@exeris/ai-bridge
OfficialClick on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@@exeris/ai-bridge what did my last build emit?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
@exeris/ai-bridge
Model Context Protocol server that gives AI agents the context of an Exeris application: what your build emitted, the SDK's authoring contract, the ADR registry, and a running kernel.
Every tool is a read. The bridge never writes to your project, never calls a model API, and runs as a separate process from the Java kernel. It is built for developers writing an application on Exeris, and works without a checkout of the ecosystem: the reference data it needs ships inside the package.
Installation
Run the server inside your project, so it finds your pom.xml.
Claude Code
claude mcp add exeris -- npx -y @exeris/ai-bridgeClaude Desktop: install exeris-ai-bridge-<version>.mcpb from the
GitHub Releases. Desktop does not
start the server inside a project, so set Project root in the extension's settings to use the
build-* and caps-* tools.
Any other MCP client (Cursor, VS Code, Copilot): the command is npx -y @exeris/ai-bridge
over stdio.
{
"mcpServers": {
"exeris": { "command": "npx", "args": ["-y", "@exeris/ai-bridge"] }
}
}The server is also listed in the MCP Registry as
io.github.exeris-systems/ai-bridge.
Related MCP server: Yasin-MCP
Requirements
Node 20 or newer.
kernel-*only: JDK 25 or newer, andeu.exeris:exeris-kernel-diagnostics-cliin your local Maven repository, which a build of an Exeris application puts there.build-preview_generationonly: forjava, JDK 25 or newer andeu.exeris:exeris-codegen-cli, at the tooling version your project pins, in your local Maven repository; forts,@exeris/codegen-tsin your project's ownnode_modules.No network at run time. The bridge reads artefacts already in
~/.m2and never downloads.
Tools
Tool names are family-tool. All 34 are listed in every session; a family whose dependency is
missing answers with a structured family_unavailable result naming the reason and the remedy.
Family | Answers | Tools |
| What your last build emitted, which files codegen owns, what a build diagnostic means, the build files for a new application, what the generator would write now, and whether your discovery briefs are ready; |
|
| The capability composition your build validated |
|
| The SDK's annotations, deprecations, scoping rules and AST schema; whether your project pins a different SDK |
|
| The Exeris ADRs, the high-level architecture, the whitepaper and the decision templates |
|
| A running kernel's providers, bootstrap DAG, subsystems and JVM ergonomics |
|
|
|
|
| The server itself: version, bundled data, and why a family is unavailable |
|
docs-list_repos, docs-list_repo_docs and docs-get_repo_doc need a checkout of the Exeris
repositories; the other docs-* tools answer from a checkout when there is one, and otherwise from a
registry snapshot bundled in the package.
lsp-* needs a locally built exeris-platform-lsp, which publishes no runnable jar yet.
Resources
The registry tier of docs-* is also available as MCP resources, all text/markdown, listed in
every session. A read goes through the same checkout-or-snapshot resolution and public-scope policy
as the tools.
URI | Content |
| The ADR registry index |
| One ADR by its three-digit number; a record from a repository that is not served publicly names its owning repository and GitHub URL |
| The high-level architecture |
| The B2B technical whitepaper |
| A decision-doc template |
| The JSON Schema of a discovery brief (ADR-106), |
exeris://docs/adr-index can be subscribed to (resources/subscribe): from an exeris-docs
checkout the server sends notifications/resources/updated once per burst of changes to the file,
including a replacement by rename; from the bundled snapshot the content cannot change, so the
subscription is accepted and never fires. Any other URI, exeris://sdk/brief-schema included, is rejected as invalid params.
A URI outside this set, or an ADR number that is not three digits, is rejected as invalid params. An
ADR that is not in the registry or has no servable content answers with the MCP resource-not-found
error (-32002); docs being unavailable answers with an internal error whose data is the
family_unavailable reason and remedy.
Prompts
Nine stateless prompt templates, listed in every session on every transport. A prompt returns
messages built from its template and the arguments you give it: the bridge keeps no state, runs no
loop and calls no model or tool. Where a step needs a tool, the prompt tells the agent to call it.
Context is embedded as resource messages (the exeris://docs/* resources) or as bundled SDK data;
where it cannot be embedded, the prompt says which resource to read instead. Arguments are strings,
at most 256 characters, except diff (65,536 characters), output (64,000) and area (64, lowercase
letters, digits and hyphens).
