mps-mcp-build

name	mps-mcp-build
description	Create and deploy MCP tools from MATLAB functions to MATLAB Production Server using the MCP Framework for MATLAB Production Server support package
argument-hint	<function path(s), server address, optional files/resources>
disable-model-invocation	true
allowed-tools	Read Bash(matlab *) Bash(which matlab) Bash(ls *) Bash(grep *) Bash(sha256sum *) Bash(cat *) Bash(curl *) Bash(mkdir *) Bash(rm *) Bash(cp *)

Create MCP Tool from MATLAB Function

You are helping the user create and deploy MCP tools from MATLAB functions to a MATLAB Production Server using the MCP Framework for MATLAB Production Server support package.

Step 1: Check Prerequisites

Before doing anything else, verify these prerequisites are met:

1. MATLAB availability: Check if MATLAB is accessible by running which matlab. Alternatively, check if the MATLAB MCP Server is available as a configured MCP tool.
2. MCP Framework availability: This is verified automatically as part of the combined build call in Step 5. Do NOT run a separate matlab -batch just to check for the support package.

If MATLAB is not available, report clearly what is needed and stop. Do NOT attempt to download, install, or otherwise acquire MATLAB or the support package.

Prefer using the MATLAB MCP Server if it is available as a configured MCP tool. Fall back to matlab -batch "..." on the command line otherwise.

Step 2: Parse User Intent

Extract the following from the user's request:

• Function path(s): One or more paths to .m files (required)
• Server address: Host and port of the MATLAB Production Server (required), e.g. localhost:9910
• Scheme: http (default) or https if the user specifies it
• Additional files: Extra files to include in the CTF archive (optional)
• Resources: MCP resources to include (optional), specified in flexible formats such as:
  • "add a resource at config://settings with the contents from settings.json"
  • "create the resource data://AuthorName with string value 'Mr. Author'"
  • "add the following resources: config://settings,settings.json, data://AuthorName,'Mr. Author'"
• Force rebuild: Whether the user explicitly wants a full rebuild regardless of input state. Triggered by keywords like "force", "rebuild", "clean", "from scratch", or "fresh".

For resources, construct a MATLAB struct array where each element has:

• uri: the resource URI (e.g. "config://settings")
• contents: the resource contents — either a file path to read from or a literal string value

Step 3: Validate File Paths (Case-Sensitive)

For all file paths provided by the user — including function .m files and any optional additional files — verify that each filename matches a file on disk exactly, including case. Windows filesystems are case-insensitive, so a path may resolve even when the case is wrong — but MATLAB function names are case-sensitive and must match the file exactly, and additional files must also be referenced by their exact on-disk name to avoid deployment issues.

For each path:

1. Check whether the path resolves at all:

   ls "<full_path>" &>/dev/null
   

2. If the path does not resolve, report an error: "File not found: <path>" and stop.
3. If the path resolves, verify the filename case by listing the parent directory and checking for an exact (case-sensitive, fixed-string) match:

   ls "<parent_dir>" | grep -Fx "<filename>"
   

4. If grep -Fx finds no match, the file exists with different case. Find the actual name:

   ls "<parent_dir>" | grep -iFx "<filename>"
   

   Then ask the user: "The file on disk is named <actual_name>, but you specified <user_name>. Would you like to use <actual_name> instead?" Wait for the user to confirm before proceeding. If they decline, stop.
5. If grep -Fx succeeds, the case matches — proceed.

Only continue to Step 4 once all file paths have been validated.

Step 4: Incremental Build Check

Skip this step if the user requested a force rebuild.

Before preparing the working directory, check whether a rebuild is necessary by comparing current inputs against a stored build manifest.

4.1: Read the manifest

Check for an existing manifest in the deployment folder:

cat "<deploy-folder>/.mcp-build-manifest.json" 2>/dev/null

If the manifest does not exist, proceed to Step 5 (full build).

4.2: Compute current input hashes

Compute SHA-256 hashes for all source files and additional files:

sha256sum "<source.m>" | cut -d' ' -f1

For the framework fingerprint, compute it dynamically from the current state of all framework .m files. The manifest stores the framework root path from the previous build:

cd "<manifest.framework_root>" && find +prodserver -name "*.m" -print0 | sort -z | xargs -0 sha256sum | sort -k2 | sha256sum | cut -d' ' -f1

If the command fails (e.g. the framework root no longer exists), treat as "framework changed" and proceed to full build.

