What is Mnestiq?
Mnestiq records AI agent activity as verifiable evidence, including model calls, tool activity and, when configured, outbound network connections. It is built for teams that need to investigate an agent’s actions, understand what inputs came before them and check whether recorded evidence has been altered. Python developers can instrument Anthropic or OpenAI clients in code, while agents in other languages can use its proxy if they support an OpenAI- or Anthropic-format API base URL.
Mnestiq Features:
Agent Run Recording: Records agent runs and steps with identifiers such as agent, operator, sandbox and parent-run IDs when those details are available.
Model Call Capture: Captures calls made through supported Anthropic and OpenAI Python SDKs, including synchronous and asynchronous clients, Chat Completions and Responses APIs.
Proxy Recording: Records model calls from outside an agent through a proxy that accepts OpenAI- or Anthropic-format API traffic.
Egress Tracking: Can record outbound connections made through Python’s socket layer and asyncio, with connection details linked to agent runs.
Input Provenance: Tags captured context by source, including system, user, tool output, retrieved documents, web content and memory.
Signed Evidence: Hash-chains records and seals them with Ed25519-signed checkpoints, which can be checked with a trusted public key.
Investigation Dashboard: Shows runs, a context-colored step timeline, findings and verification status, and can locate an agent step from a recorded IP address and port.
Mnestiq Benefits:
Clearer Incident Reviews: Teams can trace a suspicious action to a specific run and step instead of relying only on a firewall event or an agent’s final output.
More Useful Evidence: Source tags help investigators see whether untrusted content appeared before a consequential tool call or outbound connection.
Verifiable Records: Signed checkpoints make changes to sealed evidence detectable during later review.
Safer Evidence Sharing: Prompts, tool arguments and outputs can be removed for an auditor while preserving a chain that can still be verified.
Fewer Integration Changes: Teams can capture compatible agents through the proxy without rewriting their application code.
Use Cases:
Prompt Injection Review: Investigate whether untrusted web content preceded an agent action that sent data, changed an account or used a sensitive tool.
Outbound Connection Triage: Use a firewall’s recorded IP address and port to find the associated agent step and review what happened earlier in the run.
Agent Audit Preparation: Share a redacted run record with an auditor when prompts or tool contents should not be disclosed.
Support Agent Monitoring: Review a customer-support agent’s model context, tool activity and approvals after a questionable response or action.
Cross-Language Capture: Record model calls from an agent built in another language when it allows its API base URL to be pointed at the Mnestiq proxy.

