Claude Guide (2026): MCP, Agents & Automation

Claude is the AI assistant that made agents talk to tools. Anthropic released the Model Context Protocol in November 2024, and by December 2025 there were more than 10,000 active public MCP servers, 97 million monthly SDK downloads, and adoption from OpenAI, Google, and Microsoft.¹ The business behind it has grown just as fast: Anthropic reported a $65 billion annualized revenue run rate at the end of July 2026 and a confidential IPO filing in June.² For marketers and agencies, the practical takeaway is simple. Claude is no longer a chat window that writes captions. It is a system that can be plugged into your scheduler, your analytics, and your review inbox, and left to work.

This is a long read (around 18 minutes). The table of contents on the right lets you jump. Every number in this guide has a footnote, and we update the page when Anthropic changes pricing, ships a new model, or moves MCP forward. If you only want the connection steps, jump to Connect Claude to posterly.

What is Claude?

Claude is the family of large language models and end-user products built by Anthropic, the San Francisco company co-founded by Dario and Daniela Amodei. Anthropic describes its mission as building systems that people can rely on and generating research about the opportunities and risks of AI.³

In August 2026 the model family looks like this:⁴

Claude Fable 5 launched on June 9, 2026 alongside the invitation-only Claude Mythos 5, and Anthropic said Fable 5's capabilities "exceed those of any model we've ever made generally available".⁵ Claude Opus 5 followed on July 24, 2026 as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price", and it became the default model on Claude Max and the strongest model on Claude Pro.⁶ Opus 4.7, released April 16, 2026, and the rest of the 4.x line remain available as legacy models.⁷

Claude reaches users through more surfaces than most people realise. Anthropic's pricing page lists the web app, iOS, Android, desktop chat, Claude Code, Claude Cowork, Claude Design, Claude Science, connectors with remote MCP, and desktop extensions.⁸ Claude Desktop runs on macOS and Windows.⁹ Claude in Chrome became a full Cowork session on August 12, 2026: "the side panel runs the same Claude Cowork session you use on desktop, web, and mobile for longer, multi-step work". It is Chrome-only, available to Max and Team immediately, Pro over the following weeks, and off by default for Enterprise.¹⁰ Claude Tag, Claude in Slack, launched in beta on June 23, 2026 for Team and Enterprise: tag @Claude with a request and it breaks the task into stages and works through them with the tools it can reach.¹¹

Claude pricing in 2026

Consumer and team pricing, per claude.com/pricing as of August 2026:⁸

Secondary press reported that from July 20, 2026 Max and Team Premium subscribers get Fable 5 access at roughly half of regular usage limits while Pro and Team Standard use usage credits with a one-time starting credit; that report is not from Anthropic, so check the help center before relying on it.¹²

Key Claude statistics in 2026

Anthropic is private, so the hard numbers come from its own announcements, financial press, and traffic estimators. Where we could not find a primary source we say so.

Two numbers you will see elsewhere are deliberately absent. Third-party aggregators cite around 245 million monthly active users and 300,000+ business customers, but we found no primary Anthropic figure for either, so treat them as unverified.¹³

How Claude connects to tools: MCP, connectors, plugins

The reason Claude matters for automation is not the chat interface. It is the plumbing.

MCP in one paragraph

The Model Context Protocol is "an open standard enabling developers to create secure, bidirectional connections between data sources and AI-powered applications". Anthropic open-sourced it on November 25, 2024 with early adopters Block and Apollo, and tooling companies Zed, Replit, Codeium, and Sourcegraph.¹⁴ OpenAI adopted it in March 2025 (Sam Altman: "People love MCP and we are excited to add support across our products"),¹⁵ Google followed in April 2025 (Demis Hassabis: "MCP is a good protocol and it's rapidly becoming an open standard for the AI agentic era"),¹⁶ and Microsoft announced native MCP support in Windows 11 at Build in May 2025.¹⁷ On December 9, 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare, and Bloomberg.¹⁸

The current specification (2025-11-25) defines two standard transports. stdio runs a server as a local subprocess and is the one clients "SHOULD support whenever possible". Streamable HTTP is the remote transport and replaces the older HTTP+SSE transport; servers must validate Origin headers to prevent DNS rebinding.¹⁹ That split, local versus remote, is the single most useful thing to understand about Claude's tool support, because different Claude surfaces support different transports.