4.3: Compare and decide

Compare current values against the manifest:

Current vs Manifest	Action
All source hashes match AND framework fingerprint matches AND server address matches	Skip — report "Build is up to date. No changes detected." and jump to Step 7 (register/verify only)
All source hashes match AND framework fingerprint matches AND server address changed	Deploy only — upload existing CTF to new server (Step 5b)
Any source hash differs OR framework fingerprint differs	Full rebuild — proceed to Step 5

4.4: Verify outputs still exist

Even when all hashes match, verify that the CTF file still exists:

ls "<deploy-folder>/<archive>.ctf" &>/dev/null

If the CTF is missing, proceed to full build regardless of hash match.

4.5: Verify server is responding (skip path only)

For the "skip" path, verify the archive is actually deployed and responding. MCP endpoints only accept POST requests (GET returns 405), so use a JSON-RPC ping:

curl -s -o /dev/null -w "%{http_code}" -X POST \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"ping"}' \
  "<scheme>://<host>:<port>/<archiveName>/mcp"

If the server returns a non-2xx status, switch to the deploy only path (Step 5b).

If curl fails entirely (connection refused), report that the build is up to date but the server is unreachable, and suggest the user check that MATLAB Production Server is running.

Step 5: Build and Deploy

Construct and execute a single matlab -batch call that combines the prerequisite check with the build, deploy, and manifest generation. This avoids paying the MATLAB startup cost multiple times.

Arguments to prodserver.mcp.build:

• fcn: Function name(s) derived from the file path(s) (filename without .m extension)
• server: Full server URL as "<scheme>://<host>:<port>" (default scheme is http)
• files: String array of additional file paths (only if user specified them)
• resource: Struct array of resources (only if user specified them)
• folder: The deployment folder (same directory the function was copied into)

Important:

• Do NOT specify the wrapper argument — accept the framework's default behavior
• Do NOT read the MATLAB function or make any judgements about wrappers
• Always cd into the deployment folder before calling build so generated wrappers are co-located with the source function
• Always use a deploy subfolder for build output. The folder argument to prodserver.mcp.build should be <project-dir>/deploy. This keeps generated artifacts (CTF, wrappers, signatures) separate from the source .m files.
• All paths in matlab -batch scripts MUST be absolute paths — after cd changes the working directory, relative paths will resolve incorrectly. This applies to the cd target, the folder argument, file paths, and any other path references.
• Always addpath the project root (the directory containing the +prodserver package folder) before cd-ing into the deployment folder. Without this, cd moves away from the package and MATLAB can no longer resolve prodserver.mcp.* functions.
• Use a script file for complex builds — when passing string arrays or complex MATLAB expressions, write the command to a temporary .m file and run it with matlab -batch "run('<path>')". This avoids shell quoting issues that silently corrupt double-quoted MATLAB strings in -batch mode on Windows.

Prepare working directory

Before running the MATLAB build command, prepare the deployment folder:

1. Create the deployment folder if it does not exist.
2. Delete any pre-existing MCP wrapper files from the deployment folder. For each function foo, remove fooMCP.m if it exists. The wrapper name is case-sensitive: for function foo the wrapper is fooMCP.m, for Foo it is FooMCP.m. Remove only the exact match.
3. Delete any existing copy of the function file in the deployment folder before copying. On Windows (case-insensitive filesystem), cp will silently keep the case of an existing file rather than using the source's case. Always rm -f <folder>/<filename> first, then copy. This ensures the destination file has the exact case of the source.
4. Copy the specified .m function file(s) into the deployment folder.

This ensures the framework generates fresh wrappers without conflicts from stale cached files.

Combined script structure

The script uses section markers (===STEP:name===, ===OK:name===, ===FAIL:name===) so that you can parse the output and identify exactly which stage failed:

matlab -batch "
addpath('<project-root>');
fprintf('===STEP:prerequisite===\n');
try
    help prodserver.mcp.build;
    fprintf('===OK:prerequisite===\n');
catch e
    fprintf('===FAIL:prerequisite===\n%s\n', e.message);
    exit(1);
end

fprintf('===STEP:build===\n');
try
    cd('<absolute-folder-path>');
    [ctf, endpoint] = prodserver.mcp.build('<fcnName>', server='<scheme>://<host>:<port>', folder='<absolute-folder-path>');
    fprintf('===OK:build===\n');
    fprintf('CTF: %s\nEndpoint: %s\n', ctf, endpoint);
catch e
    fprintf('===FAIL:build===\n%s\n', e.message);
    exit(1);
end

