How to inspect a coding-agent session and a failed MCP call locally

Description and use

This workflow enables local inspection of coding‑agent sessions, identification of stalled tasks, and detailed tracing of failed MCP calls, enhancing debugging of coding‑agent workflows.

Problem

When multiple coding‑agent sessions run, it is hard to determine which session is waiting and what it last did.

Outcome

The workflow allows users to identify stalled sessions, inspect the last recorded action, and trace failed MCP calls, improving debugging of coding‑agent tasks.

Workflow

• Install Python 3.11+ and uv. Run uvx --from agent-observatory-cli agent-observatory. Press b to open the browser dashboard. Select a session under the Agents view to see runtime, ID, project, status, and last task. Open the MCP Calls view to inspect recorded calls, arguments, results, duration, and linked traces. Follow traces to see surrounding operations and error evidence. If the agent is unsupported, instrument it with the provided Python or JavaScript clients to export OTLP/HTTP telemetry.

Evidence

Agent Observatory is a local terminal and browser dashboard for inspecting supported Codex and Claude Code session records.

Source

The receiver listens on 127.0.0.1:14318 and stores captured data in a local SQLite file.

Source

If your agent is not a supported local runtime, send OTLP/HTTP telemetry or use the project's Python and JavaScript clients.

Source