Inventory.reserve() throws IllegalStateException. EventBus.dispatch() runs each handler on a single-thread executor as a fire-and-forget task — the Future is never inspected, so the exception is captured inside java.util.concurrent.FutureTask.run (a JDK frame) and discarded. No sandbox frame ever holds the throwable, so a breakpoint-context dump has nothing to show, and getErrorSummary() stays empty.
Debug path: Because the exception is caught (by FutureTask), an uncaught-only breakpoint never fires. Set a line BP at the test/dispatch entry as a trigger, then jdwp_set_exception_breakpoint("java.lang.IllegalStateException", caught=true, triggerBreakpointId=<trigger>) so bootstrap exceptions don't drown the signal. The throw site suspends with qty = -56 in the frame's locals.
Difficulty: Hard | Test: ConfigurationProviderTest | Package: config | Par: 4 | Exercises: field watchpoint + event history
Symptom: expected timeout=5000 but was 0 — the timeout was set during construction, yet it reads back as the default.
Hint: The value was set correctly. Something wrote over it afterward. There's no half-built object to inspect — the damage is in the order of writes.
ConfigurationProvider's constructor calls init(5000), so the timeout is correct. Then runMaintenanceSweep() starts a background config-reaper thread that wrongly treats the live instance as stale and calls resetToDefaults(), writing the timeout back to 0. By the time the test reads it, the heap shows 0 and no local holds 5000 — a single context dump looks like the value was never set.
Debug path: jdwp_set_field_breakpoint(className="…config.Configuration", fieldName="timeout", mode="modification"), then jdwp_get_events shows the two stores in order — 0→5000 on the main thread, then 5000→0 on config-reaper. jdwp_get_stack on the second event names the reaper as the clobberer.
Difficulty: Hard | Test: TransferServiceTest | Package: bank | Par: 4 | Exercises: field watchpoint + stack narration
Symptom: expected discrepancy=0 but was non-zero — money is neither created nor destroyed, yet the audit says the books don't balance.
Hint: The transfer moves money in two steps. The audit snapshot is taken between them.
TransferService.transfer() is not atomic: it calls source.withdraw(), then auditService.snapshotBalances(), then destination.deposit(). The snapshot captures the intermediate state where money has left the source but hasn't arrived at the destination — showing a total of 1500 instead of 2000.
Debug path: jdwp_set_field_breakpoint on AuditService.lastTotalSnapshot (modification). It fires mid-transfer with the stack at TransferService.transfer between the debit and the credit; jdwp_get_stack plus a live balance read confirm the captured dip.
Difficulty: Hard | Test: UserProfileTest | Package: userprofile | Par: 3 | Exercises: event history as evidence
Symptom: expected: <Alice> but was: <alice> — the welcome message rendered correctly, yet the user's stored display name has silently changed casing.
Hint: You will not find the write by ripgrep'ing for setDisplayName — the public setter is never called. A line BP on the setter never fires. Whatever path the write travels, the field's value still flips. Let the JVM tell you exactly when it changes, regardless of how the write reaches the field.
LoginNormalizer.welcomeMessage calls a private canonicalForm helper which delegates to a static DisplayNameMirror nested class. The mirror uses Field.setAccessible(true) + Field.set(profile, canonical) to write the lower-cased form straight into UserProfile.displayName — bypassing the public setter entirely. From the test's perspective the formatter is read-only; from JDI's perspective every reflective write is still a real field modification.
Debug path: jdwp_set_field_breakpoint(className="one.edee.jdwp.sandbox.userprofile.UserProfile", fieldName="displayName", mode="modification") — JDI watchpoints fire on every JVM-level field store, including stores issued through Field.set and Unsafe. The next write suspends the thread inside DisplayNameMirror.mirror; jdwp_get_stack walks back through canonicalForm to LoginNormalizer.welcomeMessage and names the culprit. A line BP on UserProfile.setDisplayName would never fire — the setter is genuinely unused.
