MCP for coding agents
Your coding agent can edit a Java service and finish a build. Reclazz's MCP tools help it answer the next questions: Which JVM is running? Were the changed bytes applied? Is something held or waiting for a restart? What should it check before telling you the change works?
Keep the application running while the agent gathers that evidence. The bridge exposes eleven tools, read-only resources and a verification prompt through stdio, backed by the attached Reclazz agent's local status socket. Use the MCP reference for installation, full schemas and target selection. This page explains the developer benefits.
Examples below come from recorded local acceptance through 10 November 2026. Tool requests are shown as an MCP tool name and argument object, not a shell command. The resource and prompt examples identify their method and parameters. Results show selected fields or selected array entries, not complete responses. Uppercase session, cursor, path, hash and owner values are placeholders to replace with evidence from your own run. Rendering this page does not rerun acceptance.
Find the right running application
An agent can compile successfully while talking to the wrong server. Start with
reclazz_status for attachment and health, then reclazz_doctor for JVM identity,
session, launch directory, capabilities and watcher evidence.
{
"name": "reclazz_doctor",
"arguments": {
"timeoutMs": "1500"
}
}
Recorded result excerpt:
{
"status": "observed",
"watcherState": "watching",
"buildHold": "none",
"verifySupported": true,
"reloadConfirmed": false
}
Compare the returned PID, session and working directory with the application you intended. With an explicit project, the server checks that connection's DOCTOR launch-directory evidence before sending the requested command.
Limit: Attachment and a healthy watcher do not prove that an edit reloaded. A sampled directory list can be incomplete; a working directory is not a repository-root assertion. See target checks.
Replace a guessed wait with a bounded result
After saving, an agent needs more than a delay or an increasing reload counter.
reclazz_await_reload creates a checkpoint before the edit, then waits for
qualifying received events and an exact-owner byte receipt.
First call reclazz_await_reload with action:"checkpoint" and require
status:"ready". Save only after that reply. Use its session, subscription and
cursor in the wait:
{
"name": "reclazz_await_reload",
"arguments": {
"action": "wait",
"sessionId": "SESSION_FROM_CHECKPOINT",
"subscriptionId": "STREAM_FROM_CHECKPOINT",
"afterCursor": "CURSOR_FROM_CHECKPOINT",
"filePath": "ABSOLUTE_SOURCE_PATH",
"ownerClassName": "com.example.demo.HelloController",
"timeoutMs": "10000"
}
}
Recorded result excerpt:
{
"status": "applied",
"reloadConfirmed": true,
"correlationScope": "received_events",
"sourcePathConfirmed": false,
"editRevisionConfirmed": false
}
Recorded acceptance also exercised compile_error, refused,
restart_required and held, so an agent can branch on a failed compile,
refusal, hierarchy limit or blocked admission. timeout and unavailable
remain explicit outcomes, never success by silence.
Limit: Received events do not identify an exact editor revision or full source path. A matching basename can be ambiguous. Require the explicit binary owner in file mode; read the receipt and still exercise the application. See checkpoint and wait.
Prove the compiled bytes before claiming completion
A successful compiler exit does not show which bytes the live JVM accepted.
Hash the changed .class file and ask reclazz_verify for that binary class
and SHA-256. Use reclazz_verify_batch for several class/hash pairs on one
connection; inspect every result.
{
"name": "reclazz_verify",
"arguments": {
"className": "com.example.demo.HelloController",
"sha256": "SHA256_OF_YOUR_CLASS_FILE",
"timeoutMs": "1500"
}
}
Recorded result excerpt:
{
"status": "applied",
"className": "com.example.demo.HelloController",
"detail": "method bodies"
}
Require isError:false, status:"applied", the expected session and matching
expected/observed hashes. A recorded batch result had allApplied:true while
still reporting atomic:false and reloadConfirmed:false: its individual
receipts are evidence, not an atomic application transaction.
Limit: A byte receipt is not proof of the feature's behavior, source revision or successful database work. Make the HTTP, UI or application test the last step. See exact-byte verification.
Inspect several edits without reading a log by eye
When an agent edits several files, reclazz_events gives it bounded, ordered
pages from a checkpoint. Filter by level, exact owner or file basename to
inspect the relevant observations.
{
"name": "reclazz_events",
"arguments": {
"sessionId": "SESSION_FROM_CHECKPOINT",
"subscriptionId": "STREAM_FROM_CHECKPOINT",
"sinceCursor": "0",
"limit": "100",
"timeoutMs": "10000"
}
}
Recorded result excerpt, showing the first two of eleven returned events:
{
"status": "observed",
"nextCursor": "11",
"hasMore": false,
"reloadConfirmed": false,
"events": [
{
"sequence": "1",
"type": "compile",
"fileBasename": "HelloController.java"
},
{
"sequence": "2",
"type": "reload",
"ownerClassName": "com.example.demo.HelloController"
}
]
}
Feed back nextCursor to continue. Filters apply to each event together; an
empty filtered page can advance over inspected nonmatches. Cursor eviction
is reported as unavailable instead of returning a misleading partial history.
