Deep Agentic Core MCP — Claude Code MCP server: profiling and chaos
One MCP server for agentic workflow profiling, fault-injection experiments and AI Operations spec checks, so your client connects once.
claude mcp add deep-agentic-core-mcp -- uvx deep-agentic-core-mcpWhat Deep Agentic Core MCP does
The server folds several product surfaces into a single MCP identity: workflow inspection and profiling in the agenticlens style, resilience testing with fault injection in the agentic-chaos style, module-surface and supervision-readiness discovery in the agentic-sidecar style, plus authenticated read-only access to Control Tower 0.7.0. It does not reimplement any of that logic. This is a thin orchestration layer over the existing Python libraries, packaged through PyPI and meant to run as a stdio server.
Tools are grouped by prefix. lens.analyze_workflow returns recommendations with a provenance block naming the artifact, the workflow step, the evidence span IDs and the producer, while provenance_summary counts how many of those were actually found in the artifact you sent. lens.diagnose_run takes a trace Run plus optional baseline runs and returns failure-taxonomy diagnoses with confidence, the faulty step and agent, an incident timeline and a Markdown narrative. spec.validate_artifact validates against the schema set named by the artifact's spec_version and rejects markers it has no schema for.
One thing to know before installing: chaos.run_experiment executes real code, so the server is built for trusted local stdio use. The hosted HTTP endpoint at https://mcp.deepagentlabs.io/mcp does not execute code — provenance, conformance reports and chaos.export_aios_run behave the same there, but experiment runs stay local.
What it helps with
- profile an agentic workflow and get a structured telemetry summary
- trace a failed run back to the faulty step
- run a controlled chaos experiment against a target workflow
- compare a normal run with a chaos run
- validate an artifact against the AI Operations spec
Example prompts for the agent
- Profile
workflow.jsonand show me where the time goes - Diagnose the trace in
run.jsonand name the step that failed - Run a chaos experiment with the
stream_cutfault on this workflow
How to install Deep Agentic Core MCP
Platform CLI
claude mcp add deep-agentic-core-mcp -- uvx deep-agentic-core-mcpuvx runs the command named after the package; if the README names another one, use uvx --from deep-agentic-core-mcp <command>
Platform CLI
codex mcp add deep-agentic-core-mcp -- uvx deep-agentic-core-mcpuvx runs the command named after the package; if the README names another one, use uvx --from deep-agentic-core-mcp <command>
Config file
{
"mcpServers": {
"deep-agentic-core-mcp": {
"command": "uvx",
"args": [
"deep-agentic-core-mcp"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from deep-agentic-core-mcp <command>
Config file
{
"servers": {
"deep-agentic-core-mcp": {
"type": "stdio",
"command": "uvx",
"args": [
"deep-agentic-core-mcp"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from deep-agentic-core-mcp <command>
Platform CLI
gemini mcp add deep-agentic-core-mcp uvx deep-agentic-core-mcpuvx runs the command named after the package; if the README names another one, use uvx --from deep-agentic-core-mcp <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about Deep Agentic Core MCP
How do I install Deep Agentic Core MCP in Claude Code?
Run claude mcp add deep-agentic-core-mcp -- uvx deep-agentic-core-mcp. The server then starts as a stdio process and its tools show up in your session.
What does Deep Agentic Core MCP do?
It puts workflow profiling, chaos testing, Sidecar module discovery and Control Tower access behind one MCP server. Your client connects once and gets all four tool groups.
How is the local setup different from the hosted one?
On https://mcp.deepagentlabs.io/mcp, provenance, conformance reports and chaos.export_aios_run work the same as locally. Code execution is off there, so chaos.run_experiment only works over a local stdio connection.
