Provedex (provedex/provedex) is an open-source Python library that creates cryptographically signed, tamper-evident audit logs for AI agent actions. Every event is Ed25519-signed at capture time and SHA-256-chained to the previous entry — edit or drop a single event and the chain visibly breaks. Logs are append-only NDJSON files with a fixed seven-event schema covering agent lifecycles: session_started, session_ended, utterance_captured, utterance_spoken, model_invoked, tool_called, and tool_returned. The signing core is written in Rust and exposed to Python via PyO3 bindings, achieving ~11 microseconds per in-process signature and ~3.8ms full cycle with fsync (~261 events/second). Verification is offline — anyone with the public key can validate the entire chain with no callback required. Also ships a sidecar HTTP mode for non-Python environments. Pre-built wheels for CPython 3.11+ on Linux x86_64, Linux aarch64, and macOS arm64.
Provedex solves the accountability gap for AI agents running in regulated or high-stakes environments. When an agent makes a decision — invoking a model, calling a tool, capturing user input — the action is cryptographically signed at the moment it happens and chained to every previous action in the session. The result is an audit trail that is mathematically provable: edit or remove a single event and the hash chain breaks visibly.
The library records seven event types that cover the full agent lifecycle: session start/end, utterance capture/spoken, model invocations, and tool calls with their returns. Each event is signed with Ed25519 and includes a SHA-256 digest of the previous entry, creating an append-only chain stored as NDJSON. The signing core is written in Rust and exposed to Python via PyO3 bindings, keeping per-signature overhead at ~11 microseconds. With fsync (ensuring durability), a full write cycle takes ~3.8 milliseconds — enough for ~261 events per second. Signing keys are loaded at startup from controlled filesystem paths with 0600 Unix permissions.
Provedex fits teams building AI agents for regulated industries — healthcare, finance, legal — where proving what the agent did (and didn’t do) matters for compliance, audits, or legal proceedings. The offline verification model means regulators or auditors can validate the entire chain with just the public key, no API access required. The sidecar HTTP mode extends coverage to non-Python agent stacks. Pairs with Pipecat voice agent pipelines and any Python-based agent framework.
Builder.io's open-source framework for building agent-native applications — shared actions, SQL-backed state, identity, tools, skills, jobs, observability, and UI surfaces that all work together.
Persistent memory layer for AI coding agents — benchmark-backed (95.2% on LongMemEval-S), 92% fewer tokens per session vs full-context pasting, zero manual memory.add() calls.
Open-source AI pair programming tool that works in your terminal to edit code across your entire repository.