Agent Observatory
A local-first session debugging and observability tool for AI coding assistants. It provides terminal and browser dashboards for inspecting Codex and Claude Code session records, MCP calls and trace chains, with data stored in a local SQLite database.
Timeline (last 90 days)
Agent Observatory is a local terminal and browser dashboard tool designed specifically for inspecting session records of Codex and Claude Code, while also receiving OpenTelemetry data from custom instrumented Agents.
Using Agent Observatory requires Python 3.11+ and the uv package manager, and the startup command is uvx --from agent-observatory-cli agent-observatory.
Agent Observatory's receiver listens on 127.0.0.1:14318, captured data is stored in a local SQLite file, and session data does not leave the user's machine.
Only when the source explicitly reports a waiting state does Agent Observatory mark a session as waiting, and the dashboard does not infer unrecorded reasons.
Agent Observatory's MCP Calls view only shows MCP servers that have actual call records, rather than listing all servers configured on the machine.
Agents with non-built-in runtimes can be instrumented and connected to Agent Observatory by sending OTLP/HTTP telemetry data, or by using the project-provided Python and JavaScript clients.
Agent Observatory's views are organized around sessions, activity, recorded MCP calls, trace chains, and host processes.
If the source does not record the reason for waiting or the MCP calls, the Agent Observatory dashboard cannot reconstruct this information.
All Facts (8)
Agent Observatory is a local terminal and browser dashboard tool designed specifically for inspecting session records of Codex and Claude Code, while also receiving OpenTelemetry data from custom instrumented Agents.
50%UnverifiedUsing Agent Observatory requires Python 3.11+ and the uv package manager, and the startup command is uvx --from agent-observatory-cli agent-observatory.
50%UnverifiedAgent Observatory's receiver listens on 127.0.0.1:14318, captured data is stored in a local SQLite file, and session data does not leave the user's machine.
50%UnverifiedOnly when the source explicitly reports a waiting state does Agent Observatory mark a session as waiting, and the dashboard does not infer unrecorded reasons.
50%UnverifiedAgent Observatory's MCP Calls view only shows MCP servers that have actual call records, rather than listing all servers configured on the machine.
50%UnverifiedAgents with non-built-in runtimes can be instrumented and connected to Agent Observatory by sending OTLP/HTTP telemetry data, or by using the project-provided Python and JavaScript clients.
50%UnverifiedAgent Observatory's views are organized around sessions, activity, recorded MCP calls, trace chains, and host processes.
50%UnverifiedIf the source does not record the reason for waiting or the MCP calls, the Agent Observatory dashboard cannot reconstruct this information.
50%