Difficulty: Moderate | Test: CheckoutTest | Package: cart | Par: 6 | Exercises: marked instances ($label)
Symptom: expected 45.0 but was 0.0 — the cart goes through pricing and discounting, yet its total never changes.
Hint: The object you get back from the pipeline is not the object you passed in. Prove it.
Checkout.process runs the cart through validate → price → discount. validate was changed to return a defensive snapshot — return new Cart(cart) — so every later stage mutates the copy. process reassigns its local to the copy as it threads the stages, so no single frame holds both the caller's cart and the copy at once. The test ignores the return value (assuming in-place mutation), so its cart stays at total 0. The copy's fields are identical to the original — only identity differs.
Debug path: Break in the test once the cart exists, jdwp_mark_instance(label="input", objectId=<id>) to pin it, then set a breakpoint in price (or applyDiscount) with condition cart != $input. It fires — the stage is working on a doppelgänger, not the caller's cart.
Difficulty: Hard | Test: RaceCounterTest | Package: race | Par: 3 | Exercises: logpoints (non-stopping)
Symptom: expected 2 but was 1 — two threads each increment a counter once, but one increment vanishes.
Hint: Two threads, one lost update. Suspending a thread to look changes the timing — watch the reads without stopping anything.
RaceCounter.increment reads the count, waits at a CyclicBarrier so both threads have read before either writes, then writes back read + 1. Both threads read 0 and both write 1 — one increment is lost. A suspending breakpoint serializes the threads and makes you juggle two parked threads to reconstruct what happened.
Debug path: Set a non-suspending jdwp_set_logpoint at the read site logging the thread name and count (or a field logpoint on count). Let the test run, then jdwp_get_events — both racer-1 and racer-2 are recorded reading 0 before either writes, which is the lost update laid out in order.
Difficulty: Warm-up | Test: DateParserTest | Package: parser | Par: 3 | Exercises: runtime mutation (set_local)
Symptom: NumberFormatException: For input string: "15 " — a date string that looks right fails to parse.
Hint: The input looks almost right. Rather than rebuild with a fix, patch the value in place and see if the test goes green.
DateParser.parse splits YYYY-MM-DD and Integer.parseInts each part. The feed delivers "2026-05-15 " with a trailing space, so parseInt("15 ") throws. The parser never trims its input.
Debug path: Break at the top of parse, observe input = "2026-05-15 ", then jdwp_set_local("input", "2026-05-15") and resume. The parse now succeeds and the test passes — confirming the trailing space is the whole story, with no rebuild.
Difficulty: Hard | Test: TransferDeadlockTest | Package: deadlock | Par: 4 | Exercises: multi-thread inspection (no breakpoints)
Symptom: The test hangs and then fails its join — both transfers never complete.
Hint: Nothing is executing. No exception is thrown. Find out what each thread is waiting for.
Account.transfer locks the source account, pauses, then locks the destination. Two transfers in opposite directions (A→B and B→A) each grab their first lock and then wait forever for the other's — a classic lock-ordering deadlock.
Debug path: No breakpoint can fire — the threads execute nothing. jdwp_get_threads shows transfer-A-to-B and transfer-B-to-A in MONITOR state; jdwp_dump_locks prints the lock graph with both monitor holders and the inverse-order cycle; jdwp_get_stack on each thread shows the lock-ordering inversion. (Trying jdwp_evaluate_expression / jdwp_to_string on these threads is refused — invoking a method on a monitor-blocked thread would hang the debugger, which is itself the tell.)
Flights solved — did you find the root cause?
Style — how close to par? Award ⭐⭐⭐ for a flight solved at its par tool count, ⭐⭐ within 2× par, ⭐ solved at all. A flawless run is 27 stars across the nine flights.
These are the capabilities this server adds on top of raw JDI — the reason to use it instead of writing JDI calls directly.
jdwp_set_breakpoint(
className="com.example.OrderService",
lineNumber=42,
condition="order.getTotal() > 1000"
)
The breakpoint fires on every hit, but the thread is only suspended when the condition evaluates to true. False hits are auto-resumed transparently. This is essential for AI agents — without conditions, a breakpoint in a hot loop would generate thousands of useless stops. Conditions (and expressions, and logpoint bodies) all accept block syntax — wrap in { ...; return X; } to use intermediate locals, try/catch, or early returns.
jdwp_set_logpoint(
className="com.example.OrderService",
lineNumber=42,
expression="\"Processing order \" + order.getId() + \" total=\" + order.getTotal()",
condition="order.getTotal() > 1000"
)
A logpoint evaluates an expression every time the line is hit, logs the result to the event history, and never suspends the thread. Combined with an optional condition, this gives the agent non-intrusive tracing without stopping execution. View results with jdwp_get_events().
By default, every line BP, logpoint, exception BP, and field BP is registered passively: if the target class isn't loaded when the breakpoint is set, the server registers a ClassPrepareRequest and promotes the breakpoint to active when the JVM loads the class on its own — same behaviour as IntelliJ / Eclipse / jdb.
jdwp_set_breakpoint(className="com.example.LazyService", lineNumber=15)
→ "Breakpoint deferred — class com.example.LazyService not yet loaded. Will activate on class load."
jdwp_overview() shows pending breakpoints with their status.
Pass forceLoad=true only when you need the breakpoint to bind immediately — the server will call Class.forName in the target VM to load the class up front. That runs the class's <clinit> (early static init, eager dependency loads), which can mask lazy-init / classloader-leak diagnostics. Use sparingly.
jdwp_set_exception_breakpoint(
exceptionClass="java.lang.NullPointerException",
caught=true,
uncaught=true
)
Catch exceptions at the throw site — before the stack unwinds. Supports deferred activation (if the exception class isn't loaded yet). Use jdwp_overview(types="exception_breakpoint") to see active and pending exception breakpoints.
Log-only mode — record the throw without suspending the thread, evaluating an expression with $exception bound to the thrown object. Use the dedicated jdwp_set_exception_logpoint tool:
jdwp_set_exception_logpoint(
exceptionClass="java.sql.SQLException",
expression="$exception.getSQLState() + \": \" + $exception.getMessage()"
)
Each hit records an EXCEPTION_LOG entry to jdwp_get_events; failures during evaluation surface as EXCEPTION_LOG_ERROR so the listener never throws. Use this to trace exception flows in long-running services without stopping them.
jdwp_set_field_breakpoint(
className="com.example.OrderState",
fieldName="status",
mode="modification"
)
Suspend whenever a field is read (access), written (modification), or both (both — IntelliJ-style, binds both directions to one synthetic ID). Conditions are evaluated against the firing frame with five synthetic bindings:
$oldValue — value before the event (the value being read, or the value about to be overwritten)$newValue — value about to be written (modification events only)$object — the instance the field belongs to, or null for static fields$fieldName — the field name as a string$mode — "access" or "modification", identifying which direction firedLog-only mode — record reads/writes without suspending, evaluating an expression with the same bindings:
jdwp_set_field_logpoint(
className="com.example.OrderState",
fieldName="status",
mode="modification",
expression="$oldValue + \" -> \" + $newValue"
)
Each hit records a FIELD_LOGPOINT entry to jdwp_get_events; failures surface as FIELD_LOGPOINT_ERROR. Filters (threadFilterId, objectFilterId) and trigger chaining (triggerBreakpointId, oneShot) work the same as for line and exception BPs. Deferred activation is supported — the watchpoint installs the moment the declaring class loads, so static-initializer writes are caught.
Hard errors (no silent fallback): invalid mode, ambiguous or missing field, objectFilterId on a static field. The deferred path can't validate static-ness until class load, so it surfaces a warning Note in the response.
Performance: field watchpoints fire on every read/write of the watched field. For hot fields they can dominate target-VM CPU — prefer narrow filters or short-lived sessions. jdwp_diagnose surfaces canWatchFieldAccess / canWatchFieldModification plus a perf warning when connected.
