OpenClaw Guide (2026): Skills, MCP & Setup

OpenClaw is the open-source project that turned "connect your chat apps to an AI agent" from a weekend hack into, by some measures, the fastest-growing open-source project in GitHub's history.¹ It is a self-hosted, MIT-licensed gateway, not a hosted service: you install it, it runs on your machine, and it bridges WhatsApp, Telegram, Discord, Slack, and 25 other channels to whichever AI model and tools you point it at.² For a marketing or agency team already scheduling posts through posterly, that means a single running process can turn a WhatsApp message into a scheduled LinkedIn post, without a browser tab or a dashboard in sight.

This is a long read (about 18 minutes). The table of contents on the right lets you jump. Every stat has a footnote, and where OpenClaw's own docs were silent we say so rather than guess. If you only want the connection steps, jump to Connect OpenClaw to posterly.

What is OpenClaw?

OpenClaw's own docs describe it as "a multi-channel gateway for AI agents that runs on any OS," and more fully as a "self-hosted gateway that connects your favorite chat apps... to AI coding agents."² Under the hood, a single Gateway process is "the single source of truth for sessions, routing, and channel connections," with a local Control UI on 127.0.0.1:18789, a CLI, a Web Control UI, a macOS app, and iOS and Android nodes.² Everything reads from one config file at ~/.openclaw/openclaw.json.

The project was created by Peter Steinberger, and its README credits him and the community: "OpenClaw was built for Molty, a space lobster AI assistant, by Peter Steinberger and the community."³ It is not affiliated with Anthropic; the official site states this directly.⁴ In February 2026 Steinberger joined OpenAI, and drew a clear line around governance: "OpenAI hired me, not OpenClaw. The OpenClaw Foundation is independent, with sponsors rather than owners."⁵ The OpenClaw Foundation, a nonprofit formed on July 8, 2026, now stewards the project, with a stated mission of "Bringing people closer to AI."⁶ Sponsors listed on openclaw.ai include OpenAI, GitHub, NVIDIA, Vercel, Blacksmith, and Convex.⁴

Name history and a naming collision worth knowing

OpenClaw has changed names twice since it first appeared. According to Wikipedia, the project released on November 24, 2025 as Clawdbot, renamed to Moltbot on January 27, 2026 after an Anthropic trademark complaint, then renamed again to OpenClaw on January 30, 2026.⁷ If you see older screenshots or blog posts referring to "Clawdbot" or "Moltbot," they are describing the same project under an earlier name.

The name also collides with two unrelated GitHub projects: pjasicek/OpenClaw, a reimplementation of the 1997 platformer Captain Claw, and Gen-Verse/OpenClaw-RL, an unrelated reinforcement-learning training repo.⁸ Neither has anything to do with the AI gateway covered in this guide.

How OpenClaw is installed and licensed

OpenClaw ships an install script (curl -fsSL https://openclaw.ai/install.sh | bash for macOS, Linux, and WSL2, or a PowerShell equivalent for Windows), an npm package (npm install -g openclaw@latest --allow-scripts=openclaw), and Docker images at ghcr.io/openclaw/openclaw:latest.⁹ It needs a fairly recent Node.js runtime (the docs recommend Node 26) and, for the Docker route, at least 2 GB of RAM.¹⁰ Its README and docs describe the license as MIT; GitHub's own repository metadata currently reports the license file as "Other/NOASSERTION," so this guide states it as "MIT-licensed per README and docs" rather than as an unambiguous fact.³

There is no OpenClaw pricing page, subscription tier, or paid plan. It is open source; the only ongoing cost is whichever model API keys you plug in, at each provider's own rates.⁴ Any guide describing OpenClaw as a hosted or managed service is describing a third-party host, not the official project, since none of the third-party hosts (clawcloud.sh, kilo.ai, getopenclaw.ai) are official.¹¹

Key OpenClaw statistics in 2026

OpenClaw has no company publishing quarterly numbers, so the figures below come from GitHub's API, npm's download API, and independent security research.

TechCrunch, covering Steinberger's move to OpenAI, called OpenClaw "the fastest-growing open source project in GitHub history."¹² Microsoft's internal "Scout" agent for Teams, Outlook, and OneDrive is reportedly built on OpenClaw, and Wikipedia notes that China restricted state agencies, state-owned enterprises, and banks from using it in March 2026.¹³ Treat the exact "fastest growing" framing as a characterization from press coverage rather than a number OpenClaw itself publishes.

How OpenClaw connects to tools: skills, ClawHub, MCP and automations

OpenClaw reaches outside tools through four overlapping mechanisms: skills, the ClawHub registry, MCP servers, and a scheduling layer now called automations.