Prompt | Arguments | Pre-loads | Needs a local server |
|
| HLA, ADR-006, ADR-024, the drift-patterns policy, the cap-tier Wall rule | no |
|
| ADR template, registry index | no |
|
| the ecosystem's routing rules, from an ecosystem checkout only | no |
|
| ADR-006 | no |
|
| SDK scoping rules, AST schema |
|
|
| nothing; the contract is the | the validation step |
| none | nothing; calls | yes |
|
| SDK deprecation list | the |
|
| SDK deprecation list | no |
route-this-task without an ecosystem checkout answers that routing needs one and names a remedy.
An unknown prompt, an unknown argument, a missing required argument or one over its bound is
rejected as invalid params.
Configuration
Nothing needs setting for an application. Every variable is optional, and none can stop the server from booting.
Variable | Default | Purpose |
| nearest | The project |
|
| Where published jars are looked up |
| newest release in the local repository | The diagnostics CLI version to launch |
| resolved from the local repository | Launch the kernel diagnostics CLI explicitly |
| none | Launch |
| the working directory | The workspace the language server indexes |
| a sibling | The |
|
|
|
How each variable resolves, the launch ladder for the Java children, and the bundled reference data are described in the configuration reference.
Troubleshooting
Call bridge-health. It reports every family's state, with the reason and the remedy for each
one that is unavailable, without starting any child process. bridge-version reports the build
and the upstream releases its bundled data came from. The server also writes a one-line summary
(mode=… docs=… lsp=… kernel=… build=… caps=… sdk=…) to stderr, which MCP clients show in their logs.
Versioning
The package follows semver. Until 1.0, a minor release may change a tool's output shape; the release notes list every change a client may notice. At 1.0 the tool names and input schemas are frozen.
Release 0.9.0 vendors exeris-sdk 0.13.3 and exeris-tooling 0.11.0 into the package; the tooling
0.11.0 application BOM pins exeris-kernel 0.13.1. All three are on Maven Central.
Links
License
Apache-2.0
Available Tools
34 toolsbridge-healthA
Report which tool families can serve requests and why the others cannot. Each family carries its state and, when unavailable, the reason and the remedy. Families backed by a child process also report that process's current state — without starting it. Call this first when a tool returns 'family_unavailable'.
Inputs: no named parameters.
bridge-versionA
Identify this MCP server: package name and version, the Node runtime it is running on, which persona mode it resolved to, and what bundled reference data it carries. Use it to confirm which bridge build an answer came from, and which upstream release that data reflects.
Inputs: no named parameters.
build-explain_artefactsA
List the files Exeris codegen has generated into this project, each with the generator that produced it and when that generator emits — handlers, services, repositories, domain errors, event publishers, graph projections, sagas, Flyway migrations, OpenAPI specs, and the project-wide bootstrap. Ask this BEFORE hand-writing SQL, a DTO or a controller for an Exeris entity: the build usually already emits it. Optional entity narrows to one entity's artefacts. Reports what the pipeline actually emitted, read from its own output manifest and explained by the generator catalogue exeris-tooling publishes with its codegen jar. Each artefact carries candidates: every catalogue row that can have written the path, first match first. One candidate means generator names it; several mean the path shape is shared and generator is null rather than a guess; none means the path is unrecognised. A project that has not run codegen returns present:false.
Inputs: entity.
build-explain_diagnosticA
Decode an error, warning or note exeris-tooling printed — an annotation-processor message, a codegen plugin goal failure or warning, a capability-graph or cap-tier Wall refusal, a missing runtime driver, a code-generation pipeline or CodegenMain line, or an exeris-initializr error. Paste the diagnostic text verbatim; returns what it means, the rule behind it, and what to do. Looks each diagnostic up by its stable EXT- identifier (ADR-095), which exeris-tooling prints from 0.9.0, and falls back to message text only for output that carries none. An identifier with no catalogue entry is named back as a catalogue gap, and a diagnostic it does not recognise is reported as not in the catalogue — never as not being an Exeris problem.
Inputs: text (required).
build-get_detach_stateA
Report whether this project's generated code is still pipeline-owned (L1, src/main/generated/java — regenerated on every build, so edits are lost) or has been detached into the developer's own sources (L2, src/main/java — owned, safe to edit). Ask this before editing anything under a generated tree. Returns the state, the observations behind it, and how many files the pipeline claims.
