Agents Observe

A real-time, fully local observability dashboard for Claude Code and Codex agent sessions — filtering, search, session replay, and token/cost breakdowns via an auto-starting MCP server.

607stars
58forks
MIT License
TypeScript

Agents Observe installs as a Claude Code plugin rather than a separate service you have to remember to run: adding it as a marketplace plugin and installing it auto-starts an MCP server that captures session events as you work, with no manual setup step beyond the initial install. The dashboard then gives real-time visibility into multi-agent sessions, with filtering, search, full session replay, and (as of v0.9.7) token usage and cost breakdowns.

Because it runs as a local MCP server rather than a hosted service, session data stays on your machine by default. A built-in /observe skill lets you check status, restart the server, or view logs directly from within Claude Code, rather than needing to shell out to manage the dashboard process separately.

MIT licensed, the project includes Docker Compose and Dockerfile support for containerized deployment, alongside the primary local plugin-based workflow.

What You Get

  • A Claude Code plugin that auto-starts an MCP server to capture session events with no manual setup
  • A real-time dashboard with filtering, search, and full session replay across multi-agent sessions
  • Token usage and cost breakdowns per session (added in v0.9.7)
  • A built-in /observe skill for checking status, restarting the server, or viewing logs from within Claude Code

Common Use Cases

  • Reviewing what a Claude Code or Codex agent actually did during a session, including full replay of its actions
  • Tracking token usage and cost across sessions to understand spend without a third-party analytics service
  • Debugging multi-agent workflows by filtering and searching session events in real time
  • Keeping session data entirely local instead of uploading transcripts to a hosted analytics platform

Under The Hood

Architecture Agents Observe ships as a Claude Code plugin (installed via claude plugin install) bundling an MCP server that Claude Code auto-starts, a web app/ for the dashboard UI, and hooks/skills directories integrating directly into Claude Code’s plugin and skill system (the /observe skill is itself part of the shipped plugin). This tight integration with Claude Code’s own extension points — rather than a standalone daemon you configure separately — is what allows zero-manual-setup event capture.

Tech Stack TypeScript/Node.js (via start.mjs as the entry point), Vitest for testing, and Docker/Docker Compose support for containerized deployment alongside the primary local-plugin workflow. A justfile suggests task automation via the just command runner for development workflows.

Code Quality The repo includes a dedicated test/ directory with Vitest configuration, a CHANGELOG.md tracking versioned releases (up to 0.9.7 with specific feature additions noted), and very active, consistently maintained commit history — signals of a project treating releases and testing seriously despite its young age.

What Makes It Unique Unlike hosted coding-agent analytics tools that require an account and upload transcripts to a cloud service, Agents Observe’s plugin-based, auto-starting MCP server design keeps observability entirely local by default, integrating into Claude Code’s own extension mechanisms (plugins, skills, hooks) rather than bolting on an external dashboard.

Self-Hosting

Licensing Model MIT licensed — fully open source with no license key.

Self-Hosting Restrictions None; it runs locally as a Claude Code plugin by default, with Docker support for containerized deployment if desired.

License Key Required No.

Join founders buildingwith open source

Opinionated takes, migration guides, cost-saving tips, and insights from the open source ecosystem.

Subscribe on Substack

No spam. Unsubscribe anytime.

Join 750+ subscribers
No spam. Unsubscribe anytime.

Search