fprintf('===STEP:manifest===\n');
try
    fw_root = prodserver.mcp.internal.packageFolder();
    fp_file = fullfile(fw_root, '+prodserver', '+mcp', '+internal', 'frameworkFingerprint.txt');
    if isfile(fp_file)
        fw_fp = strtrim(fileread(fp_file));
    else
        fw_fp = 'unavailable';
    end
    manifest = struct();
    manifest.version = 1;
    manifest.timestamp = string(datetime('now', 'TimeZone', 'UTC', 'Format', 'yyyy-MM-dd''T''HH:mm:ss''Z'''));
    manifest.server = '<scheme>://<host>:<port>';
    manifest.archive = '<archiveName>';
    manifest.framework_root = fw_root;
    manifest.framework_fingerprint = fw_fp;
    manifest.sources = struct('<fcnName>', '<source-hash>');
    manifest.additional_files = struct('file', {}, 'hash', {});
    manifest.outputs = struct('ctf', ctf, 'endpoint', endpoint);
    manifest.deployed = true;
    json = jsonencode(manifest, PrettyPrint=true);
    writelines(json, fullfile('<absolute-folder-path>', '.mcp-build-manifest.json'));
    fprintf('===OK:manifest===\n');
catch e
    fprintf('===FAIL:manifest===\n%s\n', e.message);
end
"

Where <project-root> is the absolute path to the directory containing the +prodserver package (i.e. the root of this repository), <absolute-folder-path> is the absolute path to the deployment folder, and <source-hash> is the SHA-256 hash computed in Step 4 (or computed fresh if this is the first build).

Manifest notes:

• The sources struct uses function names as field names and their SHA-256 hashes as values. Function names are always valid MATLAB identifiers, so this is safe.
• The additional_files field is a struct array with file and hash fields. Each element stores the original filename (preserving exact case and special characters) and its SHA-256 hash. When there are no additional files, use an empty struct array: struct('file', {}, 'hash', {}). When there are files, build the array element-by-element:

  manifest.additional_files(1).file = 'data.mat';
  manifest.additional_files(1).hash = '<hash1>';
  manifest.additional_files(2).file = 'config.json';
  manifest.additional_files(2).hash = '<hash2>';
  

• The ===FAIL:manifest=== is non-fatal — the build succeeded even if the manifest write fails. Report a warning but do not retry.

Output parsing

After the command completes, parse the output:

• If ===FAIL:prerequisite=== is present → the MCP Framework support package is not installed or not on the path. Report this to the user and stop.
• If ===OK:prerequisite=== is present but ===FAIL:build=== is present → the build failed. The error message follows the FAIL marker. Diagnose and report.
• If ===OK:build=== is present → extract the CTF path and endpoint URL from the lines following the OK marker.
• If ===FAIL:manifest=== is present → warn the user that the build manifest could not be written (next build will not be incremental).

Build success criteria

The build is successful only when both of the following are true:

1. No errors: The matlab -batch command exits with code 0 and output contains ===OK:build===.
2. CTF is newer than wrapper: The generated .ctf archive must have a modification timestamp strictly later than the generated MCP wrapper file (<fcnName>MCP.m). If the wrapper exists but the CTF is older or missing, the build did not complete — a stale CTF from a previous build is not acceptable.

If either criterion fails, diagnose and retry the build (e.g., delete stale artifacts and rebuild from scratch).

If the build succeeds but the deploy step within build fails (e.g., cannot find the archive), deploy separately using prodserver.mcp.deploy with the full path to the generated .ctf file. Include this as a second try/catch block appended to the script:

fprintf('===STEP:deploy===\n');
try
    endpoint = prodserver.mcp.deploy('<absolute-folder-path>/<fcnName>.ctf', '<host>', <port>, scheme='<scheme>');
    fprintf('===OK:deploy===\n');
    fprintf('Endpoint: %s\n', endpoint);
catch e
    fprintf('===FAIL:deploy===\n%s\n', e.message);
    exit(1);
end

Step 5b: Deploy Only (existing CTF, new server)

