Short answer: Claude Code, Codex and Cursor are the only three built to run unattended out of the box — headless or cron modes their own docs describe. Hermes and OpenClaw go further and put real cron scheduling inside the agent itself, whichever MCP server it's pointed at. Claude (claude.ai and Desktop), ChatGPT and Grok's consumer app connect cleanly but stay chat-first: you ask, something else holds the queue while you're away. And if you came looking for "Gemini CLI" specifically, the product changed under you in June 2026 — read that section before copying a config from anywhere, this post included.

I build one of the schedulers an agent like these would connect to, seenpaid — 21 platforms, 50 tools on its MCP server — so read this with that in mind. Every plan requirement, config format and scheduling claim below is checked against that vendor's own documentation, fetched today, 24 September 2026, not memory and not another site's recap of it. Where a third-party summary disagreed with the primary doc, the primary doc won, and I say so. Sources are listed at the end.

Claude (claude.ai and Claude Desktop)

Connects with an OAuth custom connector — no key field, you sign in. It works on every plan including Free, but Free is capped to one custom connector total; if seenpaid is that one, nothing else shares the slot (support.claude.com/en/articles/11175166). It has no scheduler of its own: ask it to post something Thursday and it goes out on time because the connected scheduler holds the queue, not because Claude ran anything while you were away. Strength: nothing to leak, since there's no key to paste anywhere. Weakness: the one-connector cap on Free, and it only ever acts mid-chat.

Claude Code

Connects with one command — claude mcp add --transport http seenpaid <url> --header "Authorization: Bearer sp_..." — key-based, no file to edit. There's no free Claude Code plan: it needs at least Pro ($17/mo billed annually, $20 monthly), Max, or API credits (claude.com/pricing). What sets it apart here is claude -p, its documented headless mode: one prompt in, one result out, built explicitly for "running Claude on a schedule" to automate repeating work like overnight failure analysis (code.claude.com/docs/en/headless). Point cron at that and it can check the queue or draft a post with nobody watching. Strength: the only harness on this list designed for cron from day one. Weakness: no free tier to try it on first.

ChatGPT

Two routes: add a scheduler from ChatGPT's plugin directory (chatgpt.com/plugins) and sign in — seenpaid was approved there on 24 September 2026 — or, on a paid plan, Developer Mode with the server's URL, OAuth only. Developer Mode is Plus, Pro, Business, Enterprise and Education, not Free, and OpenAI's own docs say it grants "full Model Context Protocol (MCP) client support for all tools, both read and write" to every one of those plans alike (developers.openai.com/api/docs/guides/developer-mode) — a few third-party recaps claim Pro only gets read-only access, but that isn't what the primary doc says, and the primary doc wins. Scheduled tasks exist too (3 active on Free/Go, 5 on Plus), but they're reminders ChatGPT runs and messages you about, and OpenAI's own guidance is that a write action inside one can still pause for your approval — not a silent standing job. Strength: the plugin directory is the easiest install here if your scheduler is listed. Weakness: Free ChatGPT can't reach any MCP server, listed or not.

Codex

Connects with a [mcp_servers.seenpaid] table in ~/.codex/config.toml, the key read from an environment variable so it never sits in the file. Codex CLI needs at least ChatGPT Go ($8/mo) — Free doesn't include it (learn.chatgpt.com/docs/pricing) — or API credits. codex exec is its non-interactive mode, "the entry point for every CI/CD integration": shell scripts, cron jobs, git hooks (developers.openai.com/codex/noninteractive.md). Its sandbox defaults to read-only, though, so a cron job can't actually call a write tool like schedule_post until you deliberately loosen it. Strength: the same CI-native design as Claude Code. Weakness: that safe default means an unattended run silently can't write until you say so.

Cursor

Connects with a url and headers block in ~/.cursor/mcp.json. MCP itself isn't plan-gated; even Hobby (free) includes limited Cloud Agent runs. Of everything on this list, Cursor's Cloud Agents (formerly Background Agents) go furthest with unattended: they run on their own VM without the editor open, and since Cursor's 19 August 2026 changelog they can subscribe to events — a pull request, a Slack thread, or a cron schedule — and wake themselves up (cursor.com/help/ai-features/background-agents, cursor.com/changelog/08-19-26). Strength: genuine cloud scheduling, not a workaround. Weakness: it's an editor first — most people reach for it mid-session, not as a standing posting queue.

