Skip to main content

MCP Server

The log inspector ships an MCP server built on the official PHP MCP SDK (mcp/sdk). Any MCP client, such as Claude Code, Cursor, Claude Desktop or the MCP Inspector, can then search your logs, trace requests and run a root-cause investigation without you writing a line of PHP.

php bin/log-inspector-mcp      # STDIO transport

How it fits together​

 MCP client (Claude Code, Cursor, Inspector, bin/log-inspector-demo)
│ JSON-RPC over STDIO
▼
bin/log-inspector-mcp ── mcp/sdk Server (LogInspectorMcpServerFactory)
│
├── log_search ──────┐
├── request_context ─┼──► LogRetriever ──► vector store (Symfony AI Store)
├── ask_logs ──► LogInspectorAgent (Symfony AI Agent, tool-calling loop)
│ └── calls log_search / request_context itself
└── index_logs ──► LogLineParser ──► LogDocumentIndexer (embeddings) ──► vector store

The server is assembled from env vars by LogInspectorRuntime (no framework or config files needed), and registered with LogInspectorMcpServerFactory. Both live in src/Mcp/.

Tools​

ToolArgumentsReturns
log_searchquery (string)success, reason (AI explanation), evidence_logs[] (id, content, timestamp, level, source, tags), search_method, log_count
request_contextidentifier (a request_id, trace_id or session_id)success, reason, root_cause, evidence_logs[] in chronological order, request_timeline, services_involved, time_span
ask_logsquestion (string)answer: a report written by the server-side agent (summary, root cause, timeline, impact, evidence ids, actions)
index_logspath (a log file or a directory of *.log files)success, indexed_files (lines per file), total_indexed_lines

log_search, request_context and ask_logs are annotated read-only; index_logs is not.

log_search vs ask_logs​

  • log_search / request_context return the raw evidence. The client's model (for example Claude in Claude Code) reasons over it. This is the most flexible way to use the tools, with a strong client model.
  • ask_logs runs the full Symfony AI agent on the server. It picks its own tools, always checks for deployments and config changes before naming a root cause, and returns a finished report. Use it with lightweight clients, or when you want the same investigation regardless of which client asks.

Resource​

URIContent
logs://index/statsJSON: indexed sources and number of lines per source

Prompt​

NameArgumentsPurpose
investigate_incidentsymptomA step-by-step playbook that tells the client model to search the symptom, trace affected requests, look for upstream causes and changes, and answer in a fixed format citing log ids

Example session​

> Use log-inspector to find out why checkout failed this afternoon and which customers need a refund.

Called log-inspector 4 times

What happened
- 14:02–14:04 v2.14.0 went out, and the config change set DATABASE_POOL_MAX from 50 to 5 (deploy_0002)
- 14:18 onward All 5 connections in use, with up to 41 requests waiting; checkouts failed with 500/504 errors
- 14:25 Error rate 38%, 412 failed orders (checkout_api_0014)
- 14:47–14:49 Rolled back to v2.13.4 (deploy_0004, deploy_0005)

Customers to refund (charged in Stripe, order never confirmed)
| ORD-55201 | ch_3QxL9 | 249.00 | payment_service_0005 |
| ORD-55263 | ch_3QxM2 | 59.00 | payment_service_0006 |

This is the bundled incident demo. Next: connect your client and configure providers and log paths.

Limits​

  • The index is in memory. It is built when the server starts (LOG_INSPECTOR_PATHS) and extended with index_logs, and it is gone when the server stops.
  • Each file contributes at most 5,000 lines, and every non-empty line becomes one document.
  • The transport is STDIO. For a remote setup, run the server in Docker on the machine that has the logs.