Solo-dev AI agent cost tracker (Langfuse-but-tiny)

dev tool real project ••• trending

Indie devs and small teams running Claude Code, Cursor, Aider, and homegrown agents are eating surprise bills with no per-feature breakdown. Existing LLM observability is built for ML platform teams (LiteLLM proxy, Langfuse self-hosted, Helicone) and feels like overkill for one person tracking one repo. Demand is for a local-first, single-binary cost tracker that hooks into the agents you actually run, attributes spend to repo/branch/task, and warns before you cross your own budget.

builder note

Don't try to be Langfuse-lite. Ship a single binary that scrapes the agents' own log files (Claude Code's ~/.claude/projects, Cursor's session JSON, OpenRouter usage API) and produces a weekly invoice by branch. The Langfuse SDK route loses every time on a one-person team.

landscape (4 existing solutions)

Real LLM observability is built for ops teams managing prod inference. Nothing in the middle gives a solo dev a single, no-config view of 'how much did I spend on this branch this week' across Cursor + Claude Code + a few API scripts. The gap is positioning, not technology.

Langfuse Self-hostable and powerful but assumes you want a Postgres + ClickHouse stack and a web dashboard. Designed around production LLM apps with eval, prompt management, RBAC. A solo dev tracking one Claude Code session shouldn't need a 5-service docker-compose.

LiteLLM Proxy Great proxy with budget enforcement, but you have to route every agent through it. Most coding agents (Cursor, Claude Code) don't speak the OpenAI proxy protocol natively and you lose model-specific features by squeezing them through.

Helicone Cloud-first, requires sending requests through their proxy, B2B pricing model. Friction and privacy concerns for a solo dev who just wants a number at the end of the day.

agenttrace Closest in spirit (local TUI, anomaly reports), but narrow (Augment Code session focus) and doesn't unify across the three or four agents most devs run in parallel.

sources (3)

other https://medium.com/@nirbhaysingh1/our-ai-bill-was-4-800-last... "Our AI bill was $4,800 last month. Nobody knew why." 2026-02-15

other https://dev.to/liv_melendez_4be3c47ea998/what-the-ai-agent-c... "Builders are debugging session burn and invisible orchestration costs" 2026-05-10

other https://www.augmentcode.com/tools/best-ai-agent-observabilit... "agenttrace, a local-first TUI for AI coding agent session observability" 2026-05-19

ai-agentsobservabilitycost-trackinglocal-firstdeveloper-tools