Set Up Hermes Agent with DeepSeek

Choose Desktop for a graphical chat window and settings. Choose CLI if you prefer working in a terminal. Both use Hermes Agent; a DeepSeek connection sends model requests to an API rather than downloading model weights to your computer.

For a direct DeepSeek connection, have a DeepSeek API-platform key and sufficient API balance. A chat-app login is not an API key. Hermes itself is free and open source; model calls and optional tool services can have separate charges.

If your key came with a Coding Plan or another provider, use that provider's delivered endpoint and model instructions. A key, Base URL and model ID belong together. Do not send a plan key to the official DeepSeek endpoint unless its delivery instructions explicitly support that route.

Use the official shell installer below on a supported system. Reopen your terminal afterward so it picks up the launcher. If you already installed a desktop package, use its bundled runtime instead of layering a second installation over it.

Windows users can use the native PowerShell installer. Open a new terminal afterward and check the version. WSL2 is a separate Linux installation route, not a prerequisite for the native CLI. Desktop package requirements are covered in the Desktop guide.

Run hermes model to open the provider and model picker. Select DeepSeek and enter the key at the prompt. This avoids putting the key into a command that your shell might retain in its history.

DeepSeek's integration page describes a Quick Setup route, but current Hermes documentation labels Quick Setup as a Nous Portal flow. To bring your own DeepSeek key, use hermes model, or choose Full Setup when running hermes setup. Follow the provider selection rather than assuming every Quick Setup screen is the same.

Before replacing a configuration, privately back up config.yaml and .env in the active Hermes profile. The default home is ~/.hermes; a launcher, --profile option or HERMES_HOME can select another home. Confirm the profile used by the actual launcher and edit that profile, not an unrelated default file. Merge only the intended provider/model settings; do not overwrite existing tools, memory or auxiliary providers with a downloaded example.

A saved default applies to new chats; an existing chat can retain its previous model. After saving, start a fresh session and check the provider and model again. Also inspect launcher overrides and auxiliary-model settings if requests still reach another provider. To recover, restore the backed-up files in the same profile, restart Hermes and test a new chat.

Open a new session with the selected model and send the short prompt below. A normal reply confirms a basic model request; it does not yet prove that file, terminal or web tools work.

For a separate tool check, open a small test folder and ask: List the names of files in the current folder using a file tool. Do not edit or delete anything. Check the visible tool activity and compare the result with the folder. A reply claiming to have checked files without a tool event is not a successful tool test.

For search, ask Hermes to find a current page in the official DeepSeek documentation and return its URL. Check a real web_search event and the returned source. Search uses Hermes tool services, which can require separate credentials or fall back to another search provider; a model key does not itself include search quota. Record chat, file tools and search separately, along with the version, endpoint, model and date.

Our October 14, 2026 test used DeepSeek Flash through the Multi-Model gateway, not the official DeepSeek endpoint or the separate DeepSeek Coding Plan gateway. Hermes chat and read_file passed; web_search passed after its search service fell back from Firecrawl to Keenable. Those results do not verify this direct-API recipe, Desktop, or tools on every model. This guide does not claim a completed official-direct live test.

Keep the working provider and model while adding one capability at a time. After chat and tool use work, build a reusable briefing workflow with Skills and Memory. Use the comparison guides to evaluate OpenClaw with the same DeepSeek model or compare your setup with Meta Muse's managed service.