mistri

Mistri is an agent harness designed for Ruby applications, enabling developers to integrate tool-using agents with advanced session management and interaction capabilities. The platform allows users to define tools and hand them to agents, which can then execute tasks such as streaming responses and handling events as they occur. Mistri is implemented as a Ruby gem and requires no additional dependencies for installation.

Key features include support for approvals that do not block execution—allowing risky actions to pause and resume upon later approval—and persistent sessions stored in a user’s database, enabling conversations to continue seamlessly across process restarts or deployments. Streaming is handled natively, with events such as text, tool calls, and internal agent thinking delivered as they arrive. Integration with ActionCable or Server-Sent Events is supported through dedicated sinks, providing real-time updates within Ruby applications.

The tool also manages context compaction near model context limits, summarizing older history while retaining full transcripts in storage. Structured output is enforced, ensuring that agent responses match specified schemas, with errors raised for mismatches. Mistri supports the delegation of tasks to sub-agents, each operating with isolated contexts and returning only their answers to the parent agent. It can interface with Model Context Protocol (MCP) servers, allowing tools and OAuth-enabled resources to be accessed by agents, and includes examples such as connecting to external services or enabling browser automation through MCP clients.

Mistri streams to three providers natively—Anthropic, OpenAI, and Gemini—supporting features like prompt caching, parallel tool calls, and constrained JSON output. The platform is aimed at Ruby developers seeking to build robust, interactive agent workflows within their applications. Installation is handled via the RubyGems package manager, requiring only the inclusion of the mistri gem.