Vvanakor
Back to research
activePython2026

ContextMesh

See where your agent's tokens go.

An open-source observability layer for AI coding agents. Every step an agent takes — every file read, every test run — is logged to a local, typed, evidence-backed ledger. It publishes one agent-agnostic number that nobody else does: the fraction of tokens spent on tasks that actually finished.

View on GitHub

Key metric

useful_context_ratio

the one number that matters

Highlights

  • 01Agent-agnostic: traces any agent CLI — Claude Code, Codex, Cursor, Aider, or your own.
  • 02Puts provider cache savings and on-top compression side by side, so you can see which actually paid.
  • 03Typed, evidence-backed ledger any agent can write to.
  • 04119 passing tests, MIT licensed, currently alpha at v0.4.0.

Tags

observabilityai-agentscontext-engineeringtoken-optimizationllm