Local servers in Claude Desktop

Claude Desktop reads a config file at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows, reachable from Settings, then Developer, then Edit Config. Each entry under mcpServers names a command and its arguments; the reference example launches a server with npx -y. Node.js is required, every tool action needs explicit approval in the UI, and logs land in ~/Library/Logs/Claude/mcp*.log.⁹ Since June 2025 there is also a one-click path: Desktop Extensions (.mcpb files, formerly .dxt) bundle an MCP server with a manifest.json so non-technical users can install without touching JSON.²⁰

Remote connectors in claude.ai, Cowork, and Chrome

Custom connectors using remote MCP are available on Free, Pro, Max, Team, and Enterprise, with Free users limited to one custom connector. The important architectural detail: "Claude connects to your remote MCP server from Anthropic's cloud infrastructure, rather than from your local device", so the server must be publicly reachable. Team and Enterprise owners add connectors in Organization settings before members see them.²¹ Anthropic also runs a connectors directory at claude.ai/directory for vetted, one-click integrations.²²

Claude Code

Claude Code, the terminal agent, adds servers with a single command: claude mcp add --transport http <name> <url> for remote servers, or claude mcp add <name> -- <command> [args...] for stdio. The SSE transport is deprecated in favour of HTTP. Servers can be scoped locally (the default, in ~/.claude.json), per project (a .mcp.json file shared through version control and requiring approval), or per user. /mcp handles OAuth, and tool output is warned at 10,000 tokens and capped at 25,000 by default.²³

Skills, plugins, and the API

Beyond MCP, Claude has three other extension points worth knowing. Agent Skills (October 2025) are "folders containing instructions, scripts, and resources" defined by a SKILL.md; Anthropic describes them as "custom onboarding materials that let you package expertise", and they work in the Claude apps, Claude Code, and the API.²⁴ Computer use in Cowork and Claude Code (research preview since March 2026 for Pro and Max) lets Claude "point, click, and navigate what's on your screen" when it lacks a tool.²⁵ And on the API, the MCP connector (beta) lets the Messages API call remote MCP servers server-side, while the Claude Agent SDK gives developers "the same tools, agent loop, and context management that power Claude Code" in Python and TypeScript.²⁶ posterly's own MCP server, CLI, and Claude Code plugin all sit on top of these primitives; the full menu is on the posterly agents page.

Connect Claude to posterly

posterly exposes 78 tools over its local stdio MCP server and 75 over the hosted HTTP endpoint, covering 18 platforms: Instagram, Facebook, TikTok, X, LinkedIn, YouTube, Pinterest, Threads, Google Business, Telegram, Bluesky, Discord, Slack, Mastodon, Dev.to, Hashnode, WordPress, and Lemmy. Both routes use the same key and the same paywall. The API and MCP add-on is $3 per month on top of any posterly plan (plans start at $7 per month); the Agency API is $29 per month.²⁷ Full reference: posterly MCP docs and the MCP overview page.

Because Claude has three different tool surfaces, pick the route that matches where you actually use Claude.

Route 1: Claude Desktop (local stdio)

1. In posterly, open Dashboard, then API and MCP, enable the add-on, and generate a key. Keys start with pst_live_.
2. In Claude Desktop open Settings, then Developer, then Edit Config.
3. Merge this into claude_desktop_config.json:

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

4. Restart Claude Desktop and ask: "Run whoami and list_accounts on posterly." Approve the tool calls when prompted.

If you have not signed up 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 Claude can walk you through checkout and password setup in the browser, then you add the key afterwards.²⁷

Route 2: Claude Code (local stdio, one command)

claude mcp add --scope user -e POSTERLY_API_KEY=pst_live_your_key_here posterly -- npx -y posterly-mcp-server@latest

--scope user makes the server available in every project rather than the current directory only.²³ Prefer a packaged install? The posterly Claude Code plugin bundles the MCP server, a skill, and a /posterly:setup flow: /plugin marketplace add awpthorp/posterly-agent then /plugin install posterly@posterly-agent. There is also a posterly CLI if you would rather script the same actions in shell.

