Cursor Guide (2026): MCP, Agents & Automation

Cursor spent 2026 becoming two things at once: the code editor most developers already had open, and the reason SpaceX now owns a $60 billion AI company that has never built a rocket. Anysphere, the startup behind Cursor, was founded in 2022 by four MIT students, crossed $1 billion in annualized revenue by November 2025, and closed a SpaceX acquisition on August 14, 2026 that SEC filings value at $60 billion in all-stock consideration.¹² For anyone running social accounts through an AI agent, the part that matters is quieter: Cursor's Model Context Protocol support, its Marketplace, and its Cloud Agents are now the plumbing behind Grok Bot too, and posterly speaks to all of it over the same MCP server.

This is a long read (about 17 minutes). The table of contents on the right lets you jump. Every stat in this guide has a footnote, and where two sources disagreed on a number we say so rather than pick one silently. If you only want the connection steps, jump to Connect Cursor to posterly.

What is Cursor?

Cursor is an AI-powered code editor built by Anysphere, incorporated in 2022 by Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger while they were MIT students. The company raised an $8 million seed round led by the OpenAI Startup Fund in October 2023.³ Cursor describes itself as "a coding agent for building ambitious software," and by 2026 that description covers far more than autocomplete.⁴

The product now spans several surfaces:⁵

• Desktop IDE, with an Agent mode and a Plan Mode (triggered with Shift+Tab) where the agent asks clarifying questions, researches the codebase, writes a plan, waits for your approval, then builds; plans can be saved to your home directory or to the workspace.
• Cursor 3.0 (April 2, 2026) introduced an "Agents Window" for running many agents in parallel, "locally, in worktrees, in the cloud, and on remote SSH," with agents able to output interactive canvases such as reports and dashboards.⁶
• CLI, installed with curl https://cursor.com/install -fsS | bash on macOS and Linux, or a PowerShell one-liner on Windows. The command is agent, and it supports a headless mode (agent -p "..." --output-format text), an --mode flag for Agent, Plan, or Ask, and session resumption with agent resume or agent --continue.⁷
• Cloud Agents, described as "isolated VMs in the cloud with full development environments." You can start one from the desktop app's Cloud dropdown, the cursor.com/agents web app, an iOS app or Android PWA, a Slack mention, a GitHub or Bitbucket PR comment, Linear, or the API. Cloud Agents are billed at API pricing for the selected model and require a paid plan.⁸
• Slack integration, where @Cursor [prompt] starts a task, @Cursor list my agents checks status, and inline options like model=, repo=, branch=, and autopr=false control the run.⁹
• Bugbot, which reviews pull requests for bugs, security issues, and code quality problems, running automatically on every PR update or on demand with cursor review or bugbot run, across GitHub (including GitHub Enterprise Server), GitLab, Bitbucket, and limited Azure DevOps support. Bugbot launched in July 2025 and is billed on usage.¹⁰
• JetBrains IDE support, added March 2026.¹¹

Models

Cursor's model picker groups options into a first-party "Cursor Models" pool and third-party models. As of August 2026, the Cursor pool includes Grok 4.6 and Grok 4.6 Fast, Grok 4.5 and Grok 4.5 Fast, and Composer 2.5 and Composer 2.5 Fast. Third-party options include Anthropic's Claude Sonnet, Opus, Haiku, and Fable 5, OpenAI's GPT-5.x line up to GPT-5.6, Google's Gemini 2.5 through 3.7, Z.ai's GLM 5.2, and Moonshot's Kimi. An Auto setting has three modes, Auto Cost, Auto Balance, and Auto Intelligence, and third-party models cost an extra $0.25 per million tokens on Teams and Enterprise plans.¹² Cursor's own Composer models scaled up through partnership with SpaceXAI: Composer 1.5 "scaled reinforcement learning by over 20x," and Composer 2 reached "frontier-level performance at reduced cost."¹³ Grok 4.6 itself launched jointly with SpaceXAI on August 12, 2026 at $2 per million input tokens and $6 per million output tokens, with double the included usage in Cursor and Grok Build for its first week.¹⁴

