4/ I'm not asking you to trust the claim. There's a live attack playground on the site. Pick an attack or write your own tool call, watch the gate decide. Published suite: 24/24 blocked, 0 false…

5/ So: callgate, a firewall for agent tool calls. Schema + policy before anything executes. Audit trail. Dry-run mode. Zero deps, native OpenAI/Anthropic/Gemini shapes, no migration.

Agent tax: verbose tool-call JSON. My test: TOAP (compact + proxy). No claims. Open source.

Industry default: fat JSON on every agent tool call. My experiment: TOAP. Compact on the wire. Expand before tools run. No victory lap. Just the work. Curious what % of your spend is structure vs…

Not claiming I solved agent token costs. Testing a smaller tool-call format + proxy path (TOAP). Same tools. Less wire. Framework-friendly. If it sticks, great. Until then - shipping.

Quiet problem in agents: you pay for repeated keys as much as for decisions. TOAP = experiment to cut that wire waste. Building anyway. Results TBD. How do you measure tool-call overhead today?

Same tools. Less wire. That's the bet with TOAP - compact agent tool calls + middleware expansion. Experimenting openly. Not selling a miracle. LangChain / CrewAI folks — would this help you?

Agent bills aren't only "smart model" cost. A lot is verbose tool-call JSON. TOAP = my experiment to shrink that layer. Still early. Still iterating. What's the dumbest token waste you've seen in…

What if tool calls didn't need a full JSON novel each time? That's the experiment behind TOAP. Compact format → proxy → your tools. No big claims. Active work. Would you try this on a side agent…

Gemini numbers (few-shot): • Format compliance: 100% • Tool accuracy: 93.8% • Output token savings: ~45% • Net savings (with prompt): ~5–6% I want the same numbers on GPT-4o and Claude 3.5/4. Even…