Route 3: claude.ai, Cowork, Claude in Chrome (remote HTTP)

These surfaces cannot spawn a local process, so use the hosted endpoint as a custom connector.²¹

• Endpoint: POST https://www.poster.ly/api/mcp
• Auth header: Authorization: Bearer pst_live_your_key_here
• Wire format: JSON-RPC 2.0 (single request, no SSE in v1)
• Discovery: GET https://www.poster.ly/.well-known/mcp/server-card.json for the server card

Add it under Settings, then Connectors, then Add custom connector, paste the URL and key, then test with "whoami". The hosted transport is JSON-only, so for large video uploads either use Route 1 or hand Claude a public media URL and let upload_media fetch it.²⁷ The same endpoint works from any remote-only client, which is why our ChatGPT guide and Grok Bot guide point at it too. If you prefer OAuth over pasting a key, the Smithery listing wraps the same hosted endpoint.

Scheduling Claude itself

Connecting is half the job; the other half is making Claude run without you. Claude Code offers cloud Routines (Anthropic-managed, no machine required, connectors chosen per task, minimum interval one hour, set up with /schedule), local Desktop scheduled tasks with a one-minute minimum, and a session-bound /loop that expires after seven days.²⁸ Cowork has its own scheduled recurring tasks.²⁹ Pair a Routine with the posterly connector and "every Monday at 8am, draft and schedule this week's LinkedIn posts from the content plan" becomes a standing job.

What you can automate

Each workflow below names the posterly MCP tool that does the heavy lifting, plus a prompt you can paste into Claude. Claude decides which tools to chain; the tool names help you audit what it did.

1. Schedule a week of LinkedIn posts. Tools: generate_captions, find_available_slot, create_post (or create_posts_batch for up to 25 posts in one confirmed call). Prompt: "Using my LinkedIn account on posterly, write five posts for next week about our new onboarding flow, one per weekday, find the best free slot each day, and schedule them. Show me the drafts before creating anything."

2. Repurpose one blog post into five platforms. Tools: generate_captions, validate_post, create_post. Prompt: "Read this article [paste URL or text], then create platform-native versions for X, Threads, Bluesky, Mastodon, and LinkedIn. Validate each against platform limits before scheduling them across Tuesday and Wednesday."

3. Upload media with captions. Tools: upload_media, create_post. In Claude Desktop the local server can read a file path and use signed upload for larger files; on the hosted endpoint pass a public URL. Prompt: "Upload ~/Desktop/launch.mp4 to posterly, write a caption for Instagram Reels and TikTok, and schedule both for Thursday at 6pm."

4. Find the best time to post. Tools: find_available_slot, get_performance_profile (Pro plan and above). Prompt: "Look at my Instagram performance profile on posterly and tell me the best two windows this week, then reserve them for the two carousels in my drafts."

5. Pull analytics into a report. Tools: get_account_analytics, list_accounts. Prompt: "For every connected account, pull the last 30 days of follower growth, reach, and engagement rate from posterly and give me a table plus three sentences on what changed." Ask for presentation: "table" if you want Markdown tables in the reply.

6. Reply to Google Business reviews. Tools: suggest_google_business_review_reply, audit_google_business_profile, get_google_business_review_link. Prompt: "Draft brand-voice replies for the three newest Google reviews on our Denver location, flag anything I should read before you post, and generate a review request link I can text to today's customers." See how one team runs this on autopilot in our GBP reviews post.

7. Generate images and video for a campaign. Tools: generate_image, generate_video, create_post. Prompt: "Make three square images for our summer sale in the brand style, pick the best, and schedule it to Facebook and Pinterest with a caption that links to the sale page."

8. Onboard a new client account. Tools: get_connect_link, list_accounts, whoami. Prompt: "Create a connect link for Instagram and LinkedIn for the Acme workspace, send it to me, then confirm when the accounts show up in list_accounts."

Anthropic's own marketing team reports the shape of the payoff. In its January 2026 write-up, ad creation went from 30 minutes to 30 seconds, case study drafting from 2.5 hours to 30 minutes, and influencer marketing saved 100+ hours a month; the advice was "All you need to know is how to explain your challenge and what you're trying to solve in a very clear, concise manner."³⁰ Anthropic has not published a case study specifically about social scheduling, so treat those figures as directional for content work rather than a promise about your calendar.

Claude vs ChatGPT vs Grok Bot

The three agents most of our users ask about differ mainly in how they reach tools.^21233132

Read the deeper dives: ChatGPT guide and Grok Bot guide. Cursor is a fourth option for developers: it supports stdio, SSE, and Streamable HTTP, reads .cursor/mcp.json or ~/.cursor/mcp.json, does OAuth for remote servers, and offers one-click installs from its marketplace.³³ posterly also ships setup pages for Hermes, OpenClaw, Poke, and Grok Bot, and every one of them speaks to the same MCP endpoint.