Pricing

Cursor's consumer and team pricing, per cursor.com/pricing as of August 2026:¹⁵¹⁶

Cloud Agents are billed separately at API pricing for whichever model you pick, and Cursor requires you to set a spend limit the first time you use one.⁸

Key Cursor statistics in 2026

Anysphere was private for most of 2026, so growth numbers come from its own blog, financial press covering the SpaceX deal, and SEC filings rather than a single investor-relations page.

The gap between the ~$3 billion trajectory figure Forbes reported for April 2026 and the roughly $2.6 billion "annualized revenue" figure Yahoo Finance reported at the August close is real and unresolved in the public record; we present both rather than picking one. What is not in dispute is the pace: Anysphere reportedly went from a $400 million valuation in its 2024 Series A to $29.3 billion in November 2025, and was in the middle of raising $2 billion at a $50 billion valuation from a16z, Thrive, and Nvidia when SpaceX preempted the round.¹⁷¹⁸

How Cursor connects to tools: MCP, plugins and the Marketplace

Cursor's tool layer rests on the Model Context Protocol, the open standard for connecting an AI client to external data sources and actions.¹⁹ Cursor's own docs describe two config locations: .cursor/mcp.json in a project "for project-specific tools" and ~/.cursor/mcp.json in your home directory "for tools available everywhere."²⁰ Each entry sits under an mcpServers key.

Three transports

Cursor supports all three MCP transports: stdio, where Cursor launches and manages a local process; SSE, pointed at a URL for a Server-Sent Events endpoint; and Streamable HTTP, pointed at a URL for a plain HTTP endpoint.²¹ Stdio configs use command, args, and env fields; HTTP and SSE configs use url and optional headers. Cursor supports variable interpolation in these configs, including ${env:NAME} for environment variables, ${userHome}, ${workspaceFolder}, ${workspaceFolderBasename}, and ${pathSeparator}.²⁰

For servers that require OAuth, Cursor documents both browser-based OAuth flows and static client credentials placed directly in mcp.json, with callback URLs of https://www.cursor.com/agents/mcp/oauth/callback for the web and Cloud Agents, and http://localhost:8787/callback for desktop.²⁰ By default, "Cursor asks for approval before using MCP tools," and an Auto-review setting lets allowlisted tools run immediately without a prompt.²¹

One-click installs

Cursor also supports a deeplink format for installing a server without touching JSON: cursor://anysphere.cursor-deeplink/mcp/install?name=$NAME&config=$BASE64_ENCODED_CONFIG, where the config is the same JSON object base64-encoded.²² This is the mechanism most Marketplace listings and vendor "Add to Cursor" buttons use.

The Marketplace and plugins

Cursor 2.5, released February 17, 2026, introduced the Marketplace: "Plugins package skills, subagents, MCP servers, hooks, and rules into a single install," with launch partners including Amplitude, AWS, Figma, Linear, and Stripe.²³ There are two plugin formats: Agent Plugins, defined by a plugin.json at the repository root, and Cursor Plugins, defined under .cursor-plugin/plugin.json, which can additionally bundle rules, agents, commands, hooks, and variables. Every plugin is manually reviewed before publishing, and anyone can submit one at cursor.com/marketplace/publish. Team marketplaces are capped at one for Teams accounts and unlimited for Enterprise, with modes of Default Off, Default On, or Required.²⁴ Cursor 2.6, released March 3, 2026, added "MCP Apps," where a connected server can render an interactive UI inside Cursor, such as an Amplitude chart or a Figma diagram, alongside team marketplaces for plugins.²⁵ Browsing the Marketplace today turns up categories such as Productivity, Data & Analytics, Payments, Agent Orchestration, Sales, and Customer Support, plus individual plugins for X, Slack, Notion, Linear, and Figma; there is no listed Buffer, Hootsuite, or posterly plugin, which is why the manual MCP setup below is the right path today.²⁶