Use jdwp_overview(types="field_breakpoint") to see active and pending field breakpoints with chain status, mode, and any pending failure reason.
jdwp_set_breakpoint(
className="com.example.OrderProcessor",
lineNumber=87,
triggerBreakpointId=12,
oneShot=true
)
Make one breakpoint depend on another. The dependent BP stays disabled until its trigger fires; from then on it's armed and behaves normally. Use this when a hit is uninteresting unless reached via a specific code path — e.g., suspend at the order-pricing line only after a "VIP customer" path has been entered, or break inside a generic utility only when called from a specific feature.
Two modes:
oneShot=false) — once the trigger fires, the dependent stays armed forever. Use when you want to filter by an early gate but observe every subsequent hit.oneShot=true) — the dependent self-disarms after firing, so the next hit requires the trigger to fire again. Matches IntelliJ's "Remove once hit" behavior.Chains compose: a trigger BP can itself be a dependent of an earlier trigger. Cycles are rejected at registration time with the offending path in the error message. Pending (not-yet-loaded) BPs are first-class participants — a chained dependent registered against a class that hasn't loaded yet is promoted with the chain bit intact, so the very first event after class load can't fire before the trigger has. Trigger fires that happen while the dependent is still pending are remembered, so a deferred dependent isn't penalised for arriving late.
Both line breakpoints and exception breakpoints can be chain dependents and chain triggers; mix them freely.
Manage chains with:
jdwp_set_breakpoint_dependency(dependentId, triggerId, oneShot?) — add an edge between two existing BPs (active or pending)jdwp_clear_breakpoint_dependency(dependentId) — detach the edge and re-arm the dependentjdwp_disarm_until_trigger(dependentId) — manually re-disable a dependent without changing its trigger (e.g., after a one-shot fired and you want another round)When the trigger BP is removed, every dependent is automatically armed and a CHAIN_BROKEN event is recorded for each — jdwp_get_events shows exactly which BPs lost their guard. jdwp_diagnose recognises the "every armed BP is waiting on a trigger that hasn't fired" state and tells you why no BP is currently firing, including any pending dependents waiting on class load.
jdwp_evaluate_expression(
threadId=25,
expression="request.getData().get(\"_domain\")",
frameIndex=0
)
→ "self.operationTypeSelect = 3"
Compiles arbitrary Java expressions to bytecode using Eclipse JDT, injects them into the target JVM via ClassLoader.defineClass(), and executes them in the context of the suspended frame. Full classpath is discovered automatically (including container classloaders like Tomcat). Results are cached for performance. Handles Guice/CGLIB proxies automatically.
Two input modes:
order.getTotal(), x + y, list.size() == 5. The compiler treats the input as the body of return ...;.{ ... } to send a multi-statement body with try/catch, intermediate locals, or early return. The agent writes explicit return X; to yield a value. Available everywhere expressions are accepted: jdwp_evaluate_expression, jdwp_assert_expression, breakpoint conditions, logpoint expressions, watchers.See docs/expression-evaluation.md for the compilation pipeline details, or docs/index.md for the full developer reference.
jdwp_assert_expression(
expression="order.getStatus()",
expected="CONFIRMED",
threadId=25
)
→ "OK" or "MISMATCH: actual='PENDING', expected='CONFIRMED'"
Evaluate and compare in one call — useful for agents running verification sequences.
jdwp_set_local(threadId=25, frameIndex=0, varName="retryCount", value="3")
jdwp_set_field(objectId=26886, fieldName="limit", value="100")
Change local variables and object fields at runtime. The agent can test hypotheses ("what if this value were different?") without restarting the application.
jdwp_mark_instance(label="input", objectId=27531)
Pin a cached object in the target heap and label it as $label, so the agent can reference it across many events from inside expressions — conditions, logpoints, watchers, assertions. Reserved bindings ($exception, $oldValue, $newValue, $object, $fieldName, $mode, _this) and label collisions are rejected. Pinned by default via disableCollection; pass pin=false to allow natural GC.
