Developer docs
Build with Zero Proof Labs
Three products. The Simulations SDK generates graded training data for any agent. zkFetch proves what an API returned. The intent model reads what the user actually asked for. Each ships with a skill file you hand to your coding agent, and each hosted endpoint takes the same X-Api-Key.
Products
Training data · Python SDK
Simulations
Diverse, graded agent training data from a spec.
Point it at any agent's tools and policy. It writes the users, rolls the agent against a simulated tool backend, grades every trajectory, and exports JSONL training rows. Four modes cover eval sets, paraphrase-robust SFT, and contrastive RL pairs. Nothing is hardcoded per domain: a new agent is a spec file, not an integration.
- Called by
- three lines of Python
- Emits
- graded multi-turn conversations with tool calls
- Modes
- explore, sft, rl, adaptive
- Use when
- an agent needs training or eval data
Proof of data · REST API
zkFetch
What the API really returned, signed.
Generate a cryptographic proof that a specific HTTPS request was made and returned a specific response, without trusting the party who made it. Credentials stay hidden, chosen fields are selectively revealed, and proofs verify off-chain in milliseconds or on-chain on any EVM.
- Called by
- plain HTTP, any language
- Verified
- off-chain in ms, or on-chain on any EVM
- Hides
- API keys, tokens, any field you name
- Use when
- an agent acts on data it fetched
Intent detection · OpenAI API + MCP
Intent model
What the user actually asked for, as JSON.
zeroproof-ecommerce-1b classifies a payment-assistant conversation into one of seven intent types and returns a single JSON verdict with typed details, confidence, and the user messages that support it. Open weights, hosted warm, sub-second responses.
- Called by
- MCP tool or OpenAI-compatible API
- Returns
- one JSON verdict, typed per intent
- Latency
- sub-second, always warm
- Use when
- an agent must know what the user wants
Platform
Authentication
API keys
One key, every hosted endpoint.
Every hosted endpoint authenticates with one key from the platform page, sent as X-Api-Key. Up to five active keys per account, one shared free daily token allowance, counters reset at midnight UTC.
- Header
- X-Api-Key: zp_...
- Free daily
- 500k input, 1M output tokens
- Limit
- five active keys per account
- Use when
- you call anything hosted here
Trace ingest · OpenTelemetry
Trace ingest
Your production agent traces, scored.
Already emitting OpenTelemetry? Three environment variables send your agent's traces here, where each batch becomes a dataset the analyzer scores for prompt grouping, gradient support, pass rate, and reward-hack artifacts. No Zero Proof code runs in your app.
- Called by
- your existing OTLP exporter
- Emits
- one dataset row per trace
- Storage
- 5 GB per account, yours to pull back out
- Use when
- an agent already runs in production
Model releases and their evaluation numbers are on the intent model release history. Questions, or a higher limit: jacob@zeroproofai.com