Rules and hooks

Two more mechanisms shape how an agent behaves once tools are connected. Rules, stored as .cursor/rules/*.mdc files, come in four types (Always, Apply Intelligently, Specific Files, Manual), and Cursor also reads an AGENTS.md file in the repository root or subdirectories, applying precedence in the order Team Rules, then Project Rules, then User Rules.²⁷ Hooks, configured in .cursor/hooks.json at the project level or ~/.cursor/hooks.json at the user level (with additional enterprise system paths), fire on events such as beforeShellExecution, beforeMCPExecution, afterFileEdit, beforeSubmitPrompt, and stop, and can return allow, deny, or ask. Cloud Agents only run the command hooks defined in a repository's own .cursor/hooks.json.²⁸

Connect Cursor to posterly

posterly ships one MCP server over two transports: a local stdio package on npm, and a hosted HTTP endpoint. Which one you use in Cursor depends entirely on which Cursor surface you are in, because Cursor's own docs draw a hard line between "runs on your machine" and "runs in a cloud VM."

Route 1: Cursor Desktop or CLI (local stdio)

Use this for the desktop IDE's Agent and Plan Mode, and for the agent CLI running on your own machine.

1. In posterly, open Dashboard, then API and MCP, enable the API and MCP add-on ($3/month on top of any plan; the Agency API tier is $29/month), and generate a key. Keys start with pst_live_.
2. Open Cursor, go to Settings, then MCP, then Add new global MCP server, or create the file yourself at ~/.cursor/mcp.json (every project) or .cursor/mcp.json inside a repository (that project only).²⁰
3. Add this entry:

{
  "mcpServers": {
    "posterly": {
      "command": "npx",
      "args": ["-y", "posterly-mcp-server@latest"],
      "env": {
        "POSTERLY_API_KEY": "pst_live_your_key_here"
      }
    }
  }
}

4. Save the file. Cursor lists the new server's tools under Available Tools, and the agent will call them automatically in chat or Plan Mode when a task needs them, asking for approval on each call unless you have added posterly to an auto-review allowlist.²¹ Ask it: "Run whoami and list_accounts on posterly" to confirm the connection.

If you have not signed up for posterly yet, install the same server without the env block. The public tools get_agent_signup_info, start_signup, and get_signup_session work before a key exists, so the agent can walk you through checkout in the browser and you add the key afterward, the same pattern posterly documents for other coding agents.²⁹

Route 2: Cloud Agents, Automations, and the @Cursor Slack bot (hosted HTTP)

Cloud Agents, scheduled Automations, and the Slack integration all run inside a cloud VM, not on your laptop, so the local npm package will not work there. Cursor's own Cloud Agent docs are explicit: "SSE and mcp-remote are not supported," and HTTP is recommended because "server configurations are never present in the cloud agent's VM environment," while a stdio server would only run if its command already existed inside that VM's image.³⁰ Use posterly's hosted endpoint instead:

{
  "mcpServers": {
    "posterly": {
      "url": "https://www.poster.ly/api/mcp",
      "headers": {
        "Authorization": "Bearer pst_live_your_key_here"
      }
    }
  }
}

Add this under Dashboard, then Integrations & MCP for a team-wide server, or directly in ~/.cursor/mcp.json with the url/headers shape shown above for a personal one; Cursor stores MCP configurations for Cloud Agents encrypted at rest.³⁰ The endpoint speaks JSON-RPC 2.0 over plain HTTP (no SSE in this version), and you can confirm it is reachable before pasting in a key with a plain GET https://www.poster.ly/api/mcp. Discovery metadata, including the full tool list, lives at GET https://www.poster.ly/.well-known/mcp/server-card.json.