Skills

A skill is a folder with a SKILL.md file. OpenClaw's docs say the file needs "at minimum a name and description in the frontmatter," with optional fields such as user-invocable, disable-model-invocation, command-dispatch: "tool", and an metadata.openclaw block that can gate a skill to a specific OS or require certain binaries, environment variables, or config keys.¹⁴ Skills load from a defined precedence order: a workspace's skills folder first, then .agents/skills, then ~/.agents/skills, then a state directory, then bundled skills, then any extra directories set in config.¹⁴

Installing a skill uses one of these documented forms:¹⁴

openclaw skills install @owner/<slug>
openclaw skills install skills-sh:owner/repo/slug
openclaw skills install git:owner/repo@ref
openclaw skills install ./path --as my-tool

Add --global to install into ~/.openclaw/skills instead of the current workspace, openclaw skills update --all to refresh everything, and openclaw skills verify @owner/<slug> to check a skill's signature before trusting it. There is no documented bare openclaw skills install <repo> shorthand; use the git:owner/repo@ref or @owner/<slug> forms above.

ClawHub

ClawHub, at clawhub.ai, is the community registry for OpenClaw skills and plugins. Its own copy promises "skills and plugins from top creators" with "signed manifests" and "moderated releases," and creators publish with npm i -g clawhub followed by clawhub skill publish.¹⁵ Its public footer showed 30 skills and 12 plugins on the day we checked, which looks like a curated highlight slice rather than a full count, since a February 2026 security audit scanned 2,857 skills from the same ecosystem.¹⁵ Because ClawHub content is community-submitted, treat every skill as untrusted code until you have read it; OpenClaw's own docs say exactly that.¹⁴

MCP

MCP, the Model Context Protocol, is the open standard OpenClaw uses to reach structured tool servers rather than loose skills. OpenClaw supports all three common transports: stdio, Streamable HTTP, and SSE.¹⁶ Servers live under mcp.servers in ~/.openclaw/openclaw.json, with per-server fields for url or command, transport, enabled, connectionTimeoutMs, requestTimeoutMs, and a toolFilter with include and exclude patterns. Servers that need it can use auth: "oauth" and openclaw mcp login <name>.¹⁶ The CLI has matching commands: openclaw mcp add to add a server interactively, openclaw mcp doctor <name> --probe to test a connection, and openclaw mcp reload to pick up config changes without restarting the gateway. Once connected, MCP tools "go through the same tool-profile and tool-policy controls as everything else," meaning they are gated the same way skills are.¹⁶

Memory and multi-agent routing

Two related pieces round out the picture. Memory is plain files: USER.md, MEMORY.md, dated files under memory/, and DREAMS.md, with the docs stating flatly, "the model only remembers what gets saved to disk; there is no hidden state."¹⁷ For teams running more than one persona or client, agents.entries defines multiple agents, each with its own workspace, and bindings route specific channels or account IDs to a specific agent, with "first match in config order wins."¹⁸

Automations (formerly cron jobs)

OpenClaw's scheduling layer is documented today as automations; the older "cron jobs" URL now redirects to the same page. A full example from the docs:¹⁹

openclaw automations create "0 7 * * *" "Summarize overnight updates" \
  --name "Morning brief" --tz "America/Los_Angeles" \
  --session isolated --announce --channel slack --to "channel:C1234567890"

Schedules can also be set with --at, --every, or --cron (5 or 6 fields), sessions can be main, isolated, current, or a specific session:<id>, and delivery can go through --announce, a --webhook URL, or --none. A job auto-disables after 10 consecutive failures, and its state lives in a local SQLite database; you can turn the whole feature off with cron.enabled: false or the OPENCLAW_SKIP_CRON=1 environment variable.¹⁹ OpenClaw also exposes webhooks directly: hooks.enabled, a hooks.token, and a path (default /hooks), with POST /hooks/wake and POST /hooks/agent endpoints authenticated by an Authorization: Bearer <token> header.¹⁹

Connect OpenClaw to posterly

posterly exposes the same MCP server two ways: a local stdio package on npm and a hosted HTTP endpoint. OpenClaw can use either, which matters because you may be running OpenClaw on a home server with no public IP, or on a VPS where a hosted endpoint is simpler to manage. Both routes use the same posterly API key and the same paywall: the API and MCP add-on is $3 a month on top of any posterly plan (plans start at $7 a month), and the Agency API tier is $29 a month.²⁰ Full reference: the posterly MCP docs and the MCP overview page.

Step 1: get a posterly API key

Sign up for posterly, then open Dashboard, then API and MCP, enable the add-on, and generate a key. Keys start with pst_live_. If you would rather have an agent handle signup itself, posterly's public get_agent_signup_info and start_signup tools work before a key exists; see how AI agents sign users up to SaaS for the pattern and the agents signup page to start.

