AgentCost

Unified AI coding agent cost tracking + budget guardrails.

Claude Code, Cursor, Codex, MCP — one command to see where your AI budget went, and stop it from blowing up.

Why

AI coding agents are amazing until the bill arrives. Costs hide across different tools, models, and projects. AgentCost reads your local agent logs (nothing leaves your machine) and produces one unified cost report — by model, by day, by agent type — with budget guardrails that warn you at 80% and yell at 100%.

Features

• ✅ Multi-agent: Claude Code, Cursor, Codex logs — unified attribution
• ✅ Local-first: logs parsed on your machine, zero upload (privacy by design)
• ✅ Accurate pricing: 23 providers built-in (reused from the TokenTracker engine), cache-read discount at 0.1x
• ✅ Budget guardrails: monthly limit → 80% WARN / 100% OVERRUN alerts
• ✅ Two forms: CLI + MCP server (ask your agent "how much have I spent?")
• ⏳ Roadmap: forecasts, anomaly detection, team panel

Install

npm install -g agentcost-cli
# then run:
agentcost scan ~/.claude/projects 50

Usage (CLI)

# Scan a log directory, $50 monthly budget
agentcost scan ~/.claude/projects 50
agentcost scan ~/.codex/sessions 50
agentcost scan ~/.cursor 50

Output: unified cost by model + by day + budget guardrail status.

Usage (MCP Server)

npx agentcost-mcp   # stdio MCP server
# tools: scan_cost / get_budget / set_budget

Add to your MCP client config — then your agent can answer "what's my spend this month?".

How it works

agent logs (JSONL) → parser → pricing engine (23 providers, $/1K tokens)
                   → aggregate (model/day/source) → budget guardrail → report

Privacy

All parsing happens locally. No telemetry, no cloud, no upload. Your logs stay yours.

License

MIT — part of the SellerTools family.