How do I connect Paperclip to an LLM on Windows?

Paste this repo's link into your AI coding agent (Claude Code, Codex, Cursor). It installs the Hermes Paperclip Adapter in WSL2, writes ~/.hermes/config.yaml for a ZAI glm-4.6 (or MiniMax) model, enables WSL mirrored networking, and adds a hermes.cmd entry point that Paperclip calls.

Can I use MiniMax instead of ZAI GLM-4.6?

Yes. The bridge is provider-agnostic. Set the HERMES_MODEL environment variable on the Paperclip agent (it wins per run) or edit model/provider in ~/.hermes/config.yaml. The only requirement is that the model's provider is configured and authenticated in your WSL ~/.hermes install.

Why does the WSL call fail with 'unexpected EOF while looking for matching quote'?

Paperclip prompts contain double quotes, which break wsl python3 -c "...". The bridge writes the prompt to a file over a stdin pipe instead, so quoting can never break the call. Make sure Paperclip calls a build made from the current launch_hermes.py.

AI-agent setup instructions · Windows + WSL2 · MIT

Paperclip × Hermes Paperclip Adapter integration — run Paperclip agents on a ZAI GLM-4.6 (or MiniMax) LLM, on Windows + WSL2

Paste the repo link into your AI agent — it wires up the whole chain for you.

repo: paperclip-hermes-paperclip-adapter-integration-glm-4.6-zai-minimax-llm-windows-wsl2

Paperclip orchestrates AI agents; each agent needs an LLM backend. This bridges Paperclip (Windows) → Hermes Paperclip Adapter (WSL2) → a ZAI glm-4.6 or MiniMax model — and documents the sharp edges that silently break it.

Set it up → ▶ Video walkthrough

✓ Windows 10/11 ✓ WSL2 · Ubuntu ✓ Node.js 20+ ✓ Python 3.10+ ✓ ZAI · MiniMax

How it works

Point your AI coding agent at the repo — it follows the instructions and glue files.

The repo carries the correct example files and the step-by-step instructions an AI agent follows to build the whole chain. One step — installing Hermes itself — is interactive and flagged for a human; the rest the AI does for you.

the call chain it builds

Paperclip (Node.js, Windows, port 3100)
    │  resolveSpawnTarget → calls the entry point directly
    ▼
hermes.cmd            # main entry point Paperclip invokes
    │  python "%~dp0launch_hermes.py" %*
    ▼
launch_hermes.py      # the bridge — all the logic
    │  pins the model · writes the prompt to a file (quote-safe) · runs hermes
    ▼
WSL2 Ubuntu — hermes chat   # ~/.hermes/config.yaml + .env
    ▼
ZAI glm-4.6  # — or a MiniMax model, provider-agnostic

The two sharp edges it solves for you

Quote-safe prompt passing. Paperclip prompts contain "; passing them through wsl python3 -c "…" makes bash choke. The bridge writes the prompt to a file over a stdin pipe, so quoting can never break the call.
Mirrored networking. WSL2's own network stack can't reach 127.0.0.1:3100. A tiny .wslconfig switches WSL to mirrored mode so the agent can call back into Paperclip.

Good to know

Runs against your own LLM key — you bring a ZAI (or MiniMax) API key.
Windows-only host (Paperclip); the agent side lives in WSL2 Ubuntu.
Everything runs locally on your machine — no third-party proxy in the middle.
Prefer manual? A full by-hand guide ships in external_files/.

What you get

Glue files that make a fiddly integration repeatable.

Each file exists to remove one sharp edge from the Paperclip → Hermes → LLM chain.

🔗

The bridge

launch_hermes.py reads the Paperclip task, writes the prompt to a WSL file (quote-safe), and runs hermes.

🪟

Windows entry point

hermes.cmd is what Paperclip calls; it delegates to the bridge sitting beside it.

⚙

One-command installer

install_hermes_wsl.ps1 installs hermes-agent in WSL, writes the config + key, adds a PATH shim, and verifies.

🔁

Provider-agnostic

Default is ZAI glm-4.6; switch to MiniMax or any backend via the HERMES_MODEL env var, per run.

📦

Build a single .exe

hermes.spec builds dist\hermes.exe with PyInstaller, so Paperclip can call a compiled binary.

🩺

Troubleshooting built in

The README documents the WSL quoting bug, the networking fix and the hermesCommand Paperclip quirk — with the exact fixes.

Set it up

Two ways: let your AI do it, or run the installer.

The repo has everything either path needs.

Let your AI agent do it recommended

Paste the repo's link into Claude Code / Codex / Cursor and say "set up this Paperclip × Hermes integration." It follows the instructions and example files to bring the whole chain up.

Open the repo ↗

Or by hand — three steps

Get a ZAI API key — sign in at z.ai/manage-apikey, create a key, copy it.
Run the installer — in PowerShell from the repo folder: .\install_hermes_wsl.ps1. It sets up hermes-agent, the config (zai/glm-4.6) and your key in WSL.
Point a Paperclip agent at the bridge — set Hermes Command to the full path of hermes.cmd, add ZAI_API_KEY, click Test Environment → Passed. Assign a task and it runs on GLM-4.6.

Requirements: Windows 10/11 with WSL2 (Ubuntu), Node.js 20+, Python 3.10+ on Windows and in WSL, and a z.ai API key.

Drive Paperclip agents on GLM-4.6 — without the yak-shaving.

Open source (MIT). The repo documents and automates the working setup so you don't have to rediscover the sharp edges.

Open on GitHub ★ ▶ Watch the walkthrough