Attach persistent expressions to breakpoints that are evaluated automatically on every hit:
jdwp_attach_watcher(breakpointId=27, label="request data", expression="request.getData()")
jdwp_attach_watcher(breakpointId=27, label="user context", expression="request.getUser().getName()")
# Later, when breakpoint 27 fires:
jdwp_evaluate_watchers(threadId=25, scope="current_frame", breakpointId=27)
Watchers are MCP-side state — they survive across breakpoint hits and are cleaned up when the breakpoint is deleted.
When an expression evaluation at a breakpoint re-enters the breakpointed line (e.g., this.compute(n - 1) evaluated inside compute), JDI would re-suspend the thread and deadlock the server. This server wraps every invokeMethod chain in a per-thread reentrancy guard:
BREAKPOINT_SUPPRESSED / EXCEPTION_SUPPRESSED / STEP_SUPPRESSED entry is recorded in event historyCovered invocation sites: jdwp_evaluate_expression, jdwp_assert_expression, jdwp_evaluate_watchers, logpoint evaluation, conditional breakpoint evaluation, jdwp_to_string, classpath discovery, and deferred class loading via Class.forName.
Threads: jdwp_get_threads() hides JVM internals (Reference Handler, Finalizer, surefire workers) by default. Pass includeSystemThreads=true to see everything. Defaults to a compact 1-line table per thread (a 200-thread Tomcat renders in ~25 lines); pass verbose=true for the legacy block-per-thread format.
Stack frames: jdwp_get_stack() collapses junit/surefire/reflection noise frames by default. Pass includeNoise=true to see the full stack.
jdwp_resume_until_event(timeoutMs=30000)
→ blocks until next breakpoint/step/exception, returns context immediately
Replaces the manual "resume → poll events → poll events" pattern. The agent resumes and gets the next stop in one synchronous call. Returns one of:
Event fired … — a BP / step / exception suspended a thread; the message includes thread + stop site[TIMEOUT] … — the wait expired without an event; the message offers wait_more / reconnect / abort and reports JDI health so the agent never blindly hangs[VM_DEATH] … — the target VM died before any event[NO_EVENT] … — the wait was released without a stop (e.g., a passive disarm); the message states whether the session is still armed or was clearedThe same soft-wait envelope wraps jdwp_wait_for_attach, so bootstrap and resume share one "wait sensibly, decide deliberately" idiom.
jdwp_dump_locks()
→ shows monitor holders, blocked threads, and any deadlock cycles
Takes a transient VM-wide suspend/resume snapshot to enumerate Java monitor ownership — who is blocked on what, who holds it, and any deadlock cycle (lock-ordering inversions are surfaced explicitly). Genuine deadlocks stay deadlocked after the snapshot resumes (the probe is read-only in practice). Only synchronized-monitor contention shows; Object.wait() and java.util.concurrent locks are out of scope by design.
jdwp_reconnect()
→ re-attaches to the last target; BP IDs, conditions, logpoints, chains, watchers all survive
When the target VM is restarted, or the JDI session wedges and you need a fresh transport, jdwp_reconnect re-attaches to the most recent endpoint and replays every BP spec onto the new VM. Synthetic BP IDs are preserved — BP #7 stays #7 — so the agent doesn't have to re-derive its bookkeeping. The marks registry, object cache, last-thread pointer, and classpath cache do not survive vm.dispose() and are cleared by design.
jdwp_get_breakpoint_context(maxFrames=5, includeThisFields=true)
Returns thread info, top stack frames, locals at frame 0, and this fields in a single call — replaces the four-call sequence get_current_thread → get_stack → get_locals → get_fields(this) that an agent would otherwise need at every breakpoint hit.
jdwp_step_over / jdwp_step_into / jdwp_step_out are wired through the same event-and-latch pipeline as breakpoints — the step resumes the thread, the next STEP event suspends it again, and jdwp_resume_until_event blocks on it. threadId is optional; omitted, the step targets the thread of the last breakpoint hit.