Limit: This is bounded received-event history, not a durable attempt journal or a full-path/revision tracker. Event order does not prove an atomic batch. An observed reload is not confirmed application behavior; use await/VERIFY and your own application oracle. See event pages.
Keep partial builds out of the reload queue
A compiler can write some classes before failing on another file. Start an
owned BUILD hold before compilation so that a failed build does not authorize
that partial output. The same owner reports ok only after the complete build
succeeds, or failed when it does not.
{
"name": "reclazz_build",
"arguments": {
"state": "started",
"owner": "YOUR_UNIQUE_BUILD_OWNER",
"timeoutMs": "1500"
}
}
Recorded result excerpt:
{
"status": "acknowledged",
"state": "started",
"reloadConfirmed": false
}
The build hold also gates automatic source/resource reload actions. The live SAP acceptance retained the previous HTTP/class behavior under both started and failed holds, rejected a competing owner, then applied the latest source after the original owner's successful build. The terminal build wrapper manages these signals around the complete compile command.
Limit: An acknowledgement is not reload completion. SCAN or resume cannot
release the hold. Retain the owner token for recovery; do not send ok just
to clear a warning. Already admitted work can finish, and framework refresh
and external side effects are not transactional.
Stage changes while keeping the application available
Use reclazz_reload_control to pause automatic reload admission while preparing
several edits. The application keeps serving requests. Resume when the changes
are ready, then obtain receipts and check behavior.
{
"name": "reclazz_reload_control",
"arguments": {
"action": "pause",
"timeoutMs": "1500"
}
}
Recorded result excerpt:
{
"status": "acknowledged",
"action": "pause",
"paused": true,
"buildHold": "none",
"reloadConfirmed": false
}
action:"status" reads the current admission state; action:"resume" resumes
it. Independent client acceptance checked held output, resume, exact-byte
verification and changed behavior in the same application JVM.
Limit: Pause does not freeze requests, roll back state or stop already admitted work. Resume does not unlock a failed BUILD hold. See pause and resume.
Give the agent a concrete next diagnostic
An unchanged response should lead to an investigation, not another optimistic
completion message. reclazz_pending exposes the restart ledger and admission
state. reclazz_diagnose describes the last known outcome for a named class.
{
"name": "reclazz_diagnose",
"arguments": {
"className": "com.example.demo.HelloController",
"timeoutMs": "1500"
}
}
Recorded result excerpt, selecting the actionable diagnostic line:
{
"status": "observed",
"className": "com.example.demo.HelloController",
"complete": false,
"lines": [
"Reclazz has not attempted a reload of this class, and the compiled file has not been rebuilt since this JVM started. The build did not reach it."
]
}
That result directs the agent back to compiler output rather than a server restart. A pending restart requirement calls for its specific recovery instead.
Limit: Pending/diagnose use a legacy broadcast diagnostic stream with no
response-end marker. complete:false is deliberate: the lines are observations,
not a guaranteed complete or exclusively request-owned response. See
diagnostic errors.
Use the same evidence loop for Spring Boot and SAP Commerce
The tools speak to the running Reclazz agent, so the workflow does not depend on an IDE. Recorded acceptance used both a Spring Boot sample and a clean, licensed SAP Commerce 2211 installation with synthetic fixtures.
For example, reclazz_doctor returned status:"observed" and
verifySupported:true on both targets. Applied byte receipts were followed by
real HTTP assertions. Await acceptance exercised eighteen outcomes per target;
event-page acceptance passed fifteen checks per target, including restored
application baselines. These are recorded runs, not new runs of this page.
Limit: Shared tools do not make every framework edit reloadable. Type-system updates, schema changes, framework refresh and restart requirements retain their own conditions. SAP events in this tool are not a dedicated SAP operation receipt API. See live SAP scope and reload limits.
Read SAP operation outcomes with their limits
A coding agent investigating a resource edit can read a stable operation and outcome instead of interpreting every log sentence. Event pages and the event resource classify recognized ImpEx, codegen, items schema, Backoffice reset and Solr messages. Java await also includes a bounded list of recent SAP observations, so the agent can see related activity while checking Java bytes.
This selected field projection comes from a recorded live ImpEx event:
{
"type": "sap_operation",
"sap": {
"operation": "impex_import",
"reportedOutcome": "applied",
"reasonCode": "imported",
"evidenceScope": "agent_message",
"applicationBehaviorConfirmed": false,
"resourceBasename": "test-data.impex"
}
}
It is an agent message, not a request receipt. The row does not identify
which edit caused it or confirm application behavior. Follow it with a native
data or endpoint check. In particular, codegen generated precedes the need
to verify applied classes, and Solr completion still carries
fieldFreshnessConfirmed:false.
