TokenTelemetry — Observability for AI coding agents: see token usage, cost & reasoning (100% local, open source)

[VasiHemanth]

Hey MemPalace community!

Building AI agents that use MemPalace for persistent memory is powerful — but one thing that's been missing is visibility into what those agents are actually doing and spending.

I built TokenTelemetry to fill that gap — a 100% local, open-source observability dashboard for AI coding agents.

What it tracks

• Token usage & cost per session, per model, in real-time
• Reasoning traces — see the full chain of thought your agent used
• Subagent activity — which sub-agents spun up and what they did
• Tool calls & artifacts — every action your agent took

Supported agents (zero config)

Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, Qwen, Vibe, Antigravity, OpenCode — reads directly from local log files.

Why it's relevant here

If you're running agents with MemPalace memory backends, TokenTelemetry gives you the observability layer on top — so you can see not just what your agent remembered, but how much it cost and what reasoning it did to get there.

Install

curl -fsSL https://raw.githubusercontent.com/VasiHemanth/tokentelemetry/main/install.sh | bash

MIT licensed · 100% local · no signup · Node 18+, Python 3.9+

GitHub: https://github.com/VasiHemanth/tokentelemetry
Website: https://tokentelemetry.com

Would love feedback, especially from folks running long agentic sessions with memory!

[ajdelaguila]

Full disclosure: I maintain Data Olympus, an adjacent governed project-memory MCP server.

The observability angle resonates. One thing I would want in any memory benchmark or dashboard is a distinction between retrieved and allowed to govern. A memory can be relevant but still be stale, superseded, or only an unreviewed agent observation.

In Data Olympus v0.4.0 we made that state explicit: proposal provenance, human promotion, validity windows, supersession edges, and a maintenance ledger. That has been useful for auditing why an agent saw a rule and whether it should have been allowed into the prompt at all.