Your agent just burned $200. AgentGuard stopped it at $5.

Runtime cost guardrails for AI agents: budget caps, loop detection, kill switch. Drop it in front of any LLM call. Zero dependencies. MIT.

fromagentguard47importAgentGuardguard=AgentGuard(budget=5.00)
withguard:
response=client.chat(...) # enforced at runtime

Works with OpenAI, Anthropic, LangChain, or anything that makes LLM calls.

"Done" is a claim. showwork makes agents prove it: falsifiable claims checked deterministically, refused when reality disagrees. Not observability. Proof. Every repo in my fleet closes through it.

Offline, CPU-only PDF OCR. Rasterize every page, OCR the pixels, treat embedded text as a hint and never as truth. Tesseract + RapidOCR, bounded retries, atomic JSONL output. MIT.

Air-gapped document-intake pipeline for the 8090 MIB Doc Challenge. Scored 137.23/150 on the public train set with zero catastrophic false approvals.

RTX 5090 + RTX 5070 Ti + RTX 3070
llama.cpp · measured, not guessed

I benchmark local models on this hardware and publish the numbers: tokens per second, wall clock, watts, refusal rates. The sizing desk at bmdpat.com/desk runs on that data.

Measured local-LLM data and agent-fleet notes at bmdpat.com:

• My local models refused zero of 50 security tasks
• The faster local model run took 83x longer
• Why local LLM benchmarks need power data