One-click install

Because Cursor supports the cursor://anysphere.cursor-deeplink/mcp/install scheme, you can also base64-encode either config above and share it as a link or button; teammates click it, Cursor pre-fills the server name and JSON, and they only need to paste in their own posterly key.²² This is the same mechanism vendor "Add to Cursor" buttons on the Marketplace use.

What you get either way

Both routes reach the same 18 platforms: Instagram, Facebook, TikTok, X, LinkedIn, YouTube, Pinterest, Threads, Google Business, Telegram, Bluesky, Discord, Slack, Mastodon, Dev.to, Hashnode, WordPress, and Lemmy. The hosted HTTP endpoint exposes 75 tools; the local stdio package adds posterly's public setup and signup tools on top, for 78 total.²⁹ Full reference: posterly's MCP docs, the MCP overview page, and the REST API reference for scripts that skip MCP entirely.

What you can automate: agents, automations and social scheduling

Cursor's agent decides which tools to call in what order; naming the posterly tool in your prompt just makes it easier to check what actually happened.

1. Turn a shipped feature into a week of posts. Tools: generate_captions, find_available_slot, create_post. Prompt in Cursor's Agent, pointed at your own changelog file or PR description: "Read CHANGELOG.md, write five platform-native posts about this release for LinkedIn, X, and Threads, find a free slot for each this week with posterly, and show me the drafts before scheduling."

2. Turn a Cloud Agent PR review into a launch post. Tools: create_post, validate_post. Trigger a Cloud Agent from a GitHub PR comment; once it merges the change, have it validate and schedule an announcement post with validate_post before create_post, so a broken caption never reaches posterly's queue.

3. Run a scheduled content check with Automations. Tools: list_posts, get_account_analytics. Set up an Automation (cursor.com/automations) on a weekly schedule that spins up a cloud sandbox, calls list_posts to check for calendar gaps and get_account_analytics for the past week's performance, and posts a summary to Slack.³¹

4. Reply to Google Business reviews from a Slack thread. Tools: suggest_google_business_review_reply, audit_google_business_profile. @Cursor in Slack: "using posterly, draft replies for our three newest Google reviews and post the drafts in this thread for me to approve."

5. Repurpose a blog post across platforms. Tools: generate_captions, create_posts_batch. "Read this markdown file, write native versions for X, Bluesky, Mastodon, and LinkedIn, validate each, and schedule all four with posterly's create_posts_batch in one call."

6. Upload media generated in the same session. Tools: upload_media, create_post. Because a local Cursor Agent can read files on disk, it can hand a freshly rendered image or export straight to upload_media by file path; on Cloud Agents or the hosted endpoint, pass a public URL instead, since that transport is JSON-only.

7. Onboard a new client workspace. Tools: get_connect_link, list_accounts, whoami. "Create a connect link for Instagram and LinkedIn for the Acme workspace, then confirm the accounts appear in list_accounts once they're connected."

8. Pull analytics into a PR-ready report. Tools: get_account_analytics, list_accounts. "For every connected posterly account, pull the last 30 days of engagement and reach, and write it into REPORT.md as a table I can attach to next week's status update."

Cursor vs Claude Code vs Codex vs Grok Build

All four are coding-first agents that also happen to be capable MCP clients, and the comparison matters because posterly's setup steps differ by transport support.^20323334

The last row is worth calling out: xAI documents Grok Build as reading Cursor's own .cursor/mcp.json and project .mcp.json files directly, alongside Claude's ~/.claude.json, so a posterly server you configure for Cursor can already be visible to Grok Build without re-entering it.³⁴ Grok Bot goes the other direction, using your Cursor sign-in rather than a separate account, and inherits Cursor's MCP and plugin policy for teams.³⁵ For the consumer-facing chat assistants rather than coding agents, see our Claude guide, ChatGPT guide, and Grok Bot guide; posterly also has dedicated setup pages for Hermes, OpenClaw, and Poke, all speaking to the same MCP endpoint.

