株式会社オブライト
AI2026-05-17

Hermes Agent Now Connects to X Premium / Grok Subscriptions and Searches X Posts — What Nous Research and xAI Shipped in May 2026

Nous Research's open-source self-improving AI agent Hermes Agent gained direct access to xAI's Grok subscriptions (including X Premium and SuperGrok) on May 15, 2026, plus an `x_search` tool that queries X posts, threads, and profiles via the xAI Responses API. This column reads the official sources to lay out how the two features actually work, what's authenticated against what, which Grok model variants are supported, and how the combination realistically fits into business workflows.


What Just Shipped

On May 15, 2026, xAI published Connect Grok to Hermes Agent, wiring Nous Research's open-source agent (GitHub: `NousResearch/hermes-agent`) into the Grok subscription stack. Two things landed at once:

1. Direct use of Grok subscriptions — X Premium, X Premium+, and SuperGrok tiers can now drive Hermes Agent without users manually pasting API keys 2. `x_search` tool — an xAI Responses API built-in that lets Grok search X posts, threads, and profiles server-side and return cited synthesis, callable from Hermes Agent

Hermes Agent now ships as "a self-improving, persistent-memory agent with X as a first-class real-time data source."

Hermes Agent in Context

Hermes Agent was released by Nous Research in February 2026 as an open-source, self-improving AI agent. Where OpenAI's coding agents and Anthropic's Claude Code Agent View emphasize agents that live inside a codebase, Hermes Agent's differentiators have been:

- Persistent memory — skills, procedures, and preferences accumulate across sessions - Model-agnostic backend — Grok, Claude, GPT, Llama-family models can be swapped; this update positions Grok as the lowest-friction default

It's open source on GitHub and also runs as a hosted SaaS at hermes-agent.nousresearch.com.

Feature 1 — Direct Grok Subscription Access

Auth is OAuth 2.0 PKCE via accounts.x.ai. A user clicks "Connect with X" inside Hermes Agent and they're done — no copy-pasting API keys. Compared with the usual "create API key in console → paste into agent" flow, the onboarding distance just got materially shorter.

Tiers and capabilities (per xAI's announcement):

TierGrok 4.3 text chatAdvanced reasoningTTSX Premium bundled
X Premium-(is the bundle)
X Premium+(is the bundle)
SuperGrok-
xAI API (metered)-

An existing X Premium subscriber can now run Grok from Hermes Agent at no marginal cost. Strategically, xAI is fishing for AI users among its existing X paying base — a different angle of attack than the coding-agent vendors.

Feature 2 — `x_search` Tool

Nous Research's official docs (Hermes Agent — X Search) describe `x_search`, a tool that runs inside xAI's Responses API (`api.x.ai/v1/responses`) and lets Grok search X server-side, returning a synthesized answer with citations to the underlying posts.

Capabilities: - Handle filters — restrict to a specific account, e.g. `@elonmusk` - Date filters — last 24 hours / week / month - Image/video understanding — multimodal, so media-rich posts get parsed - Cited responses — output includes URLs to source X posts - Recommended model: `grok-4.20-reasoning` - Auth: xAI OAuth (SuperGrok / X Premium+) or paid xAI API key

Unlike web-search agents that read public Google results, `x_search` sources the X graph itself, including real-time posts. For tracking specific accounts in real time, this is a different capability class from Perplexity or Google AI Mode.

Practical Business Workflows

Combined with persistent memory, common workflow patterns become straightforward:

- Automated social listening — periodic searches for brand / competitor / category terms, with summarized and sentiment-classified output piped to Slack - Launch-day reception analysis — handle-scoped queries with media parsing, surfacing user reactions to a product launch in real time - Analyst / KOL tracking — long-running follow of industry voices, with notes flowing into Notion or Linear - Crisis response — Grok's reasoning layer classifies whether a spike in negative mentions is a contained pile-on or a real product issue - Competitive intelligence — inferring strategic shifts from public posts by competitor employees or recruiters

At Oflight we'd typically wire these patterns into our AI consulting engagements, often alongside parallel-agent orchestration like Claude Code Agent View or OpenAI Symphony.

Caveats and Limits

Worth flagging:

- The `x_search` GA date isn't explicitly published in the docs. Third-party observers (TestingCatalog) place it at the same May 15–16, 2026 rollout. API-level SLA and rate limits aren't yet public - X's developer terms still apply to redistribution of fetched posts — internal analysis and public content generation have very different compliance profiles - Model lock-in for this feature — Hermes Agent is model-agnostic in general, but `x_search` is Grok-only; swapping the backend kills the capability - Light Japanese-language coverage — as of May 17, 2026 we couldn't find Japanese first-party reporting (PR TIMES / ASCII.jp / Impress Watch). Coverage is mostly English and some Chinese-language outlets - Persistent memory × social search is privacy-flammable — letting an agent silently accumulate personal posts in its memory tends to trip internal governance. Memory scope needs explicit design upfront

How This Reshapes the Agent Market

Market shape as of mid-May 2026:

AgentStrengthDifferentiating data layer
Claude Code (Anthropic)Local supervisor + worktree isolationCodebase + MCP tools
OpenAI Symphony / CodexLinear-ticket-driven flowGitHub + Linear
Cursor Background AgentsCloud-VM isolationCodebase
Devin (Cognition)High autonomyGeneral web + sub-agents
Hermes Agent (Nous)Persistent memory + model-agnosticX itself (via `x_search`) + general web

Hermes Agent is now competing on a different axis from the coding agents: the X graph as a primary, real-time, person-driven data layer is something the others can't easily reach.

FAQ

Q1. Can an existing X Premium subscriber run Grok inside Hermes Agent at no extra cost? A. Yes. xAI's announcement says the OAuth flow at accounts.x.ai handles billing automatically. What you can actually call (which Grok model, which features) still depends on whether you're on Premium, Premium+, or SuperGrok. Q2. Is `x_search` free? A. It's a built-in tool of the xAI Responses API and requires either xAI OAuth (SuperGrok / X Premium+) or a paid xAI API key. Whether plain X Premium reaches `x_search` isn't explicit in the docs — test before relying on it. Q3. Can I call `x_search` from a Claude- or GPT-backed run of Hermes Agent? A. No. `x_search` is a Grok-only Responses API tool. Switch the backend off Grok and the feature is gone. Q4. Can I redistribute fetched X posts in customer-facing content? A. X's Developer Agreement / Content Redistribution Policy applies. Internal analysis is one thing; public republication needs legal review. Q5. Is Hermes Agent open source or SaaS? A. Both. Self-hostable from the GitHub repo; SaaS at hermes-agent.nousresearch.com. The X Premium connection appears to have rolled out on the SaaS first. Q6. Why would a Japanese enterprise care? A. X is still a primary real-time channel in Japan, so social listening, PR, crisis response, and talent-market analysis are natural fits. Memory-scope governance is the prerequisite to making it deployable.

Bottom Line

This update is more than "another agent integration." It signals that the AI agent market has begun competing on which data layer the agent can see natively. Coding agents own the codebase; web agents own the open web; Nous × xAI just claimed X as a real-time, person-driven layer that the others can't directly touch.

For Japanese organizations, the practical hook is that an existing X Premium subscription gets you a usable Grok-backed agent at no marginal cost. The non-obvious work is upstream: defining how fetched posts are stored, scoping persistent memory against privacy obligations, and avoiding accidental lock-in to a Grok-only path. Solve those three first, then this becomes a real production capability rather than a demo.

References

Feel free to contact us

Contact Us