Recorded live acceptance observed ImpEx applied/refused/skipped, codegen generated, items schema applied, Backoffice reset applied, and Solr finished/failed. Native data, generated classes/OCC, schema, stored documents and configuration media supplied separate behavioral checks. Other recognized outcomes have unit/fake-agent coverage only. Backoffice media checks do not establish browser rendering or widget/socket behavior.
Limit: SAP observations cannot turn a Java await into success. Await keeps at most 16 SAP rows and reports truncation; invalid observation history clears them. Read the full typed contract and live coverage for outcome meanings and recovery boundaries.
Connect the coding client you use
Claude Code and Codex CLI each discovered all eleven tools and made one real
reclazz_status call against the Spring sample in recorded local acceptance:
{
"name": "reclazz_status",
"arguments": {}
}
Recorded result excerpt:
{
"attached": true,
"agent": "1.7.0",
"protocol": 1
}
Use the Claude Code setup, Codex TOML entry or Cursor configuration. The client matrix records exact versions, instrumentation and which configuration route was exercised.
Limit: Cursor setup is documented, not verified. Its desktop installation did not provide an installed non-GUI Agent CLI for that acceptance. The two CLI results do not certify desktop GUIs, other operating systems or every tool in each client. Client permissions and provider authentication remain separate.
Read context and choose a verification workflow
A coding client may want attachment or restart context as a resource, or a
reusable prompt before the next edit. reclazz://status and reclazz://pending
provide the same diagnostic JSON as the corresponding tools. The event URI
template shares the tools' existing checkpoint and bounded event history.
For example, packaged SDK acceptance used these resources/read parameters:
{
"uri": "reclazz://status"
}
Recorded attachment fields decoded from the resource's contents[0].text:
{
"attached": true,
"agent": "1.7.0",
"protocol": 1
}
This is a projection of the recorded result, not its port and full health data. Resource reads stay on the server's startup target; a URI cannot select another project or port. A successful read can contain unavailable evidence, so inspect its JSON before treating it as useful context.
Clients can also retrieve verify_my_last_change through prompts/get:
{
"name": "verify_my_last_change"
}
The prompt guides doctor -> checkpoint before a new edit -> real BUILD signals
for external compilation -> await reload -> verify the bytes -> check the
application. It explains already-saved edits and tells the agent never to send
BUILD ok merely to clear a hold. Fetching the prompt executes none of those
steps and grants no permission to perform them.
Limit: The client decides how to surface these capabilities. Recorded Claude Code acceptance listed both resources and the prompt, but did not request the template, read a resource or retrieve the prompt in its constrained run. The official SDK separately exercised those operations. Event resources remain observations, not byte receipts, source-revision proof or application tests. Tools-only clients keep the same eleven tools. See resources and prompts for selectors, errors and exact client coverage.
Keep the Reclazz bridge local
The transport is coding client ↔ stdio MCP server ↔ 127.0.0.1 status socket.
The MCP server has no hosted endpoint or telemetry; the agent opens its status
server on loopback and does not dial out. The bridge does not require a cloud
account to contact the application.
For example, the local status request above returned the agent's attachment and health. It did not need to upload source to a Reclazz service. The agent network guard checks built agent calls, while the bridge's socket client explicitly selects loopback.
Limit: Your coding client can contact its model provider and send tool results there. Local Reclazz transport is not a claim that the whole AI workflow stays offline. Loopback is not authentication against other local processes. Mutating tools can trigger application callbacks and external side effects; inspect their conservative hints and require appropriate permissions.
Follow the proof, then check your application
The documented behavior is present in matching locally built agent/MCP 1.5.0. Source presence does not establish that an older published release contains it; follow the matching-build instructions.
| Evidence area | What was independently accepted | Public contract or test |
|---|---|---|
| Status, doctor, build, verify, batch, pause and diagnostics | Local Spring/SAP tool probes and real application checks | MCP reference, packaged workflow |
| Automatic actions during BUILD holds | Started/failed hold, same-owner recovery and restored SAP behavior | Automatic action tests |
| Await reload | Received-event correlation, exact-owner receipts and negative outcomes | Await tests |
| Event pages | Ordered pagination, exact filters, gaps and honest observation limits | Event tests |
| Resources and verification prompt | Packaged resource/prompt contracts and SDK reads; installed Claude Code listing with explicit limits | Resource and prompt reference |
| Typed SAP observations | Parsed messages from all five families, with separate native oracles and explicit unit-only outcomes | Typed SAP contract |
| Real coding clients | Eleven-tool discovery and read-only status in Claude Code and Codex CLI | Dated client acceptance |
The bridge provides tools, read-only resources and a verification prompt. SAP-specific operation receipts remain outside this work. No automatic restart, durable edit journal, source-revision proof or universal rollback is promised.
Start with the setup reference, verify the intended JVM, and finish the edit loop with the application's own result.