Where Cursor stands out for a social team specifically: it is the only one of the four with all three MCP transports in one client, a real plugin Marketplace with team-wide install policies, and both a desktop app and a cloud agent product under one account, so an agency can go from "test this connector on my laptop" to "run it as a scheduled cloud job" without switching tools.

Limits, security and gotchas

Usage is metered in dollars, not requests. Pro, Pro Plus, and Ultra plans each come with a dollar-denominated pool of API-priced usage. When you hit the cap, "You'll see a notification in the editor," and you can either enable pay-as-you-go on-demand usage or upgrade; unused usage does not roll over to the next period.¹⁵ The Hobby tier has no published fixed quota beyond "limited Agent requests," so check your dashboard rather than assume a number.³⁶

MCP tool lists can be capped in practice. Community reports on Cursor's GitHub and forum describe agents losing reliability once a client has roughly 40 MCP tools active across all connected servers; this is not documented in Cursor's current docs, so treat it as a practical ceiling to test for, not a hard published limit.³⁷

Cloud Agents cannot use Legacy Privacy Mode. If Privacy Mode is on, Cursor does not train on your code; on non-Business plans, some model providers may retain prompts for roughly 30 days regardless. Cloud Agents specifically require moving off Legacy Privacy Mode to a currently supported data setting.³⁸

Grok Build has had a real security incident. Reporting in August 2026 found Grok Bot's sibling CLI, Grok Build, uploading entire git repositories, including committed secrets, as part of its normal operation; that is a reason to audit what any coding agent sends off-device, not a reason to avoid MCP itself.³⁹

posterly-side limits. Create-post calls are capped at 100 requests per hour per key, with separate higher limits for media writes and reads. create_posts_batch bundles up to 25 posts into a single confirmed call. Keys are scoped, and a key limited to one workspace will reject calls for another. The hosted HTTP endpoint accepts base64 media up to roughly 5MB decoded; larger files need a public URL or the local stdio route, since Route 2 above is JSON-only.²⁹

Gotchas we see in support. A stray comma in mcp.json silently disables every server in that file; Node.js must be installed for the npx command to run; and a server added under .cursor/mcp.json inside a repo only applies to that project, so teammates cloning the repo will not automatically see it unless it is checked into version control and they trust the source, per Cursor's own advice to verify MCP servers before adding them.²⁰

Best for, not best for

Best for:

• Development teams that want the same MCP servers available in a local IDE, a CLI, and a cloud agent under one account
• Agencies packaging repeatable client workflows as Marketplace plugins with team-wide install policies
• Teams that also use Grok Bot, since Cursor's sign-in, MCP policy, and plugin marketplace carry over automatically
• Workflows that need to move a task from "test on my laptop" to "run on a schedule in the cloud" without re-authenticating tools
• Developers who want all three MCP transports (stdio, SSE, Streamable HTTP) in one client

Not best for:

• Non-technical marketers who want a chat-first assistant with no code editor attached; a plain assistant like Claude or ChatGPT is a better starting point
• Teams that need SSE or the mcp-remote bridge inside Cloud Agents; Cursor's own docs say neither is supported there
• Anyone expecting a built-in social scheduling plugin on the Marketplace today; posterly is not currently listed, so manual MCP setup is required
• Large video uploads over Cloud Agents or Automations, where the JSON-only hosted transport is a poor fit compared to a local stdio install

Final word

Cursor's 2026 has been unusually eventful for a code editor: a $60 billion acquisition, a new owner in SpaceXAI, a joint agent product in Grok Bot, and a Marketplace that turned MCP servers into one-click installs. None of that changes what actually happens when you add posterly as a tool. The agent still reads .cursor/mcp.json, still asks before it calls a write action, and still reaches the same 18 platforms whether it is running on your laptop or in a cloud sandbox triggered from Slack.

If you want to try it, create a posterly account and add the posterly MCP server to Cursor Desktop, the CLI, or your Cloud Agent config, then start with a read-only prompt like whoami before you let an agent publish anything.

Footnotes