Where Claude wins for social teams: it is the only one of the three that supports both transports across consumer surfaces, MCP is available on the Free plan, and the Skills plus Routines combination means an agency can package "how we post for client X" once and run it on a schedule. Where it does not: if your team already lives in ChatGPT or wants a cloud agent that never touches a laptop, the hosted posterly endpoint gives them the same tools without switching assistants.

Limits, security, and gotchas

Plan and usage limits. On consumer plans, usage across claude.ai, Claude Code, and Claude Desktop counts toward the same limit, and Pro and Max have five-hour session windows plus a weekly cap.³⁴ Long agentic runs that call posterly tools dozens of times will burn through a session faster than chat does. On the API, limits are per tier; the Start tier example for Opus 5 is 1,000 requests, 2M input tokens, and 400k output tokens per minute, with monthly spend caps of $500, $1,000, and $200,000 as tiers rise. Cached input tokens do not count toward the input limit.³⁵

Context. Fable 5, Opus 5, and Sonnet 5 accept 1M tokens of context and emit up to 128k; Haiku 4.5 is 200k and 64k.⁴ Big enough for a quarter of analytics, but remember Claude Code caps MCP tool output at 25,000 tokens by default (MAX_MCP_OUTPUT_TOKENS raises it), so ask for compact or paginated results.²³

Transport reach. claude.ai and Desktop custom connectors are remote-only: the server must be public. Local stdio servers run only in Claude Desktop through config or extensions, or in Claude Code.²¹ This is why posterly maintains both a package and a hosted endpoint rather than one.

Security. Anthropic's guidance is blunt: "only connect to trusted servers", because "malicious MCP servers may include hidden instructions".²¹ Its August 2025 agent framework spells out the threat, "Attackers could trick an agent into ignoring its original instructions (also referred to as 'prompt injection')", and the principle, "Humans should retain control over how their goals are pursued, particularly before high-stakes decisions are made", along with "Humans need visibility into agents' problem-solving processes".³⁶ Claude Desktop enforces per-action approval,⁹ and Claude Code has documented sandboxing.³⁷ On the posterly side, keys are tier-limited (Starter 1, Pro 2, Power 3, Agency 4), can be scoped to a single workspace, and create-post calls are capped at 100 requests per hour per key, with separate higher limits for media writes and reads.²⁷ Publishing a post is a high-stakes, public action: keep the "show me before you schedule" habit from the prompts above until you trust the workflow.

Gotchas we see in support. Config files use mcpServers in camel case (Hermes uses mcp_servers with an underscore); a stray comma in claude_desktop_config.json silently disables every server; Node.js must be installed for npx; and Windows paths in the config need escaped backslashes. Check ~/Library/Logs/Claude/mcp*.log first when a server does not appear.⁹

Best for, not best for

Best for:

• Teams that want one assistant across chat, terminal, browser, and Slack with the same connectors everywhere
• Agencies that need to package repeatable client workflows as Skills and run them on Routines
• Anyone who wants MCP without paying: remote connectors work on the Free plan
• Developers who want both local stdio and remote HTTP, plus an SDK that mirrors Claude Code
• Long-context jobs such as pulling 90 days of analytics across many accounts into one analysis

Not best for:

• Users who never leave ChatGPT and do not want a second assistant (use the hosted endpoint from there instead)
• Fully hands-off cloud agents with no human approval step; Anthropic's own framework argues against that for high-stakes actions
• Enterprise Chrome deployments where the side panel is off by default and needs admin enablement
• Very large media uploads through remote-only surfaces, where a local install or public URL is the better path

Final word

Claude's real advantage in 2026 is not any single benchmark. It is that Anthropic built the protocol the rest of the industry adopted, then shipped it in every surface it owns, including the free one. For a social team that means the gap between "Claude wrote my captions" and "Claude scheduled my month, pulled the numbers, and answered the reviews" is a JSON snippet or one claude mcp add command.

If you want to test it, create a posterly account through Claude or add the posterly MCP server to the Claude surface you already use, then start with the "show me before you schedule" prompts above.

Footnotes