Each step is one round-trip, so prefer a breakpoint at the destination + jdwp_resume_until_event whenever you can predict where execution will go. Stepping pays off in three narrow situations:
step_into — polymorphic dispatch is unclear; you can't tell from source which override will run.step_out — an exception/early-abort left you deep in a frame you don't care about and finding the caller line for a breakpoint is awkward.step_over — the single next line, when observing a state mutation is faster than predicting it. More than ~3 in a row → set a breakpoint instead.(For listing or bulk-clearing breakpoints across any combination of kinds, see jdwp_overview and jdwp_clear in the Debug state section below.)
In addition to tools, the server exposes two read-only MCP resources. In Claude Code, type @ in the prompt and pick them from the autocomplete (URI form: @jdwp-inspector:jdwp://...). Attaching a resource pulls its rendered text straight into the prompt without spending a model turn on a tool call — handy for "is my target up, and on which port?" without involving the agent.
Note: resource updates are not pushed to the client — re-attach the URI to see fresh content. For probe-the-world style refreshes (briefly attach to every same-user JVM to learn its port), use the jdwp_diagnose tool with inspectAll=true; the resources do not run those attaches.
1. Launch your app with JDWP enabled
2. In Claude Code:
"Set a breakpoint in OrderService.createOrder line 42 and wait for a hit"
3. Claude:
jdwp_connect()
jdwp_set_breakpoint("com.example.OrderService", 42)
jdwp_resume_until_event(timeoutMs=60000)
jdwp_get_breakpoint_context()
→ "Thread http-nio-8080-exec-3 stopped at OrderService:42.
Local 'order' has total=0.0, status=null — looks like
the order wasn't initialized before reaching this line."
1. "Add a logpoint to trace every order over $1000"
2. Claude:
jdwp_set_logpoint(
"com.example.OrderService", 42,
"\"order=\" + order.getId() + \" total=\" + order.getTotal()",
"order.getTotal() > 1000"
)
jdwp_resume()
3. Later:
jdwp_get_events(count=20)
→ Shows LOGPOINT entries with evaluated expressions, no thread was ever stopped
1. "I'm getting a NullPointerException somewhere in the order flow"
2. Claude:
jdwp_set_exception_breakpoint("java.lang.NullPointerException", caught=true, uncaught=true)
jdwp_resume_until_event(timeoutMs=60000)
jdwp_get_breakpoint_context()
→ "NullPointerException thrown at OrderValidator:87.
Local 'customer' is null — the order was submitted without a customer reference."
The jdwp-sandbox module includes a deterministic scenario (one.edee.jdwp.sandbox.recursion package) that reproduces the recursive breakpoint case:
# Terminal 1 — launch sandbox, suspended on port 5005
./mvnw -pl jdwp-sandbox test -Dtest=RecursiveCalculatorTest -DskipTests=false -Dmaven.surefire.debug
# From Claude Code:
jdwp_wait_for_attach()
jdwp_set_breakpoint("one.edee.jdwp.sandbox.recursion.RecursiveCalculator", 22)
jdwp_resume_until_event()
# → BP fires inside compute(5)
jdwp_evaluate_expression(threadId, "this.compute(3)")
# → returns 2 without deadlock
jdwp_get_events()
# → shows BREAKPOINT_SUPPRESSED entries for each recursive hit
Claude Code ──MCP/STDIO──> Spring Boot MCP Server ──JDI──> Target JVM (port 5005)
The server is SYNC mode, web-application-type=none — JSON over STDIO, no HTTP.
