Compare Harnesses
The four harnesses that come with Link, side by side. Every one takes a hosted model from Anthropic, OpenAI or OpenRouter with your key, and speaks Telegram, Slack, Discord and Signal.
lnk agent use --harness openclaw # switch to one, installing it the first timeWhere they differ
Each row is one way they differ: who makes it, how it installs, which local models it takes, and what it remembers.
| Link Harness | OpenClaw | Hermes Agent | DeepSeek Harness | |
|---|---|---|---|---|
| Made by | Local Link | The OpenClaw project | Nous Research | DeepSeek |
| Switch to it | lnk agent use --harness link | lnk agent use --harness openclaw | lnk agent use --harness hermes | lnk agent use --harness deepseek |
| Install | In a second, nothing from a third party | Its release, with its own Node | Its release, with its own Python and Node | dsh and Node, about 800 MB |
| Local models | Any runtime | Ollama only, 32k context | Any runtime, 64k context | Any runtime |
| Connections | All four, carried by Link | All four, carried by Link | All four, carried by Link | All four, carried by Link |
| Asks for more than Link's defaults | Never | No sandbox with OpenAI on a Mac | Never | Never |
| Its memory | Plain files in your files folder | MEMORY.md in its workspace, and what's kept elsewhere | Its MEMORY.md, and what's kept elsewhere | A dsh session per chat |
| Link Scheduler: subagents and schedules | Yes | No; its own scheduled jobs reach no chat | No; its own scheduled jobs reach no chat | No |
| Its own site | local.link | openclaw.ai | nousresearch.com | GitHub |
Pick one
Your agent runs Link Harness unless you pick another.
- Link Harness to start. It's small and readable, installs in a second, asks for nothing beyond Link's defaults, and keeps each conversation as files you can read. It's the one harness that starts subagents and schedules messages to itself.
- OpenClaw for OpenClaw's personal assistant, with its own personality files. It runs local models only through Ollama.
- Hermes Agent for Nous
Research's agent, with a model in Ollama, LM Studio, llama.cpp, vLLM
or Jan. Its
hermes claw migratebrings over OpenClaw's memory and skills. - DeepSeek Harness for DeepSeek's open-source agent, with DeepSeek's models on OpenRouter or any other. Link turns off everything it would send DeepSeek, its web search among them.
Switch Harnesses says what comes along when you switch, and Harnesses what Link does for every one.