Add an MCP to TMLL by MatthewKhouzam · Pull Request #15 · eclipse-tmll/tmll

MatthewKhouzam · 2026-03-10T14:49:36Z

What it does

This PR adds Model Context Protocol (MCP) server integration to TMLL, enabling the trace analysis library to be used as an MCP tool in AI assistants and other MCP-compatible applications.

Key additions:

tmll_cli.py: A command-line interface that exposes all TMLL functionality (experiment management, anomaly detection, memory leak detection, change point analysis, correlation analysis, idle resource detection, capacity planning, and clustering)
mcp_server_cli.py: An MCP server wrapper that exposes all CLI commands as MCP tools, allowing AI assistants to interact with TMLL programmatically.

How to test

Test the CLI directly:

bash

Start a trace server instance

./tracecompass-server -data /path/to/workspace -vmargs -Dtraceserver.port=8080

Create an experiment

python3 tmll_cli.py create /path/to/trace -n test_experiment

List experiments

python3 tmll_cli.py list

Run anomaly detection

python3 tmll_cli.py anomaly <experiment_uuid> -k "cpu usage" -m iforest

Test the MCP server:
bash

Run the MCP server

python3 mcp_server_cli.py tmll_cli.py

Configure in an MCP-compatible client (e.g., Kiro CLI, Claude Desktop)

Test tool invocations through the client
Verify all ML modules work through CLI:
- Test each subcommand: anomaly, memory-leak, changepoint, correlation, idle-resources, capacity, cluster
- Verify output formatting and error handling

Follow-ups

Review checklist

As an author, I have thoroughly tested my changes and carefully followed the instructions in this template

Implement tmll_cli.py with commands for experiment management and ML analysis: - Experiment operations: create, list, delete, list-outputs, fetch-data - Anomaly detection: anomaly, memory-leak - Performance analysis: changepoint, correlation - Resource optimization: idle-resources, capacity planning - Clustering analysis Note: this code was created with the assistance of claude sonnet 4.5 Signed-off-by: Matthew Khouzam <matthew.khouzam@ericsson.com>

Add the following MCP mcp.json { "mcpServers": { "tmll": { "command": "/usr/bin/python3", "args": "path-to/tmll/mcp_server_cli.py"], "env": { "PYTHONPATH": "path-to/tmll" } } } } This code creation was assisted by claude-sonnet-4.5 Signed-off-by: Matthew Khouzam <matthew.khouzam@ericsson.com>

kavehshahedi

Thanks a lot Matthew for the MCP! I tested it with Claude Code (with some minor changes I put comments for), and it was working very well!

Vibe coders are now ready to become tracing masters, right?

kavehshahedi · 2026-04-03T17:24:02Z

tmll_cli.py

+    client = TMLLClient(args.host, args.port, verbose=args.verbose)
+    traces = [{"path": os.path.expanduser(path)} for path in args.traces]
+    experiment = client.create_experiment(traces=traces, experiment_name=args.name)
+    print(f"Created experiment: {experiment.name} (UUID: {experiment.UUID})")


The experiment.UUID should be experiment.uuid

tmll/mcp/cli.py

kavehshahedi · 2026-04-03T17:26:31Z

tmll_cli.py

+    outputs = experiment.find_outputs(keyword=args.keywords, type=['xy'])
+
+    if not outputs:
+        print("No outputs found")
+        return
+
+    mld = MemoryLeakDetection(client, experiment, outputs)


The memory leak detection module doesn't accept outputs. So, you should just use:

mld = MemoryLeakDetection(client, experiment)

kavehshahedi · 2026-04-03T17:27:00Z

tmll_cli.py

+        return
+
+    mld = MemoryLeakDetection(client, experiment, outputs)
+    result = mld.detect_memory_leak()


It should be mld.analyze_memory_leaks()

kavehshahedi · 2026-04-03T17:28:42Z

tmll_cli.py

+        return
+
+    ca = CorrelationAnalysis(client, experiment, outputs)
+    correlations = ca.analyze_correlation(method=args.method)


This should be: ca.analyze_correlations(method=args.method)

kavehshahedi · 2026-04-03T17:46:30Z

tmll_cli.py

+    if args.output:
+        for key, df in data.items():
+            df.to_csv(f"{args.output}_{key}.csv", index=False)


This might fail for cases where the data is not a dictionary of <key, df>. There're some cases that the output may be dict<key, dict<key, df>>. So, probably you should handle those cases.

kavehshahedi · 2026-04-03T17:48:34Z

tmll_cli.py

+
+    # anomaly command
+    anomaly_parser = subparsers.add_parser("anomaly", help="Detect anomalies")
+    anomaly_parser.add_argument("experiment", help="Experiment UUID or name")


I guess here we should only pass the UUID, and not the name. Same comment for the other commands below.

kavehshahedi · 2026-04-03T17:51:30Z

mcp_server_cli.py

+
+
+@server.call_tool()
+async def call_tool(name: str, arguments: dict) -> list[TextContent]:


As I checked, I guess some commands may send None for arguments. So, it might be better to make it Optional[dict] = None, and then, check for it:

arguments = arguments if isinstance(arguments, dict) else {}

so we don't get an exception when trying to .get() from it.

kavehshahedi · 2026-04-03T17:53:32Z

mcp_server_cli.py

+
+server = Server("tmll-cli-mcp-server")
+
+CLI_PATH = sys.argv[1] if len(sys.argv) > 1 else "tmll_cli.py"


If the path is not given, the "tmll_cli.py" is relative to the cwd. So, probably it's better to handle it.

