Anthropic, OpenAI, Google, Amazon, Microsoft — every AI coding runtime emits its hook events in a different shape. fastpace normalises them to the same three boring fields at the developer's laptop, signs the result with a key only that laptop holds, and never phones home.
Shell, BP, Chevron, Exxon, and Mobil all sell the same gasoline but each prints a wildly different receipt. Your accountant doesn't memorise five receipt formats — she runs each through a one-page cheat sheet that says "copy the octane number into grade, the dollars into total." Now every receipt looks identical before she touches it. fastpace is the cheat sheet for AI runners.
fuel: 91
dollars: 42.10
station: chicago-loopoctane_rating: 91
total_usd: 42.10
location: "Chicago Loop BP #421"gas: regular-91
charge: $42.10
addr: 401 N Wackerclaude-code metadata.model "claude-opus-4.7" metadata.endpoint "https://api.anthropic.com" metadata.do_not_train true codex metadata.model_name "gpt-5" metadata.host "api.openai.com" metadata.x_anthropic_no_training true gemini-cli metadata.gemini_model "gemini-2.5-pro" metadata.vertex_endpoint "…aiplatform.googleapis.com" metadata.user_data_opted_out true bedrock metadata.modelId "anthropic.claude-opus-4-7-v1:0" metadata.bedrock_endpoint "bedrock.us-east-1.amazonaws.com" metadata.bedrock_training_disabled true azure-openai metadata.openai_model "gpt-5" metadata.azure_endpoint "westus3.openai.azure.com" metadata.user_data_opted_out true { model, endpoint, training_opt_out, runtime } When GPT-6 ships next year on a new endpoint, you add four lines to the adapter map — not forty lines spread across twelve downstream consumers.
Auditors reading fastpace/manifests/ see one consistent shape, regardless of which engineer used which AI tool that day. No polyglot mess.
Portkey · Helicone · Langfuse — normalise on a server in their cloud. Your traffic flows through them.
Normalises at the developer's laptop. Signs the canonical envelope with a key only that laptop holds. Never phones home.
