ctxbudgeter

Control what AI agents know before they act

Open Source

Artificial Intelligence

GitHub

Hey Product Hunt 👋 I built ctxbudgeter around a problem I kept seeing as AI agents became more capable: we spend enormous effort choosing models, tools, memory systems and orchestration frameworks, but the actual context entering the model is still often treated like one big opaque blob. That becomes a production problem surprisingly quickly. Context grows, token costs rise, irrelevant information gets included, secrets can leak into prompts, cache layouts change, MCP tool definitions consume thousands of tokens, and when something goes wrong it can be difficult to answer a basic question: What exactly was this agent allowed to know when it made that decision? That’s the problem ctxbudgeter is trying to solve. I think of it as a ContextOps layer that sits before the model call. It can deterministically compile context against a budget, explain what was included or excluded, scan for PII and secrets, preserve provenance, generate a Context Bill of Materials, diff context between builds, budget MCP tools, optimize prompt-cache layout, and visualize the final context through Context MRI. It isn't another agent framework. The goal is to work alongside LangGraph, CrewAI, OpenAI Agents SDK, PydanticAI, Microsoft Agent Framework, or a custom agent loop. The core is local-first and makes zero LLM calls. The larger idea is simple: agent observability tells you what an agent did. ContextOps should tell you what the agent was allowed to know before it did it. ctxbudgeter is open source, and I’d especially love feedback from people building production agents: where is context management breaking down for you today—cost, retrieval noise, security, MCP tool bloat, caching, memory, or auditability? That feedback will heavily influence where I take the project next. I’d keep the launch developer-focused rather than hype-focused. The “what was the agent allowed to know?” framing is the memorable differentiator; features like Context BOM, Context MRI, deterministic compilation, security scanning, MCP budgeting and pytest-style assertions then provide evidence that this is a real engineering layer rather than another prompt utility.