Kit by Speakeasy

Product Introduction

1. Definition: Kit by Speakeasy is a high-performance, open-source coding agent runtime and orchestration platform. It is a static binary that functions as a unified environment for AI-powered software development, combining a terminal client, an Agent Client Protocol (ACP) server, an Agent-to-Agent (A2A) endpoint, and a subagent orchestrator.
2. Core Value Proposition: Kit exists to dramatically accelerate and streamline AI-assisted coding by minimizing model round trips through a unique single-tool architecture. Its primary value is enabling faster, cheaper, and more deterministic AI coding workflows compared to alternatives like Claude Code or Codex CLI, while providing open-protocol compatibility for seamless integration into existing developer toolchains.

Main Features

1. The compose Tool & Runlet Language: Kit's core innovation is exposing only one tool to the AI model: compose. The argument for this tool is a program written in Runlet, a domain-specific language. This single call can chain complex operations—reading files, executing shell commands, applying code edits, running tests, handling errors with retry logic, and delegating to subagents—within one model interaction. This reduces context repetition, token usage, and latency.
2. Subagent Orchestration: Kit can act as a central orchestrator, spawning and managing "subagents" using different AI models and harnesses (like Claude via ACP, Cursor, or another Kit instance). These subagents are reusable, forkable values that can be prompted, inspected, and closed. Kit supports structured JSON output from subagents via JSON Schema validation and allows for explicit context sharing, enabling complex, multi-agent workflows.
3. Open Protocol Support & Interoperability: Kit is built on open standards for maximum compatibility. It natively supports Agent Client Protocol (ACP) v1/v2 over stdio, HTTP/SSE, and WebSocket, allowing it to work in ACP-compatible editors. It also supports A2A v1 for agent-to-agent communication, Model Context Protocol (MCP), and Agent Skills/Plugin packages, preventing vendor lock-in.
4. Persistent & Crash-Safe Sessions: Kit sessions are designed for longevity and reliability. It synchronizes each transcript item to disk as append-only JSONL before acknowledgment and employs automatic context window compaction at 80% capacity. Sessions can be resumed from the TUI, prompt command, or any ACP client. Crash-safe locks and automatic retries for transient OpenAI subscription errors (up to 24 hours) ensure robust operation.

Problems Solved

1. Pain Point: Inefficient AI coding workflows with excessive model round trips. Traditional AI coding assistants expose dozens of individual tools (edit, run, search), each requiring a separate LLM call, leading to high latency, token cost, and context fragmentation.
2. Target Audience: Software engineers and developers leveraging AI coding assistants for production work; teams building AI-powered developer tools or agentic workflows; users of Claude Code, Codex CLI, or Cursor seeking more efficient orchestration.
3. Use Cases: Codebase Refactoring: Orchestrating multi-step changes across a codebase in a single, auditable program. Specialist Agent Workflows: Delegating specific tasks (e.g., UI design, security review) to specialized subagents (Claude Opus, Codex) while maintaining central control. Headless Automation: Running Kit as a daemon (kit serve) for CI/CD pipelines, automated code reviews, or batch processing. Migration Assistant: Using Kit's own AI to help users migrate configurations and skills from other platforms like Claude Code.

Unique Advantages

1. Differentiation: Unlike Claude Code or Codex CLI which are tightly integrated, single-model environments, Kit is a runtime-agnostic orchestrator. It separates the agent client from the runtime via ACP, allowing it to leverage the best model or harness for each subtask. Benchmarks indicate it uses approximately half the input tokens and active time per hand-written line compared to Codex CLI or Claude Code.
2. Key Innovation: The single-tool (compose) architecture combined with the Runlet language. This shifts complexity from the LLM's prompt (managing many tools) to a deterministic, programmable runtime. The model writes a short script, and Kit handles the concurrent execution, error handling, and composition, leading to significantly more work completed per model interaction.

Frequently Asked Questions (FAQ)

1. How does Kit by Speakeasy compare to Claude Code? Kit is a runtime and orchestrator, while Claude Code is an integrated assistant. Kit can use Claude as a subagent via ACP but focuses on reducing round trips via its compose tool. It is generally faster, cheaper per task due to fewer tokens, and more flexible due to multi-harness support.
2. Is Kit by Speakeasy secure for running arbitrary code? Kit has no built-in permissions framework or sandbox; it is a runtime, not a security boundary. It executes commands with the same permissions as the user running it. The recommended security model is to run Kit inside a trusted security boundary like a Docker container, VM, or sandboxed environment, using the --root flag to limit its working directory.
3. What AI models does Kit support? Kit supports OpenAI ChatGPT subscriptions via native OAuth, OpenRouter, and the Speakeasy AI Control Plane. Through its ACP client capability, it can also orchestrate any compatible harness, effectively allowing the use of models from Anthropic (Claude), Google, and others via their respective ACP adapters.
4. Can I use Kit with my existing editor like VS Code or Cursor? Yes, if your editor supports the Agent Client Protocol (ACP). Kit can run as a headless ACP server (kit acp or kit serve), and the editor can connect to it as the runtime. This decouples the client interface from the agent execution environment.
5. How do I handle dependencies and tooling inside Kit? The base Kit binary and Docker images are minimal and do not include language toolchains (e.g., npm, python, go) or Git. You must either run Kit in an environment where these tools are already available or extend the official Docker image to install necessary dependencies.