CLI_PATH = sys.argv[1] if len(sys.argv) > 1 else Path(__file__).resolve().parent / "tmll_cli.py"

kavehshahedi · 2026-04-03T17:54:27Z

mcp_server_cli.py

+        Tool(
+            name="cluster_data",
+            description="Perform clustering analysis on trace data (kmeans, dbscan, hierarchical)",
+            inputSchema={
+                "type": "object",
+                "properties": {
+                    "experiment_id": {"type": "string"},
+                    "keywords": {"type": "array", "items": {"type": "string"}, "default": ["cpu usage"]},
+                    "n_clusters": {"type": "integer", "default": 3},
+                    "method": {"type": "string", "default": "kmeans", "enum": ["kmeans", "dbscan", "hierarchical"]},
+                },
+                "required": ["experiment_id"],
+            },
+        ),
+    ]


Please checkout my comment for the clustering module.

kavehshahedi · 2026-04-03T18:04:08Z

Also, I think you should add the mcp package to the "requirements.txt" in order to run the MCP script:

mcp==1.27.0

And, could you please place the scripts in a proper package within the tmll src files? Maybe tmll/mcp?

- Fix experiment.UUID -> experiment.uuid - Add None check for experiment in create_experiment - Fix MemoryLeakDetection: remove outputs param, use analyze_memory_leaks() - Fix ChangePointAnalysis: method -> methods (list of analysis modes) - Fix CorrelationAnalysis: analyze_correlation -> analyze_correlations, plot_correlation -> plot_correlation_matrix - Fix IdleResourceDetection: single threshold -> per-resource thresholds - Fix CapacityPlanning: plan_capacity -> forecast_capacity(forecast_steps=) - Remove clustering command (module not meaningful) - Fix help text: 'UUID or name' -> 'UUID' - Handle nested dict data in fetch_data_cmd - Move scripts to tmll/mcp package - Fix MCP server CLI_PATH to use Path(__file__) for reliable resolution - Make MCP call_tool arguments Optional[dict] with None guard - Add mcp==1.27.0 to requirements.txt

When the trace server returns a non-200 status for both datatree and timegraph tree endpoints, response.model is None. Accessing .model on None caused an AttributeError, failing CI tests.

kavehshahedi

LGTM! Thanks for the fixes! Feel free to merge whenever you want!

MatthewKhouzam requested a review from kavehshahedi March 10, 2026 14:49

MatthewKhouzam added 2 commits March 10, 2026 11:00

kavehshahedi requested changes Apr 3, 2026

View reviewed changes

MatthewKhouzam force-pushed the mcp branch from 1f36f13 to e76ab3f Compare April 7, 2026 15:57

[FIX] Handle None response model in fetch_outputs_with_tree

fbf6d83

When the trace server returns a non-200 status for both datatree and timegraph tree endpoints, response.model is None. Accessing .model on None caused an AttributeError, failing CI tests.

MatthewKhouzam force-pushed the mcp branch from ded73b1 to fbf6d83 Compare April 7, 2026 19:58

kavehshahedi self-requested a review April 8, 2026 01:48

kavehshahedi approved these changes Apr 8, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add an MCP to TMLL#15

Add an MCP to TMLL#15
MatthewKhouzam wants to merge 4 commits intoeclipse-tmll:mainfrom
MatthewKhouzam:mcp

MatthewKhouzam commented Mar 10, 2026

Uh oh!

kavehshahedi left a comment

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi Apr 3, 2026

Uh oh!

kavehshahedi commented Apr 3, 2026

Uh oh!

kavehshahedi left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants



		@server.call_tool()
		async def call_tool(name: str, arguments: dict) -> list[TextContent]:


		server = Server("tmll-cli-mcp-server")

		CLI_PATH = sys.argv[1] if len(sys.argv) > 1 else "tmll_cli.py"

Conversation

MatthewKhouzam commented Mar 10, 2026

What it does

How to test

Start a trace server instance

Create an experiment

List experiments

Run anomaly detection

Run the MCP server

Configure in an MCP-compatible client (e.g., Kiro CLI, Claude Desktop)

Test tool invocations through the client

Follow-ups

Review checklist

Uh oh!

kavehshahedi left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kavehshahedi commented Apr 3, 2026

Uh oh!

kavehshahedi left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants