Provably

Provably is a company focused on runtime verification for AI agent tool calling. Their primary product is SourceryKit, an open-source Python SDK designed to verify that AI agents are calling the correct tools, receiving accurate data from the intended sources, and that outputs have not been altered or hallucinated during execution. SourceryKit works by intercepting agent interactions with external systems — including APIs, databases, and MCP (Model Context Protocol) servers — and generating cryptographic proofs that downstream agents or systems can verify in real time. This allows production agentic workflows to detect errors, hallucinations, and wrong-source data without re-querying source systems, and to perform automated retries or workflow healing. Key capabilities of SourceryKit include: - Runtime verification of agent tool calls during execution, before errors propagate - 100% error detection rate for verifiable guardrails against hallucinations and wrong-source data - Integration with existing observability tools such as Galileo, Arize Phoenix, LangFuse, and LangSmith - Support for source connectors including MCP servers, APIs, and databases - Verifiable SQL queries via their QEDB technology, which avoids heavy SNARK/circuit overhead The product is distributed as a pip-installable Python package and is accompanied by developer documentation and an API. Provably targets engineering teams building production-grade agentic AI systems where data accuracy and reliability are critical, such as customer service automation. The company also publishes research on verifiable databases and cryptographic verification methods.