Inputs: no named parameters.
build-get_domain_metadataA
Fetch the DomainMetadata JSON the annotation processor emitted for one declaration of this project — the authoritative AST the code generators consume, including fields, actions, events, relationships and the system-field flags. name is the Java simple name (e.g. 'Station'), as listed by build-list_domains. Returns the record verbatim, plus the schemaVersion and (when the build could compute one) the ADR-042 sourceDigest.
Inputs: name (required), kind.
build-get_starterA
Return the build files for a NEW Exeris application and write none: for 'backend', a pom.xml whose parent is exeris-app-parent and which depends on exeris-app-starter; for 'frontend', a package.json with @exeris/codegen-ts and its exeris-gen generate script; for 'both', the two under backend/ and frontend/. Ask this BEFORE hand-writing a pom.xml or package.json for an Exeris application: the starter already wires the annotation processor, the codegen goals and the dependency set. Every version comes from the exeris-app-bom the bridge was released with, which names one tested tooling + kernel + SDK triple; the tool reports it as bomVersion and triple. supplied lists what the application still adds itself, starting with the JDBC driver. Needs no project: it reads bundled data only, and you write the files with your own tools.
Inputs: variant (required).
build-list_domainsA
List what this project's last build emitted into target/classes/exeris-metadata: every @ExerisDomain entity, plus the enums, capabilities and views the processor wrote alongside them, each with its kind. Use it to discover names before calling build-get_domain_metadata. A project that has not been built, or has no @ExerisDomain sources, returns present:false rather than an error.
Inputs: no named parameters.
build-preview_generationA
Preview what the code generator would write for this project's current metadata (target/classes/exeris-metadata), without writing anything into the project. It runs the project's own generator in a separate process — for 'ts' the @exeris/codegen-ts installed in the project's node_modules, confined by Node's permission model to write only its scratch directory; for 'java' the eu.exeris:exeris-codegen-cli jar from the local Maven repository at the version the project pins — never downloaded and never substituted; when none resolves, a structured generator_unavailable with a reason and remedy. Returns the files it produced as added, changed, unchanged and removed against the committed generated tree (src/main/generated/java for Java, the codegen outputPath for TypeScript), a unified diff for each added or changed file, and the run's diagnostics with every EXT- identifier decoded as build-explain_diagnostic does. 'unchanged' includes hand-written files already in the committed tree, which the run starts from; the generator's own manifest file is not listed. Unlike the present:false answers of the other build-* tools, this one cannot answer without metadata, so a project that has not been built returns the error metadata_missing: run the project build first (mvn compile). Metadata whose names could not be a Java identifier or a package are refused as metadata_invalid before any generator starts. Bounded by a timeout and by output and file-size caps; a run over one is stopped and the limit is named. Local only; the output exists in a temporary directory removed when the call returns.
Inputs: target (required).
build-validate_briefA
Check the pre-construction brief(s) of this project, docs/briefs/.brief.yaml, against the brief schema (ADR-106) and say whether each is ready for authoring. Reads only the brief files, the project's pom.xml (and the app BOM poms it inherits from, read from the local Maven repository), the brief schema and, when present, target/classes/exeris-metadata; writes nothing, starts no process and uses no network, and is local only: the family needs the project root, so private product intent never reaches a hosted endpoint. The schema is the one inside the project's pinned eu.exeris:exeris-sdk-source-model jar in the local Maven repository (never downloaded), else the schema bundled with this bridge; skew.schemaSource says which. A project pinned below SDK 0.13.0 has no brief contract and gets a reason instead. A brief whose briefSchema is newer than the SDK of the schema used is returned valid:false, ready:false with the skew reported, and is not validated. valid means the schema passes. ready means valid and no blocking open question is still 'open'; an absent, empty or invalid brief is never ready. assumed lists the questions decided on credit with their owner. EARS criteria are checked by structure, never by text. Where the project has been built, each fulfilledBy reference is resolved against the emitted DomainMetadata (an entity by qualified class name; an action, rule or saga as #); unresolved names those that match nothing and does not change readiness, and without metadata every reference is unchecked. Optional area names one brief; without it every brief directly under docs/briefs is validated, up to 50.
Inputs: area.
caps-describe_compositionA
Report the composition stamp from cap-manifest.json: whether the capability graph validated, the composition version, and the content binding digest, alongside the manifest schema version, the module count and initOrder. Answers 'is this build's composition trustworthy, and which build is it'. A project with no capabilities returns a present:false answer, not an error.
Inputs: no named parameters.
caps-list_capabilitiesA
List the capability modules composed into this project, from the build-time cap-manifest.json: each module's qualified name, the services it provides with their versions, and which module owns the lifecycle. Also returns initOrder exactly as the manifest states it. A project with no capabilities returns an empty list, not an error.