GitHub Copilot

Connects with .vscode/mcp.json; the key is requested once through a masked promptString input rather than written into a file you might commit (code.visualstudio.com/docs/agents/reference/mcp-configuration). That wiring is for agent mode inside the editor — interactive, nothing scheduled. Separately, Copilot's async coding agent runs unattended inside GitHub Actions once you assign it an issue (docs.github.com/copilot/concepts/agents/coding-agent/about-coding-agent) — a real background agent, but a different product surface, with its own MCP setup we haven't verified as the same seenpaid connection. Strength: the masked-key prompt is the safest file-based setup on this list. Weakness: the unattended half and the MCP-connected half aren't the same feature yet.

Windsurf

Connects with serverUrl and headers in ~/.codeium/windsurf/mcp_config.json — but that file configures Cascade, which Windsurf's own docs (now hosted at docs.devin.ai) call the legacy agent. Cascade was scheduled for end-of-life on 1 July 2026, and Devin Local has replaced it as the default; Devin Local reads MCP servers from Devin's own CLI config instead, a format we haven't verified, so we don't publish one for it. Devin Cloud, reachable from Windsurf's Agent Command Center, does run tasks on its own machine with the editor closed, but whether it shares Cascade's or Devin Local's MCP servers isn't confirmed either. Strength: was among the earliest code editors to speak MCP at all. Weakness: the ground moved twice in three months — confirm which agent you're actually running before trusting any config for it, ours included.

Gemini CLI

Here's the honest complication: the product most people mean by "Gemini CLI" stopped taking free and individual requests on 18 June 2026. Google's own developer blog says Gemini CLI and the Gemini Code Assist IDE extensions stopped serving Free, Google AI Pro and Google AI Ultra accounts that day, moved that traffic onto a new CLI called Antigravity CLI, and kept the original Gemini CLI running only for Google Cloud customers on a Gemini Code Assist Standard or Enterprise license (developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/).

So the setup that's actually current is for Antigravity CLI. It connects with a serverUrl and headers block in ~/.gemini/config/mcp_config.json (global) or .agents/mcp_config.json (per workspace) — note the field name: Antigravity's docs say the old Gemini CLI fields url and httpUrl "are not supported" anymore (antigravity.google/docs/mcp/). It still has a free Individual tier — unlimited Tab completions plus a weekly agent-request quota Google doesn't publish a number for and has cut more than once — with Pro ($19.99/mo) and Ultra ($99.99–$199.99/mo) raising the ceiling (antigravity.google/pricing/). Unattended posting isn't something its docs describe for MCP specifically — there's an interactive /mcp panel inside the CLI, no documented cron mode — so treat it as chat-first until shown otherwise. Strength: still free to start, and the same straightforward JSON shape as most of this list. Weakness: search "Gemini CLI" today expecting the tool Google shipped through mid-2026, and you land on a different product.

Grok

Two routes: a custom OAuth connector at grok.com/connectors — paste the URL, sign in, no key field — or, for developers, the xAI Responses API's mcp tool, with server_url and headers any script can call. Business and Enterprise need an admin to provision the connector first; the API route bills against your own xAI key. Neither the chat app nor the connector schedules anything on its own, but the API route is just an HTTP call, so anything you can cron — a script, a Lambda, a GitHub Action — can drive it unattended; that's work you'd build, not a feature Grok ships. Strength: the API route means a developer isn't stuck waiting on a chat window. Weakness: no documented static-key option in the consumer app — OAuth or nothing.

Hermes

Connects with a url and headers block in ~/.hermes/config.yaml (key), or auth: oauth in the same file to sign in instead. It's open source and self-run, so there's no seenpaid-relevant plan gate. Hermes goes further than anything else here on scheduling: a built-in cron scheduler, exposed to the agent itself as a single cronjob_manage tool, fires jobs while you're offline and delivers the result to Telegram, Discord or Slack (hermes-agent.nousresearch.com/docs/user-guide/features/cron). Each cron run starts a fresh session with no memory of your other chats, so a scheduled prompt has to carry everything it needs on its own. Strength: cron is a first-class tool call, not a workaround — "every weekday at 8, suggest a post and wait for my OK" is a request it's built to take literally. Weakness: you're the one keeping its Gateway daemon running.