When the incremental check determines only the server address changed, skip the full build and deploy the existing CTF directly:

matlab -batch "
addpath('<project-root>');
fprintf('===STEP:deploy===\n');
try
    endpoint = prodserver.mcp.deploy('<absolute-folder-path>/<archiveName>.ctf', '<host>', <port>, scheme='<scheme>');
    fprintf('===OK:deploy===\n');
    fprintf('Endpoint: %s\n', endpoint);
catch e
    fprintf('===FAIL:deploy===\n%s\n', e.message);
    exit(1);
end

fprintf('===STEP:manifest===\n');
try
    fw_root = prodserver.mcp.internal.packageFolder();
    fp_file = fullfile(fw_root, '+prodserver', '+mcp', '+internal', 'frameworkFingerprint.txt');
    if isfile(fp_file)
        fw_fp = strtrim(fileread(fp_file));
    else
        fw_fp = 'unavailable';
    end
    manifest = struct();
    manifest.version = 1;
    manifest.timestamp = string(datetime('now', 'TimeZone', 'UTC', 'Format', 'yyyy-MM-dd''T''HH:mm:ss''Z'''));
    manifest.server = '<scheme>://<host>:<port>';
    manifest.archive = '<archiveName>';
    manifest.framework_root = fw_root;
    manifest.framework_fingerprint = fw_fp;
    manifest.sources = struct('<fcnName>', '<source-hash>');
    manifest.additional_files = struct('file', {}, 'hash', {});
    manifest.outputs = struct('ctf', '<absolute-folder-path>/<archiveName>.ctf', 'endpoint', endpoint);
    manifest.deployed = true;
    json = jsonencode(manifest, PrettyPrint=true);
    writelines(json, fullfile('<absolute-folder-path>', '.mcp-build-manifest.json'));
    fprintf('===OK:manifest===\n');
catch e
    fprintf('===FAIL:manifest===\n%s\n', e.message);
end
"

Example MATLAB commands:

In all examples below, <project-root> is the absolute path to the directory containing the +prodserver package, and <folder> is the absolute path to the deployment folder.

Single function, no extras:

matlab -batch "addpath('<project-root>'); fprintf('===STEP:prerequisite===\n'); try, help prodserver.mcp.build; fprintf('===OK:prerequisite===\n'); catch e, fprintf('===FAIL:prerequisite===\n%s\n', e.message); exit(1); end; fprintf('===STEP:build===\n'); try, cd('<folder>'); [ctf, endpoint] = prodserver.mcp.build('fcnName', server='http://localhost:9910', folder='<folder>'); fprintf('===OK:build===\n'); fprintf('CTF: %s\nEndpoint: %s\n', ctf, endpoint); catch e, fprintf('===FAIL:build===\n%s\n', e.message); exit(1); end; fprintf('===STEP:manifest===\n'); try, fw_root = prodserver.mcp.internal.packageFolder(); fp_file = fullfile(fw_root, '+prodserver', '+mcp', '+internal', 'frameworkFingerprint.txt'); if isfile(fp_file), fw_fp = strtrim(fileread(fp_file)); else, fw_fp = 'unavailable'; end; manifest = struct(); manifest.version = 1; manifest.timestamp = string(datetime('now', 'TimeZone', 'UTC', 'Format', 'yyyy-MM-dd''T''HH:mm:ss''Z''')); manifest.server = 'http://localhost:9910'; manifest.archive = 'fcnName'; manifest.framework_root = fw_root; manifest.framework_fingerprint = fw_fp; manifest.sources = struct('fcnName', '<hash>'); manifest.additional_files = struct('file', {}, 'hash', {}); manifest.outputs = struct('ctf', ctf, 'endpoint', endpoint); manifest.deployed = true; json = jsonencode(manifest, PrettyPrint=true); writelines(json, fullfile('<folder>', '.mcp-build-manifest.json')); fprintf('===OK:manifest===\n'); catch e, fprintf('===FAIL:manifest===\n%s\n', e.message); end"

With additional files:

matlab -batch "addpath('<project-root>'); fprintf('===STEP:prerequisite===\n'); try, help prodserver.mcp.build; fprintf('===OK:prerequisite===\n'); catch e, fprintf('===FAIL:prerequisite===\n%s\n', e.message); exit(1); end; fprintf('===STEP:build===\n'); try, cd('<folder>'); [ctf, endpoint] = prodserver.mcp.build('fcnName', server='http://localhost:9910', folder='<folder>', files=[""data.mat"", ""config.json""]); fprintf('===OK:build===\n'); fprintf('CTF: %s\nEndpoint: %s\n', ctf, endpoint); catch e, fprintf('===FAIL:build===\n%s\n', e.message); exit(1); end; fprintf('===STEP:manifest===\n'); try, fw_root = prodserver.mcp.internal.packageFolder(); fp_file = fullfile(fw_root, '+prodserver', '+mcp', '+internal', 'frameworkFingerprint.txt'); if isfile(fp_file), fw_fp = strtrim(fileread(fp_file)); else, fw_fp = 'unavailable'; end; manifest = struct(); manifest.version = 1; manifest.timestamp = string(datetime('now', 'TimeZone', 'UTC', 'Format', 'yyyy-MM-dd''T''HH:mm:ss''Z''')); manifest.server = 'http://localhost:9910'; manifest.archive = 'fcnName'; manifest.framework_root = fw_root; manifest.framework_fingerprint = fw_fp; manifest.sources = struct('fcnName', '<hash>'); manifest.additional_files(1).file = 'data.mat'; manifest.additional_files(1).hash = '<hash1>'; manifest.additional_files(2).file = 'config.json'; manifest.additional_files(2).hash = '<hash2>'; manifest.outputs = struct('ctf', ctf, 'endpoint', endpoint); manifest.deployed = true; json = jsonencode(manifest, PrettyPrint=true); writelines(json, fullfile('<folder>', '.mcp-build-manifest.json')); fprintf('===OK:manifest===\n'); catch e, fprintf('===FAIL:manifest===\n%s\n', e.message); end"

With resources:

matlab -batch "addpath('<project-root>'); fprintf('===STEP:prerequisite===\n'); try, help prodserver.mcp.build; fprintf('===OK:prerequisite===\n'); catch e, fprintf('===FAIL:prerequisite===\n%s\n', e.message); exit(1); end; fprintf('===STEP:build===\n'); try, cd('<folder>'); r(1).uri='config://settings'; r(1).contents='settings.json'; [ctf, endpoint] = prodserver.mcp.build('fcnName', server='http://localhost:9910', folder='<folder>', resource=r); fprintf('===OK:build===\n'); fprintf('CTF: %s\nEndpoint: %s\n', ctf, endpoint); catch e, fprintf('===FAIL:build===\n%s\n', e.message); exit(1); end; fprintf('===STEP:manifest===\n'); try, fw_root = prodserver.mcp.internal.packageFolder(); fp_file = fullfile(fw_root, '+prodserver', '+mcp', '+internal', 'frameworkFingerprint.txt'); if isfile(fp_file), fw_fp = strtrim(fileread(fp_file)); else, fw_fp = 'unavailable'; end; manifest = struct(); manifest.version = 1; manifest.timestamp = string(datetime('now', 'TimeZone', 'UTC', 'Format', 'yyyy-MM-dd''T''HH:mm:ss''Z''')); manifest.server = 'http://localhost:9910'; manifest.archive = 'fcnName'; manifest.framework_root = fw_root; manifest.framework_fingerprint = fw_fp; manifest.sources = struct('fcnName', '<hash>'); manifest.additional_files = struct('file', {}, 'hash', {}); manifest.outputs = struct('ctf', ctf, 'endpoint', endpoint); manifest.deployed = true; json = jsonencode(manifest, PrettyPrint=true); writelines(json, fullfile('<folder>', '.mcp-build-manifest.json')); fprintf('===OK:manifest===\n'); catch e, fprintf('===FAIL:manifest===\n%s\n', e.message); end"

Multiple functions:

matlab -batch "addpath('<project-root>'); fprintf('===STEP:prerequisite===\n'); try, help prodserver.mcp.build; fprintf('===OK:prerequisite===\n'); catch e, fprintf('===FAIL:prerequisite===\n%s\n', e.message); exit(1); end; fprintf('===STEP:build===\n'); try, cd('<folder>'); [ctf, endpoint] = prodserver.mcp.build([""fcn1"", ""fcn2""], server='http://localhost:9910', folder='<folder>'); fprintf('===OK:build===\n'); fprintf('CTF: %s\nEndpoint: %s\n', ctf, endpoint); catch e, fprintf('===FAIL:build===\n%s\n', e.message); exit(1); end; fprintf('===STEP:manifest===\n'); try, fw_root = prodserver.mcp.internal.packageFolder(); fp_file = fullfile(fw_root, '+prodserver', '+mcp', '+internal', 'frameworkFingerprint.txt'); if isfile(fp_file), fw_fp = strtrim(fileread(fp_file)); else, fw_fp = 'unavailable'; end; manifest = struct(); manifest.version = 1; manifest.timestamp = string(datetime('now', 'TimeZone', 'UTC', 'Format', 'yyyy-MM-dd''T''HH:mm:ss''Z''')); manifest.server = 'http://localhost:9910'; manifest.archive = 'fcn1'; manifest.framework_root = fw_root; manifest.framework_fingerprint = fw_fp; manifest.sources = struct('fcn1', '<hash1>', 'fcn2', '<hash2>'); manifest.additional_files = struct('file', {}, 'hash', {}); manifest.outputs = struct('ctf', ctf, 'endpoint', endpoint); manifest.deployed = true; json = jsonencode(manifest, PrettyPrint=true); writelines(json, fullfile('<folder>', '.mcp-build-manifest.json')); fprintf('===OK:manifest===\n'); catch e, fprintf('===FAIL:manifest===\n%s\n', e.message); end"

