Announcing agent-gov: Open-Source AI Agent Cost Governance
Stop waking up to surprise $500 bills from your AI agents.
The 3 AM Wake-Up Call
It was 3:47 AM on a Tuesday. My phone buzzed — a Cloudflare bill alert. Then another. Then Stripe. By the time I stumbled to my laptop, three different providers had collectively racked up $487 in just six hours.
What happened? A single AI coding agent had gotten stuck in a loop. It was re-analyzing the same bug, calling the same expensive LLM endpoint over and over, spawning sub-agents that spawned their own sub-agents. Nobody put a governor on it. Nobody thought they needed to.
If you've built anything with AI agents — auto-PR reviewers, customer-support bots, code-gen pipelines, web-research assistants — you've either had this nightmare or you're one sleep cycle away from it. The fundamental problem is simple: agents spend money the same way junior devs write code — enthusiastically, autonomously, and without asking permission.
Most teams solve this with spreadsheets and hope. Some bolt on a cloud budget alert after the first blowup. A few give up on agents entirely.
We wanted a real answer.
Enter agent-gov
Today I'm releasing agent-gov — an open-source, MIT-licensed cost governance platform purpose-built for AI agents. It's a lightweight reverse proxy that sits between your agents and their LLM providers, tracking every cent, enforcing daily budgets, and auto-pausing agents that overspend.
pip install agent-gov-saas
agent-gov start
That's it. Thirty seconds from zero to governance.
How It Works
Agent-gov is transparent to your agents. You point them at a local proxy endpoint instead of directly at the API, and agent-gov handles the rest:
Intercept — every model call routes through agent-gov's FastAPI proxy
Look up real costs — the built-in tool registry knows exact per-token pricing for hundreds of models
Enforce budgets — if an agent exceeds its daily allocation, agent-gov blocks the call
Persist everything — all usage is written to SQLite via aiosqlite
No lock-in. No cloud dependency. No per-seat licensing.
Features in v0.5
Reverse proxy middleware — drop-in for any OpenAI-compatible client
SQLite persistence — full audit trail of every call, token, and dollar
Tool registry with real cost tables — priced by model, provider, and input/output token rates
Daily budgets with auto-reset — configure per-agent, per-workspace, or globally
Auto-pause on over-budget — agents get a structured policy block
Multi-tenant workspaces — isolate costs by team or project
Docker support —docker compose up
45 tests, 0.3 seconds — tight, fast, well-tested codebase
Quickstart
pip install agent-gov-saas
agent-gov start
Set a $5 daily budget:
agent-gov config set budget 5.00 --agent code-review-bot
Why We Built This
The AI agent ecosystem is exploding. But the operational maturity around it is where web apps were in 2009. We're all running agents without guardrails because nobody has built the guardrails yet.
Agent-gov is the circuit breaker for your agent infrastructure — the thing that prevents a single runaway loop from costing you a week of GPU credits.
The project is MIT-licensed because cost governance isn't a moat — it's table stakes.
Get Involved
GitHub: github.com/sschelliah2026-source/agent-gov
Install:pip install agent-gov-saas
Don't learn you need cost governance at 3 AM.
Built with FastAPI, SQLite, and a healthy fear of surprise bills.