Step 2a: local stdio, if OpenClaw runs on the same machine as Node

Add this under mcp.servers in ~/.openclaw/openclaw.json:

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

Step 2b: hosted HTTP, for a remote gateway or a simpler config

{
  "mcp": {
    "servers": {
      "posterly": {
        "url": "https://www.poster.ly/api/mcp",
        "transport": "streamable-http",
        "enabled": true,
        "connectionTimeoutMs": 5000,
        "requestTimeoutMs": 20000
      }
    }
  }
}

For the hosted route, pass your posterly key as an Authorization: Bearer pst_live_your_key_here header rather than a config literal. OpenClaw's own MCP docs advise this directly: "keep credentials out of config literals; store sensitive headers and environment values through the supported secret mechanisms."¹⁶ GET https://www.poster.ly/api/mcp returns basic server info unauthenticated, which is a quick way to confirm the URL before wiring in a key, and GET https://www.poster.ly/.well-known/mcp/server-card.json returns discovery metadata for the server.

Step 3: add and reload

Whichever form you used, run openclaw mcp add to walk through the same fields interactively if you prefer not to hand-edit JSON, then openclaw mcp reload to apply the change without restarting the gateway. Use openclaw mcp doctor posterly --probe to confirm the connection is live.¹⁶

Step 4: verify with a read-only message

Message your OpenClaw gateway from any connected channel: "using the posterly MCP server, run whoami and then list_accounts." whoami confirms the key, workspace, and scopes you are working with; list_accounts lists your connected social accounts. A failed call almost always means either the add-on is not enabled on your posterly plan or the key was pasted into the wrong field.

Prefer a skill over raw MCP config?

If you would rather not hand-edit openclaw.json at all, install posterly's own skill from its GitHub repository using OpenClaw's documented git form:

openclaw skills install git:awpthorp/posterly-agent@main

The skill runs in hybrid mode: it calls posterly's MCP tools when your OpenClaw setup has MCP configured, and falls back to posterly's REST API using the same POSTERLY_API_KEY when it does not, so it works even on a minimal OpenClaw install.²¹

What you can automate from WhatsApp, Telegram, Slack and Discord

Every workflow below is a chat message OpenClaw routes to a posterly MCP tool. Because OpenClaw supports 29 channels, the same prompts work whether you type them in WhatsApp, Telegram, Slack, Discord, iMessage, Signal, or Microsoft Teams.²

1. Schedule a week of posts by texting a brief. Tools: generate_captions, find_available_slot, create_post. Message: "Write five LinkedIn 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 first."

2. Repurpose a blog post from a Slack thread. Tools: generate_captions, validate_post, create_post. Message: "Take the article linked above and create native versions for X, Threads, Bluesky, Mastodon, and Facebook. Validate each against platform limits before scheduling across Tuesday and Wednesday."

3. Upload media straight from a chat attachment. Tools: upload_media, create_post. In a local stdio setup, OpenClaw can hand a downloaded file path directly to upload_media; on the hosted route, use upload_media_from_url with a public link. Message: "Upload this video, write a caption for Instagram Reels and TikTok, and schedule both for Thursday at 6pm."

4. Pull analytics into a Discord update. Tools: get_account_analytics, list_accounts. Message: "For every connected account, pull the last 30 days of follower growth and engagement rate and post a short summary here."

5. Reply to Google Business reviews from Telegram. Tools: suggest_google_business_review_reply, audit_google_business_profile. Message: "Draft brand-voice replies for our three newest Google reviews and send them here for approval before anything posts."

6. Run a Monday content check as an automation. Combine list_posts with find_available_slot inside an openclaw automations create job that fires every Monday at 7am and announces to a Slack channel: "check this week's content plan for gaps and propose three posts to fill them."

7. Onboard a new client account by text. Tools: get_connect_link, list_accounts, whoami. Message: "Create a connect link for Instagram and LinkedIn for the Acme workspace and text it to me."

8. Generate and schedule a campaign image. Tools: generate_image, create_post. Message: "Make three square images for our summer sale in the brand style, pick the best, and schedule it to Facebook and Pinterest."

Before any of the create calls, validate_post checks a full payload against a platform's rules without consuming quota, which is worth keeping in an automation's prompt since a scheduled job might attempt the same task again on its own next run.

OpenClaw vs Hermes vs Poke vs Claude and ChatGPT

Hermes Agent, Nous Research's similarly self-hosted MIT-licensed alternative, describes itself as "one agent, one memory, every surface," connects to any MCP server, and installs with curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash.²² Poke, from The Interaction Company of California (founders Marvin von Hagen and Felix Schlegel), takes the opposite approach: a fully hosted, closed product reachable over iMessage, SMS, Telegram, and WhatsApp, raised on a $15 million seed plus $10 million more at a $300 million post-money valuation, positioned explicitly as a "no terminal" alternative to OpenClaw.²³ Claude and ChatGPT both reach posterly through the hosted MCP endpoint described in the Claude guide and posterly's other agent pages, but neither offers OpenClaw's chat-app-native, self-hosted model. posterly maintains dedicated setup pages for all of these: Hermes, Poke, OpenClaw, and Grok Bot, all built on the same MCP endpoint.