Inputs: no named parameters.
docs-get_adrA
Fetch the full markdown body of a specific ADR by number. Resolves the link from the registry against the ecosystem root, so cross-repo ADRs (those whose content lives in a sibling repo) are read from their authoritative source.
Inputs: number (required).
docs-get_hlaA
Fetch the High-Level Architecture document (high-level-architecture.md) from exeris-docs. Canonical narrative of the three-tier model.
Inputs: no named parameters.
docs-get_repo_docA
Fetch a specific markdown document from <repo>/docs/<path>. Use docs-list_repo_docs to discover valid paths.
Inputs: path (required), repo (required).
docs-get_templateA
Fetch one of the canonical decision-doc templates from exeris-docs/templates/. Three kinds: ADR (decision made), RFC (multi-option strategic question), RESEARCH (falsifiable hypothesis, lab-notebook shape).
Inputs: kind (required).
docs-get_whitepaperA
Fetch the B2B technical whitepaper (b2b-technical-whitepaper.md) from exeris-docs.
Inputs: no named parameters.
docs-list_adrsA
List all ADRs from the Exeris ecosystem registry. Each entry carries number, title, owning repo, scope, visibility, status, and link. Optional status filter narrows the result by the first token of the status cell (e.g. 'accepted', 'proposed', 'superseded', 'reserved').
Inputs: status.
docs-list_repo_docsA
List markdown documents under <repo>/docs/** for an Exeris sibling repo. The adr/ subdirectory is excluded — those are the registry's territory and reachable via docs-list_adrs / docs-get_adr. Use docs-list_repos to discover valid repo names.
Inputs: repo (required).
docs-list_reposA
List sibling Exeris repos under the ecosystem root that publish a docs/ directory. Returned names are usable as repo input to docs-list_repo_docs and docs-get_repo_doc.
Inputs: no named parameters.
docs-searchA
Search exeris-docs for a literal substring (case-insensitive). Returns the path, line number, and snippet of each match, capped by maxResults. Optional pathFilter narrows to files whose docs-root-relative path contains the substring (e.g. 'adr/' to scope to ADR files). Scope is the docs root only; cross-repo content is reachable via get_adr.
Inputs: query (required), maxResults, pathFilter.
kernel-describe_subsystemB
Detail for a single kernel subsystem by name (e.g. memory, crypto, persistence, graph, transport, events, flow, http, security).
Inputs: name (required).
kernel-get_bootstrap_dagA
Snapshot of the kernel bootstrap dependency DAG — nodes (subsystems) with their phase, declared dependencies, and running state.
Inputs: no named parameters.
kernel-get_jvm_ergonomicsA
What the JVM running the kernel actually resolved at startup: selected GC, max/committed heap, visible processors, and the container limits it detected (cgroup CPU quota/period, memory cap, cpuset) plus large-page, CDS and AOT-cache state. Null fields mean the kernel could not determine that value. A kernel that does not implement this answers a snapshot with gcName 'unknown' and -1 byte counts.
Inputs: no named parameters.
kernel-list_providersA
List all SPI providers registered with the running kernel, including driver origin (community/enterprise priority).
Inputs: no named parameters.
lsp-describe_domainA
Return the source-model description of an @ExerisDomain type: fields, actions, relationships, generated artefacts, and the sourceDigest that lsp-preview_mutation takes as its concurrency token.
Inputs: qualifiedName (required).
lsp-list_actionsB
List all @Action methods across the workspace, with their owning @ExerisDomain.
Inputs: no named parameters.
lsp-list_domainsB
List all @ExerisDomain types known to the active LSP session.
Inputs: no named parameters.
lsp-preview_mutationA
Preview how the SDK's canonical writer would change an @ExerisDomain source file for one mutation, applied in memory. Returns { result, sourcePath, diff }: a unified diff and the SDK outcome in result.outcome — SUCCESS, CONFLICT (the source drifted from the baseline; the result carries baseline, current and intended values), VALIDATION_ERROR, or NO_BASELINE. Nothing is written: apply the diff with your own file tools. For a stale-free preview, pass the entity's last-emitted metadata as baselineJson and the sourceDigest from lsp-describe_domain as concurrencyToken.
