A curated, verified, cross-vendor directory of AI agent harnesses — the runtime scaffolding that turns a model into an agent.
◆ 75 harnesses profiled · 9-axis capability schema · last verified 2026-07-08
Mario Zechner's radically minimal MIT/TypeScript harness (~68k stars): four tools (read/write/edit/bash), a ~300-word system prompt, deliberately no built-in MCP/subagents/permission-popups, with everything else added via TypeScript extensions, skills, and npm/git-shareable packages across interactive, print/JSON, RPC, and SDK modes.
Kilo's open-source terminal agent (Kilo CLI 1.0, built on the MIT-licensed OpenCode foundation) distinguished by support for 500+ models with zero-markup BYOK pricing, an MCP marketplace, and cross-surface continuity where a task can move between Slack, VS Code/JetBrains, cloud, and terminal.
Mistral's minimal Apache-2.0 CLI coding agent (launched December 2025 alongside the open-weight Devstral 2 models, which score a vendor-reported 72.2% on SWE-bench Verified), distinguished by a deliberately lean self-hostable design — file read/write/patch, stateful shell, grep, and todo tools — with fully open, locally runnable model weights.
Open-source (Rust, BSD-3) Codex-native personal agent platform whose defining trait is a hardware-isolated microVM per agent — Firecracker on Linux, Hyper-V on Windows, not just a container — wrapped in zero-trust controls: 'Pipelock' keeps secrets and credentials out of an agent's reach and an egress-allowlist proxy governs outbound traffic. Agents are authored, validated and launched entirely from Codex in-repo ('everything is a skill'), build their own capability-gated WASM tools at runtime, persist memory in a shared knowledge graph, and can run always-on via cron-style schedules and heartbeats.