Jonathon's AI Wiki

Set Up a Grok Subscription in Hermes Agent (xAI OAuth, No API Key)

10 min read

Source: raw/How_to_Set_Up_a_Grok_Sub_in_Hermes_Agent_And_Some_Cool_Ways_to_Use_it.md — YouTube tutorial (sPK8oIGOOLY, AI Garage / onchainaigarage.com). Recorded “May 16th” per the narration. Model names, prices, and timings below are creator-stated from the video, not pulled from xAI or Nous Research’s official docs — verify load-bearing figures before quoting in production.

xAI announced that a SuperGrok subscription can be used as an official model provider inside Hermes Agent via OAuth — no xAI API key required. Once authorized, Hermes can run Grok for text turns, Grok TTS for spoken responses, and Grok Imagine for image and video generation, with the same provider-selection UX as any other Hermes model. This is the Grok-subscription analog of the Codex App-Server Runtime pattern (run OpenAI turns against a ChatGPT subscription, no API key); both let an operator point Hermes at a flat-rate consumer subscription instead of metered per-token API billing. The standout differentiator is Grok’s native access to X data — it can summarize X posts that agents like Claude Code cannot fetch directly because x.com returns HTTP errors to crawlers.

Key Takeaways

  • SuperGrok subscription, not X Premium, not the xAI direct API. The integration is tied to a Grok subscription (“Super Grok”) created at grok.com — distinct from (a) the Grok access bundled with an X Premium account, and (b) the existing xAI direct-API provider in Hermes. As of the recording (May 16th) X-Premium Grok access did not work for this; only the dedicated SuperGrok account did. The creator notes this could change.
  • Pricing (creator-stated): SuperGrok is ~**10/month). A limited-time offer gives 3 days free, but a credit card is required up-front and you are charged if you do not cancel before the trial ends.
  • OAuth flow, no API key. Selecting the xAI Grok OAuth provider opens a long authorization URL in the browser, you authorize “Grok build,” and the connection returns successful — Hermes is then authorized against the subscription. No API key is pasted at any point. The provider only appears after running hermes update (it was not in the older provider list).
  • Grok 4.3 with a 1M-token context window. The creator selected Grok 4.3 and observed a 1,000,000-token context window — noting that even other frontier models they use (referred to as “GPT 5.5”) do not expose the full 1M window in Hermes. Responses were reported as noticeably faster than what the creator was used to.
  • Native X-post access is the killer feature. Grok summarized an X post (a Lighthouse-attention research paper) directly. By contrast, the same “summarize this tweet” request in plain Claude Code (Opus 4.7) failed because x.com returns an HTTP error to direct agent fetches and the model’s web-search fallback could not retrieve it. The only other programmatic route to X data is the X API, which gets expensive at volume — so the subscription delivers a lot of value from this one capability.
  • Grok TTS, Grok Imagine images, and Grok Imagine video are all reachable once configured. Grok TTS sounded markedly more natural than the default Microsoft Edge TTS in an A/B comparison. Grok Imagine video is the differentiator over Codex/GPT Images — the Codex image path can generate images but not video; Grok Imagine produced a video that even added audio (characters speaking) and rendered on-screen text correctly.
  • Capable on a coding-spec planning task. Asked to read an advanced spec file (an “MTP drafter” experiment) and design an execution plan, Grok-powered Hermes followed the spec, proposed a coherent plan, and identified weaknesses in the spec itself — a task the creator says even Codex could not complete on their hardware. Actual code-building with Grok was not tested (flagged for a possible future video).

Setup

The flow shown in the video (creator running Hermes under WSL):

  1. Create a SuperGrok account. Go to grok.com, sign up (email), and subscribe to Super Grok (~$30/mo; 3-day free trial available, credit card required). This must be the Grok subscription — the Grok access from an X Premium account did not work as of the recording.
  2. Update Hermes so the new provider appears. Run hermes update. The xAI Grok OAuth provider only shows up after updating; before that, xAI in the provider list is just the direct-API entry.
  3. Open the provider setup. Run hermes setup and go to the provider list. Alongside the existing xAI (direct API) entry you will now see xAI Grok OAuth — select that one (this is the SuperGrok subscription path).
  4. Authorize via OAuth in the browser. Hermes prints a long authorization URL and opens a browser. Authorize the “Grok build” connection; the browser confirms “Connection is successful,” then return to Hermes — you are now authorized. No API key is entered.
  5. Pick the model. Select Grok 4.3 (the most recent at recording). The provider is now set. Start a chat (hermes chat) and confirm a normal text turn returns a response.
  6. (For TTS) set the spoken-response engine. In hermes setup you can switch the text-to-speech engine to the xAI / Grok TTS voice (the default is the Microsoft Edge “free” voice; a Grok voice such as the “E” voice is selectable).
  7. (For images) switch the image-generation provider. Run hermes tools → reconfigure an existing tools provider / API key → image generation, switch from OpenAI to xAI Grok Imagine (choose high-quality or the faster tier). In the video Hermes did not prompt for an API key when switching — the OAuth authorization already covered it.
  8. (For video) switch the video-generation provider AND enable the tool. Run hermes tools → video generation → xAI Grok Imagine. Critically, go into configure all platforms and make sure video generation is checked on — it may be off by default (it was for the creator, because no video model had been configured before). “If you don’t check it, Grok is not going to be able to use the tool at all or even know that it exists.”