evaluation/)URLClassLoaders). Composes JDWP_EXTRA_CLASSPATH (override), a depth-5 scan of target/classes / target/test-classes under the server's CWD, and mvn dependency:build-classpath. Set JDWP_EXTRA_CLASSPATH=/path/extra.jar:/path/more.jar (colon/semicolon-separated) to plug specific gaps; jdwp_diagnose shows a per-source breakdown. Details in docs/expression-evaluation.md.ClassLoader.defineClass() and invokes it.watchers/)mcp-jdwp-java/
├── pom.xml # Parent POM (reactor)
├── mvnw / mvnw.cmd # Maven wrapper
├── README.md
├── CHANGELOG.md
├── .mcp.json # MCP server configuration
├── .claude-plugin/
│ ├── plugin.json # Claude Code plugin metadata
│ └── marketplace.json # Plugin marketplace registry
├── hooks/
│ └── hooks.json # SessionStart hook: auto-builds JAR if missing or git HEAD moved
├── skills/
│ ├── java-debug/ # Debugging skill (workflows, recipes, gotchas)
│ └── release/ # Release-cutting skill
├── docs/ # Developer reference (architecture, lifecycle, …)
│
├── jdwp-mcp-server/ # The MCP server
│ ├── pom.xml
│ └── src/main/java/one/edee/mcp/jdwp/
│ ├── JDWPMcpServerApplication.java
│ ├── JDWPTools.java # 47 @McpTool methods + 2 @McpResource
│ ├── JDIConnectionService.java # JDI connection + object cache
│ ├── BreakpointTracker.java # Breakpoint registry + deferred state + chains
│ ├── JdiEventListener.java # JDI event consumer
│ ├── EvaluationGuard.java # Recursive breakpoint protection
│ ├── EventHistory.java # Event ring buffer
│ ├── JdiHealthMonitor.java # Transport liveness for soft-wait
│ ├── DeadlockAnalyzer.java # Monitor-lock graph + cycle detection
│ ├── ClassNameMatcher.java # Glob → JDI ReferenceType matcher
│ ├── ThreadFormatting.java # Thread/frame noise filtering
│ ├── evaluation/ # Expression compile-and-inject pipeline
│ │ ├── JdiExpressionEvaluator.java
│ │ ├── RemoteCodeExecutor.java
│ │ ├── InMemoryJavaCompiler.java
│ │ ├── ClasspathDiscoverer.java
│ │ └── JdkDiscoveryService.java
│ ├── marks/ # Named, pinned heap references ($label)
│ │ ├── MarkedInstanceRegistry.java
│ │ ├── MarkInfo.java
│ │ └── ReservedBindings.java
│ ├── discovery/ # Local-JVM inventory + JDWP port probing
│ │ ├── JvmDiscoveryService.java
│ │ ├── DiagnoseReportRenderer.java
│ │ ├── JdwpAgentArgParser.java
│ │ ├── JdwpEndpoint.java
│ │ ├── JvmDescriptor.java
│ │ └── exceptions/
│ ├── transport/ # STDIO transport overrides
│ │ ├── MultiVersionStdioServerTransportProvider.java
│ │ └── StdioTransportConfig.java
│ └── watchers/ # Breakpoint-attached expression watchers
│ ├── WatcherManager.java
│ └── Watcher.java
│
└── jdwp-sandbox/ # Debugging targets (test flights)
├── pom.xml # Tests skipped by default
└── src/ # Deliberately broken scenarios
jdk.jdi module) — Java Debug InterfaceFor Java developers who want to understand the internals — safety guarantees, memory allocation, threading, bootstrapping, state clearing, the test architecture — the docs/ folder is the developer reference.
Start at docs/index.md for navigation and a reading guide. The chapters cover:
INVOKE_SINGLE_THREADED, the EvaluationGuard reentrancy mechanismdefineClass, invokeMethod)jdwp_diagnose, JVM discovery, the handshake probe, transport-loss envelopes, loggingMIT
MCP server giving AI agents full debugger control over running Java applications — inspect state, set breakpoints, evaluate expressions, and mutate values at runtime via JDWP/JDI
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
MCP Jdwp Java 透過 stdio 在本機執行。
上次檢查時間：2026-10-27。 MCP server giving AI agents full debugger control over running Java applications — inspect state, set breakpoints, evaluate expressions, and mutate values at runtime via JDWP/JDI
