Get Started

lnk-openclaw lets a Link agent run on OpenClaw, the personal AI assistant. It answers Link's harness adapter commands: installing OpenClaw, configuring it, and saying how to run it. It comes with Link: lnk agent use openclaw adds it and installs OpenClaw the first time, in its own home, ~/Link/Agents/<agent>/Harnesses/openclaw. Every other command is on Link's Agents page. How it compares with Link's other harnesses: Harnesses.

Quickstart

lnk up agents
lnk agent use openclaw        # installs it the first time, then starts it
lnk agent new work --harness openclaw   # or another agent, on OpenClaw

Use a hosted model

Anthropic, OpenAI or OpenRouter, with a key. With no --model, it uses Link's default for the provider (Agents).

ANTHROPIC_API_KEY=sk-ant-... lnk agent start

Use a local model

OpenClaw runs local models only through Ollama, registered with its address and Ollama's own API, and a context window of 32,768 tokens on both sides (OpenClaw's budget and Ollama's num_ctx): Ollama's default, a few thousand, is smaller than OpenClaw's instructions. It needs the models plugin. For a model in LM Studio, llama.cpp, vLLM or Jan, use Hermes.

lnk up models
ollama serve
lnk agent start --model qwen3

Clean up an old OpenClaw install

An OpenClaw installed before Link, or run outside its sandbox, leaves files and its own service outside its home (listed in Security). show finds them; remove deletes them with its home, then start installs it fresh.

lnk agent openclaw show
lnk agent openclaw remove
lnk agent start

How it runs

lnk agent start installs OpenClaw's own release, the one Link pins, with its own Node, in its home, and runs its gateway there, in the sandbox. What it installs and which ports it uses: Security.

You're its command owner on every channel you connect, so its commands, such as /restart and /update on Telegram, work from your phone.
WhatsApp and Signal link from a QR code OpenClaw shows on your terminal. With WhatsApp on your own number, you talk to it in your chat with yourself.
Its memory (its personality, SOUL.md, AGENTS.md, memory/) stays in its workspace, in its home; your files are in the agent's Home.

Limits

Slack needs lan. OpenClaw 2026.9.6's Slack plugin gives its socket (Socket Mode, undici 7) a proxy dispatcher from OpenClaw's own undici 8, which refuses it before it connects, whenever a proxy is set. Behind Link's proxy the socket never opens: the log shows socket-mode … WebSocket error! SMWebsocketError every 15 seconds, and no message arrives. So with Slack or Discord connected, OpenClaw asks for lan before it starts (lnk-openclaw needs), which connects it without the proxy's filtering (the agent reaches this machine and its network too), and asks about its managed browser, whose ports are then on the machine's loopback (security). lnk agent openclaw deny lan takes it back.
No sandbox with OpenAI on a Mac. There OpenClaw runs Codex's runtime, which runs ps, a setuid program no sandbox on macOS runs: with OpenAI on a Mac, OpenClaw asks to run unsandboxed, and every start warns while it does. On Linux it runs sandboxed.

Troubleshooting

Symptom	Fix
"OpenClaw runs local models through Ollama"	The model is in LM Studio, llama.cpp, vLLM or Jan: pull it in Ollama, or use Hermes.
No local model offered	`lnk up models`, then `ollama serve`.
It answers "Auto-compaction could not recover this turn"	The model's context window was too small: `lnk agent start` again sets 32,768 tokens. A model that can't hold that much needs more memory, or a hosted model.
Slack: `SMWebsocketError` every 15 seconds, no messages arrive	OpenClaw's Slack socket can't use Link's proxy (Limits): `lnk agent openclaw allow lan`, or answer yes when it asks.
Something else on your Mac uses port 18789 or the ports after it	Stop it: `main`'s ports are fixed. Another agent's are in its own block (`lnk agent list --json`).
An old OpenClaw install keeps starting outside Link	See Clean up an old OpenClaw install.

Files it keeps

File	What it is
`<harness home>/.openclaw/.lnk-tools.json`	The MCP servers this adapter added to OpenClaw's settings, one per tool, so it removes only its own.
`<harness home>/.lnk-doctored`	The release `openclaw doctor --fix` last ran for.

Why it works this way: Decisions. What it may do: Security.

Work on its code

Build and test

It depends on link-plugin from link-core, checked out beside this folder.

cargo fmt --check
cargo clippy --all-targets --locked -- -D warnings
cargo test --locked
# OpenClaw itself, installed and run by link-core's lnk-harness in its
# sandbox, on the wire (Linux, bubblewrap; a few hundred MB)
# and the conformance kit (link-core's), run against it
LNK_TEST_REAL=1 cargo test --locked --test on_the_wire -- --test-threads 1

Release

Link's release builds lnk-openclaw from this folder and ships it as one of Link's plugins, at the core's version, signed with Link's key.

License

MIT or Apache-2.0, at your option.