Step 6: Verify Deployment

After a successful build and deploy (or deploy-only), verify the deployment using a single matlab -batch call that combines ping and list. Use a 2-minute timeout for verification. Only use a longer timeout if the user explicitly requests one.

Skip this step if the incremental check determined the build is up to date AND the server responded to the curl check in Step 4.5.

matlab -batch "
addpath('<project-root>');
fprintf('===STEP:ping===\n');
try
    ok = prodserver.mcp.ping('<scheme>://<host>:<port>/<archiveName>/mcp');
    if ~ok, error('prodserver:mcp:pingFailed', 'Ping returned false'); end
    fprintf('===OK:ping===\n');
catch e
    fprintf('===FAIL:ping===\n%s\n', e.message);
    exit(1);
end

fprintf('===STEP:list===\n');
try
    tools = prodserver.mcp.list('<scheme>://<host>:<port>/<archiveName>/mcp', 'Tool');
    for i=1:numel(tools)
        fprintf('%s - %s\n', tools{i}.name, tools{i}.description);
    end
    fprintf('===OK:list===\n');
catch e
    fprintf('===FAIL:list===\n%s\n', e.message);
    exit(1);
end
"

Where <project-root> is the same absolute path used in Step 5.

Output parsing

• If ===FAIL:ping=== is present → the server is not responding. Retry once after 10 seconds (re-run the entire combined verification call). If it still fails, stop immediately — do NOT continue to Step 7 or Step 8.
• If ===OK:ping=== is present but ===FAIL:list=== is present → the archive loaded but tools couldn't be enumerated. Report the error message.
• If ===OK:list=== is present → extract tool names and descriptions from the lines between ===OK:ping=== and ===OK:list===.

On failure

If verification fails after the retry, report an error to the user explaining that the MCP server is not responding, provide the endpoint URL, and suggest they check that:

• The MATLAB Production Server is running
• The archive was loaded successfully (check MPS logs)
• The server address and port are correct

Step 7: Register MCP Server in Claude Code

After successful deployment verification, register the new MCP server so the tool is available to Claude.

By default, register at user scope in ~/.claude.json. If the user explicitly requests project-level registration, use .mcp.json in the project root instead.

Add an entry using the deployed endpoint URL. The server name should be the archive name (typically the function name for single-tool deployments).

User scope (default): ~/.claude.json

Read ~/.claude.json. If it does not exist, create it. If it already has an mcpServers object, add the new entry to it. If not, create the mcpServers key. The format is:

{
  "mcpServers": {
    "<serverName>": {
      "type": "http",
      "url": "<scheme>://<host>:<port>/<archiveName>/mcp"
    }
  }
}

For example, after deploying cleanSignal to http://localhost:9910:

{
  "mcpServers": {
    "cleanSignal": {
      "type": "http",
      "url": "http://localhost:9910/cleanSignal/mcp"
    }
  }
}

Project scope: .mcp.json

If the user requests project-level registration, read .mcp.json in the project root. If it does not exist, create it. Add the server entry under the mcpServers key using the same format as above.

Step 8: Report Results

Report to the user based on which path was taken:

Full build or deploy-only:

• The endpoint URL
• The tool name(s) and description(s) deployed
• That the MCP server has been registered in Claude Code settings
• Inform the user that they must restart Claude Code (or start a new session) for the new MCP tool to become available

Up-to-date (skipped):

• Report: "Build is up to date. No changes detected since last build at <manifest.timestamp>."
• Show the endpoint URL and confirm the server is responding
• If the MCP server is already registered, confirm it. If not, register it and inform about restart.

--- Copyright 2026 The MathWorks, Inc. ---