Use Cases

The “cool ways to use it” demonstrated in the video:

  • Summarize an X / Twitter post that other agents can’t fetch. Pasting an x.com post link and asking Hermes to summarize it works with Grok because of its native X integration. The same request fails in plain Claude Code (x.com returns an HTTP error to direct fetches; web-search fallback also failed). This is presented as the single highest-value feature of the subscription, since the only other route to X data is the (expensive-at-scale) X API. Useful for trend, research, and news monitoring off the X firehose.
  • Natural-sounding text-to-speech (Grok TTS). Generating spoken responses with Grok TTS produced clearly more natural audio than the default Edge TTS in a same-line A/B test — no obvious mispronunciations on a sample paragraph. Good fit when Hermes output needs to be voiced rather than read.
  • Image generation with Grok Imagine. A detailed steampunk movie-poster prompt rendered in ~27 seconds and followed the prompt closely (e.g., exactly four figures on a mechanical elephant). Comparable in quality to the GPT-Images output the creator had generated previously from the same prompt, with a somewhat different style.
  • Video generation with Grok Imagine (the differentiator). A “cartoon robots and shrimp in a garage, text falls and bounces” prompt produced a short video that the creator judged better than a prior MiniMax-generated comparison — Grok Imagine added audio (characters speaking) and rendered the on-screen text correctly. Codex/GPT-Images cannot do video, making this Grok-exclusive within the subscription-provider options.
  • Read a spec file and design an execution plan (coding-adjacent). Given an advanced spec, Grok-powered Hermes produced a full plan that followed the spec, offered its own opinions on what it would approach differently (earlier memory guardrails, data-pipeline ordering, tightening feedback loops), and flagged weaknesses in the spec — a planning task the creator says Codex couldn’t finish on their hardware. Full autonomous code-building was not yet tested.

Try It

  1. Start the trial. Create a SuperGrok account at grok.com and start the 3-day free trial (have a card ready; set a cancel reminder).
  2. Update, then add the provider. hermes update → hermes setup → select xAI Grok OAuth → authorize in the browser → pick Grok 4.3.
  3. Smoke-test text. hermes chat → send a trivial message and confirm a response, and note the context-window size reported.
  4. Test the X edge. Paste an x.com post URL and ask Hermes to summarize it — this is the capability you cannot get from Claude Code or Codex.
  5. A/B the TTS. Switch the spoken-response engine to Grok TTS in hermes setup and compare a sample line against the default Edge voice.
  6. Wire up media. hermes tools → image generation → xAI Grok Imagine; then hermes tools → video generation → xAI Grok Imagine, and confirm video generation is checked on under configure all platforms (it may be off by default).
  7. Decide whether $30/mo pays off for your workflow before the trial lapses — the X-data access and Grok-Imagine video are the parts no other subscription-provider option in Hermes replicates.

Open Questions

  • Model name “Grok 4.3” and “GPT 5.5” are creator-stated. The video labels the selected model Grok 4.3 (and references a “GPT 5.5” the creator normally uses). Confirm the exact provider/model identifiers against xAI’s and Nous Research’s official docs before relying on them.
  • X-Premium vs SuperGrok eligibility may change. As of May 16th only the dedicated SuperGrok subscription worked, not X-Premium Grok access. Worth re-checking — the creator explicitly flagged this as subject to change.
  • Rate limits and X-data quotas under heavy Hermes use are unspecified. The same open question that applies to the Codex/ChatGPT-subscription runtime applies here: what happens when a long-running Hermes session burns through the SuperGrok plan’s limits mid-task? Not covered in the video.
  • Does Grok Imagine media generation consume the same subscription quota as text turns, or a separate pool? Image (~27s) and video generations ran without an API-key prompt, but the video does not disaggregate how they meter against the SuperGrok plan.
  • The official xAI blog post / announcement was referenced but not captured. A follow-up Research pull of the xAI announcement and any Nous Research docs page for the xAI Grok OAuth provider would let confidence move from medium to high.
  • Autonomous coding with Grok in Hermes is untested. Only spec-reading and plan-design were demonstrated; the creator floated a dedicated build video. Open until there’s a real multi-file coding trace.

Graph View

Backlinks