Where OpenClaw wins for a marketing team: it is free, it reaches more chat channels than any competitor on this list, and its skill and automation system means an agency can package "how we post for client X" once and run it unattended on a schedule. Where it does not: you are the one who has to install, secure, and keep the Gateway patched, and the third-party skill ecosystem carries real, documented risk that a fully hosted product like Poke does not expose you to.

Limits, security and gotchas

Exposed instances are a real, documented problem. SecurityScorecard found 40,214 exposed OpenClaw instances in February 2026, 12,812 of them exploitable for remote code execution, and Censys counted 21,639 separately at the end of January.²⁴ A named vulnerability, ClawJacked (CVE-2026-25253), scored 8.8 on CVSS.²⁴ OpenClaw's own security docs are direct about why: "Your AI assistant can execute arbitrary shell commands, read/write files, access network services, and send messages to anyone," and the default configuration binds to loopback with pairing required for unknown senders, with an explicit instruction: "Never expose unauthenticated on 0.0.0.0."²⁵ Run openclaw security audit [--deep|--fix] before exposing the Gateway to any network beyond your own machine.

ClawHavoc: malicious skills in the ecosystem. Koi Security audited 2,857 ClawHub skills in early February 2026 and found 341 malicious, an incident researchers named "ClawHavoc"; Antiy later counted 1,184 in a separate pass.²⁶ Cisco separately tested a skill in January 2026 and reported both data exfiltration and prompt injection, concluding bluntly: "The skill we invoked is functionally malware."²⁷ OpenClaw's docs tell users to "treat third-party skills as untrusted code" and to read them before enabling, which is the single most important habit for anyone installing from ClawHub or a random GitHub link.¹⁴

The WhatsApp channel is not the official Business API. OpenClaw's WhatsApp integration goes through WhatsApp Web via the unofficial Baileys library, with QR-code-only login: "Login is QR-only." The docs recommend a separate WhatsApp number for this reason.²⁸ Meta banned general-purpose AI chatbots from the official WhatsApp Business Platform starting January 15, 2026, the same day OpenAI's ChatGPT left WhatsApp; OpenClaw sidesteps that specific ban by using the personal-account protocol instead, though the terms-of-service exposure of that route is not addressed in OpenClaw's own documentation.²⁹

Sandboxing helps but is not a hard boundary. agents.defaults.sandbox.mode can be off (the default), non-main, or all, backed by Docker, Podman, SSH, or an open shell, with Docker defaulting to no network access and a read-only root filesystem. The docs are honest about the limit: "not a perfect security boundary."³⁰

A security researcher's summary is worth repeating. SecurityScorecard's Jeremy Turner: "Don't just blindly download one of these things and start using it on a system that has access to your whole personal life."²⁴

posterly-side limits. Create-post calls are capped at 100 requests per hour per key, with separate higher limits for media writes and reads, and create_posts_batch bundles up to 25 posts into one confirmed call. Keys can be scoped to a single workspace, so a compromised OpenClaw install with a scoped key cannot reach every client's accounts, only the one it was issued for.

Best for, not best for

Best for:

• Technical teams and agencies who already run their own servers and want a free, self-hosted agent that reaches WhatsApp, Telegram, Slack, and Discord from one process
• Power users comfortable editing a JSON config and running a CLI to add MCP servers, skills, and scheduled automations
• Anyone who wants the widest chat-channel coverage of any agent on this list, without a subscription
• Teams that want to audit the exact code running their agent, since OpenClaw is open source

Not best for:

• Anyone who does not want to install, patch, and secure a server themselves; a fully hosted option such as Poke or the Claude and ChatGPT connectors avoids that entirely
• Teams that would install ClawHub skills without reading them first, given the documented rate of malicious submissions in past audits
• Regulated environments where the underlying company or agency policy restricts self-hosted AI tools with broad system access
• WhatsApp workflows where using a personal-account protocol instead of the official Business API is a compliance concern

Final word

OpenClaw's pitch is unusual for 2026: instead of a managed product with a monthly bill, it is a piece of software you run yourself, and it has grown into one of the largest open-source projects on GitHub doing exactly that. For a social team, the appeal is real: one process, 29 chat channels, and a config file that can point at posterly's MCP server the same way it points at any other tool. The catch is equally real: you are the security team now, and the numbers on exposed instances and malicious skills are not hypothetical.

If you already run OpenClaw, create a posterly account or add the posterly MCP server to your openclaw.json, start with the read-only whoami and list_accounts check above, and keep any automation's first few runs on "show me before you schedule" until you trust it.

Footnotes