OpenClaw

Connects with openclaw mcp set seenpaid '{...}' (key, with the transport explicitly set to streamable-http), or openclaw mcp add --auth oauth followed by openclaw mcp login to sign in instead. It's an open runtime with no plan gate documented. Like Hermes, it's built to run unattended by default: its Gateway is a persistent daemon, and its docs describe cron-driven openclaw agent runs with the runtime tracking which job is still active (docs.openclaw.ai/cli/daemon). Strength: infrastructure-first design means unattended is the normal case, not the edge case. Weakness: more to operate than a hosted chat app — a daemon you keep alive yourself.

Meta Muse

Ask it, in plain language, to add a custom connector for the URL, and it asks whether the server needs a key. seenpaid isn't in Meta's own connector directory yet — we submitted a listing on 20 September 2026 and it hasn't been approved — so the custom-connector route above is what's actually verifiable today, not a directory listing either of us can point to. Muse itself launched 8 September 2026, United States only, 18 and over. There's no evidence in Meta's own materials of a scheduled or cron mode: it's a personal-assistant app and WhatsApp bot, built for asking it things, not for standing jobs. Strength: it writes and tests the connector itself, conversationally — nothing to configure by hand. Weakness: the youngest product on this list, by far the smallest verified surface, and the directory approval that would make this easier isn't live.

Which one to actually use

Want it to post while you're not watching → Claude Code or Codex, both documented for cron; Hermes or OpenClaw if you want the schedule to live inside the agent itself.
Want zero config files → claude.ai, Claude Desktop or ChatGPT's plugin directory. OAuth sign-in, nothing to paste.
Building your own posting bot → Grok's xAI API, or any other MCP client's raw URL. Script it, cron it yourself.
Already live in an editor all day → Cursor's Cloud Agents are the most capable scheduled option built into one.
Searched for "Gemini CLI" → you're choosing Antigravity CLI now. Same idea, different product, different config field name.
On Windsurf → confirm whether you're on Cascade or Devin Local before pasting any MCP config, including the one above.
Want to try it in Muse → the custom-connector route works today; the app-directory listing doesn't exist yet.

The honest limits of mine

Since I'm a vendor on this list: seenpaid has no free plan, and the 7-day trial asks for a card — Solo starts at $19.99/mo. X is bring-your-own-key, so you pay X on top of us. TikTok videos land in your drafts; you publish from the app. New users can't connect Instagram or Facebook yet, pending Meta's review. The ChatGPT plugin is version 1.0.0, one day old at the time of writing. A Claude plugin is submitted but not yet listed — for now, Claude connects the way described above, as a custom connector.

Setup guides and further reading

Sources, all checked 24 September 2026: support.claude.com/en/articles/11175166, code.claude.com/docs/en/mcp and /headless, claude.com/pricing, developers.openai.com/api/docs/guides/developer-mode, help.openai.com (Scheduled tasks in ChatGPT), chatgpt.com/plugins, learn.chatgpt.com/docs/pricing, developers.openai.com/codex/noninteractive.md, cursor.com/docs/mcp, cursor.com/help/ai-features/background-agents, cursor.com/changelog/08-19-26, code.visualstudio.com/docs/agents/reference/mcp-configuration, docs.github.com/copilot/concepts/agents/coding-agent/about-coding-agent, docs.devin.ai/desktop/cascade/mcp, developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/, antigravity.google/docs/mcp/ and /pricing/, docs.x.ai/grok/connectors and /developers/tools/remote-mcp, github.com/NousResearch/hermes-agent (website/docs/reference/mcp-config-reference.md), hermes-agent.nousresearch.com/docs/user-guide/features/cron, docs.openclaw.ai/tools/mcp, /cli/mcp/registry and /cli/daemon, muse.ai/platform.

Which AI agent should run your social media? 12 compared