Cohort: Local agent runtime for auditable, evidence-driven automation
Cohort by Congchuanling Dot is a local-first agent runtime that connects large language models to local system resources for secure automation. The app records agent actions and provides replayable execution traces, targeting reproducible testing and governance. It bundles session recording, verifiable memory, and protocol bridges for browser and desktop automation. Cohort is aimed at AI developers, researchers, and power users who need auditable, testable agent workflows on their own hardware.
What tasks can you actually use it for?
The tool maps LLM outputs to concrete web and desktop actions through specific bridges and APIs. Its Chrome Bridge operates over the Chrome DevTools Protocol so agents can read the DOM, execute JavaScript, take screenshots, and interact with page elements. Desktop controls use macOS Accessibility (AX) integration for window and UI manipulation. Native MCP server support also enables connecting MCP-compatible clients for complex multi-tool workflows.
How verifiable and reproducible are its agent traces?
Cohort records full execution traces with cryptographic proof, producing SHA-256 proof bundles that let developers verify and replay sessions offline. The Time Machine lets developers fork a recorded session at a chosen point to run counterfactual experiments and observe alternate outcomes. Long-term memory entries require tool-backed evidence, enforcing an evidence-driven memory model that ties stored facts to verifiable tool outputs.
What are the platform and integration requirements?
The runtime expects an MCP-capable environment and uses a Node.js CLI with NPM packages for developer interaction, while the core runtime is implemented in Go. The architecture is local-first, so all processing runs on the host machine rather than a remote control plane. Cross-platform support exists for core functions, though desktop automation features are currently optimized for macOS Accessibility APIs.
Does it fit into developer workflows and governance needs?
Cohort includes a Runtime Governor that enforces controls over context windows, caching, and tool-related risk, giving teams a way to audit and limit agent actions. Governance extends to usage policies and reproducibility practices by recording the full toolchain used in a session. The design targets engineering workflows where traceability, repeatable tests, and mitigated tool risk are primary concerns rather than casual, non-technical experimentation.
A focused choice for engineering teams prioritizing verifiable agent experiments
Given its local-first execution model and the Time Machine's cryptographic traces, Cohort is a practical option for AI developers and researchers who need auditable agent behavior. The project is well-regarded among open-source MCP adopters, making it suitable for community-driven experimentation. Expect an engineering-oriented setup and plan to integrate it into development and testing pipelines rather than end-user applications.