Inputs: op (required), qualifiedName (required), baselineJson, concurrencyToken.
sdk-describe_annotationA
Describe a specific Exeris SDK annotation: simple and qualified name, owning package rationale, targets, retention, purpose, detailed description, all attributes (with types, defaults, required-ness, descriptions, and deprecations), and annotation-level deprecation status.
Inputs: annotation (required).
sdk-get_ast_schemaA
Get JSON Schema for the Exeris build-time DomainMetadata AST and sibling records. If 'definition' is omitted, returns the top-level schema envelope, reader requirements (such as Jackson 3 constraints), and definition list. If 'definition' is specified (e.g. 'DomainMetadata', 'FieldMetadata', 'ActionMetadata'), returns the schema definition for that specific record.
Inputs: definition.
sdk-get_scoping_rulesB
Get canonical authoring contract scoping rules for Exeris SDK annotations (including @Field vs @Validation split, FieldMetadata as single AST carrier per ADR-054, derived SQL NOT NULL and API not-blank generator semantics, form lifecycle scoping, deprecations, and sibling vs nested form trap). Supports optional filtering by rule ID or query string.
Inputs: rule.
sdk-list_annotationsA
List Exeris SDK annotations from the authoring contract: name, qualified name, package, targets, retention, and purpose. Supports optional filtering by package (e.g. 'system', 'security', 'capability', or full package name), target (e.g. 'METHOD', 'TYPE', 'FIELD'), or query string.
Inputs: query, target, package.
sdk-list_deprecationsA
List all deprecated Exeris SDK annotations and attributes across the authoring contract, including replacement guidance, since version, and forRemoval status. Supports filtering by package or forRemoval only.
Inputs: package, forRemovalOnly.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 34 tools
The seven family prefixes (docs-, lsp-, build-, kernel-, caps-, sdk-, bridge-) draw clean lines between domains, and most tools within a family target a distinct artefact. However, `lsp-list_domains` vs `build-list_domains` and `lsp-describe_domain` vs `build-get_domain_metadata` genuinely overlap — both surface fields, actions and relationships for an @ExerisDomain — and only long descriptions (live LSP session vs emitted target/classes metadata) separate them. `lsp-preview_mutation` and `build-preview_generation` are also adjacent preview concepts, though clearly differentiated.
Every tool follows a strict `family-verb_noun` pattern in snake_case (docs-get_adr, build-explain_artefacts, sdk-list_annotations, lsp-preview_mutation). The rare exceptions (docs-search, bridge-version, bridge-health) are still readable and consistent with the hyphen-prefix convention, so tool selection by name alone is predictable.
34 tools is heavy and sits above the comfortable band, but the surface is unusually broad (docs, live LSP, build/codegen, kernel runtime, capabilities, SDK contract, meta) and each family carries a small, coherent 2-9 tool set that mostly earns its place. It is still at the upper edge where an agent must first pick a family before picking a tool.
Coverage across inspection domains is strong: docs reading, ADR registry, domain/action introspection, mutation preview, build metadata, artefact and diagnostic explanation, kernel state, capability composition, SDK contract and health checks. The notable gap is that the docs root has `docs-search` and a few named fetchers but no generic `docs-list_docs`/`docs-get_doc` by path, and mutations are preview-only with no post-apply verification tool.
Maintenance
Related MCP Connectors
Read-only MCP tools for AI agent discovery, structured resources, and NIULAI information.
Read-only Remote MCP for externally grounded AI agent trust receipts.
Read-only MCP for AI usage profiles, leaderboards, stats, and docs; no writes or private data.
Read-only discovery of ChatGXT's integration catalog and public MCP capabilities.
Related MCP Servers
- AlicenseAqualityBmaintenanceRead-only MCP server that exposes the agentic-os governance, SDLC, and Quality Engineering methodology to any MCP host. It never writes to your repository and never executes code — it serves the methodology, plans an install, and verifies it, handing any commands back to the host to run.72Apache 2.0
- FlicenseDqualityBmaintenanceProvides read-only, structured access to Yasin ecosystem information (project registry, documentation, GitHub state, diagnostics) through the MCP protocol, enabling AI agents to query without direct repository access.223-
- AlicenseNot gradedqualityCmaintenanceEnables AI clients to observe local codebases in real time through a secure, read-only MCP interface, providing workspace snapshots, git history and diffs, cross-project search, and project reality coverage without write access.MIT
- AlicenseAqualityBmaintenanceEnables MCP-capable coding agents to access SkeletIQ architecture releases, including reading designs and build orders, generating and critiquing architectures, and reporting build progress and drift.6385 npm1MIT