Which MCP clients can operate pendpost?

Any MCP client. pendpost speaks MCP over stdio (npx -y pendpost --stdio) and over streamable-HTTP at http://127.0.0.1:8090/mcp, so Claude Desktop, Claude Code, and other MCP clients all drive the same tools. The web dashboard and the MCP surface are kept at parity by a test, so anything an agent can do you can also do by hand, and the reverse.

Can an agent publish without my approval?

Not unless you let it. By default every post is a draft and fail-closed: with no approval it will not publish, and by default that approval is you. You can opt into an owner-controlled auto-approve policy to let pendpost publish on its own, but only the owner can turn it on, and you can switch it off at any time.

How does an agent know how to operate pendpost?

pendpost ships an AGENTS.md contract: the rules an agent follows (the approval gate, why secrets never flow through the agent) plus a developer-portal playbook for each platform. Clients that read AGENTS.md, like Claude Code, pick it up from the project; any client can also call pendpost_health to get the same playbook data live. The dashboard Setup page adds a Copy AI prompt button that hands an agent the exact steps to connect one platform.

Do my API tokens pass through the agent?

No. Secrets are entered only by you, in a local ceremony you run yourself - the dashboard Setup page or your terminal - which writes your .env on your own machine. The agent never reads, types, or stores a token; it can set non-secret identifiers only. Your keys stay in your own .env, and pendpost never phones home.

Agents

Operate your social media from Claude or any MCP client

pendpost is MCP-native, so an AI agent operates it the same way you do. Connect Claude or another MCP client, hand it the setup prompt, and it drafts and schedules posts across Instagram, Facebook, LinkedIn, YouTube, and X. You approve before anything publishes; opt into owner-controlled auto-approve to scale toward hands-off. Below: the end-to-end loop, the copy-paste prompt, and three agent recipes you can run today.

End to end

How an agent operates pendpost, start to finish

Five steps, the same whether you drive Claude, Claude Code, or another MCP client. The agent does the work; you hold the gate.

Connect over MCP

Install the one-click Claude Desktop bundle, or run npx -y pendpost --stdio and point any MCP client at it. The agent can now see pendpost's tools alongside the same dashboard you use.
The agent drafts

Ask for a week of posts, a launch campaign, or one idea reworked per platform. pendpost stores them as drafts, and its read-only tools can never publish, so drafting carries no risk of an accidental post.
It schedules

The agent picks times per platform. Facebook and YouTube schedule natively and fire even when your computer is off; Instagram, LinkedIn, and X publish while pendpost is running.
You approve

Review the queue and approve what ships. Nothing publishes until it's approved, and by default that's you, so a human stays in control of what goes live.
Scale to auto-approve

Once you trust a lane, opt into owner-controlled auto-approve and pendpost publishes those posts on its own. Only you can turn it on, and you can switch it off at any time.

Connect

One command, then hand the agent the prompt

pendpost runs on your machine and exposes the same tools to any MCP client. Connect once, then paste the prompt below to put the agent to work.

One-click, Claude Desktop

# install the pendpost .mcpb bundle; it self-boots:
$ npx -y pendpost --stdio

Any stdio MCP client

# point any stdio client at this command
$ npx -y pendpost --stdio

Claude Code over HTTP

# with the server already running via npx pendpost
$ claude mcp add --transport http pendpost http://127.0.0.1:8090/mcp

Paste this into your agent

You can operate my social media through pendpost, connected over MCP. Read the
AGENTS.md contract first, or call pendpost_health for the live rules. Then:
1. Check which platform lanes are set up.
2. Draft a week of posts for the lanes that are ready (mock mode if none are).
3. Schedule them at sensible times per platform and queue them for my approval.
Hold everything as a draft. Never publish, and never approve your own post.
Tell me what you queued.

pendpost ships an AGENTS.md contract: the rules an agent follows to operate it, and a developer-portal playbook for each platform. To connect a single platform, the Setup page in the dashboard has a Copy AI prompt button that hands a browser-driving agent the exact steps, with no secret ever passing through the agent. For the full safety model, see the MCP server page.

Guardrails

What an agent can and cannot do through pendpost

The approval gate is a dial you own, not a wall the agent can move. The safety model is the same whether you approve by hand or opt into auto-approve.

The agent can

Draft and edit posts, and rework one idea into platform-native variants.
Schedule across Instagram, Facebook, LinkedIn, YouTube, and X.
Read insights, queues, and account health to plan the next move.
Publish on its own, but only once you have enabled auto-approve for that lane.

The agent cannot

Publish from a read-only tool. The read surface and the write surface are separate.
Push a draft live on its own. Nothing publishes until it's approved, and by default that's you.
Grant itself autonomy. Only the owner can turn auto-approve on.
Read or paste your secrets. Tokens are minted by a local CLI, never through the agent.
Keep firing into a platform that blocked you. A circuit breaker halts the lane.

Recipes

Three agent recipes you can run today

Copy a prompt, adapt it to your brand, and let the agent work. Each one holds posts for your approval until you decide otherwise.

Recipe 01

Weekly content agent

Run a standing agent that fills next week every Monday. It drafts to your cadence, schedules across the week, and leaves the queue for your approval.

Prompt

Every Monday, draft next week's posts across my set-up platforms: 3 for
LinkedIn, 5 for X, 2 for Instagram. Pull themes from my notes below. Schedule
them across the week at the times that did best last month, and queue them for
my approval. Do not publish anything.

Trigger it with your MCP client's scheduling, or run pendpost always-on so the agent keeps the week full without you.

Recipe 02

Scale to autonomous in 3 steps

Move from approving every post to hands-off, at your own pace. Autonomy is a dial only the owner can turn, and you can turn it back any time.

Review everything. The default. The agent drafts and schedules; you approve each post before it goes live.
Auto-approve one trusted lane. In Automation settings, turn on auto-approve for a single lane, say X, and watch a week. Auto-approval is recorded under a separate authority, never the agent's.
Go hands-off. Extend auto-approve to the lanes you trust and run pendpost always-on, so the agent drafts, schedules, and publishes around the clock while you supervise by exception.

Every publish still passes brand-lint, the anti-ban circuit breakers, and your schedule. Nothing publishes until it's approved, and auto-approve is owner-only and revocable.

Recipe 03

Zero-credential dry run

See the whole loop before you connect a single account. With no credentials in your .env, every platform runs through a mock driver: no network, nothing leaves your machine.

Start in mock mode

$ npx pendpost
# no credentials in .env means every lane runs mock

Prompt

We are in mock mode. Draft a 3-post launch campaign for a new feature, schedule
it over three days, and queue it for my approval, so I can watch the full draft
to publish loop end to end with no real API calls.

FAQ

Agent questions, answered

Which MCP clients can operate pendpost?: Any MCP client. pendpost speaks MCP over stdio (npx -y pendpost --stdio) and over streamable-HTTP at http://127.0.0.1:8090/mcp, so Claude Desktop, Claude Code, and other MCP clients all drive the same tools. The web dashboard and the MCP surface are kept at parity by a test, so anything an agent can do you can also do by hand, and the reverse.
Can an agent publish without my approval?: Not unless you let it. By default every post is a draft and fail-closed: with no approval it will not publish, and by default that approval is you. You can opt into an owner-controlled auto-approve policy to let pendpost publish on its own, but only the owner can turn it on, and you can switch it off at any time.
How does an agent know how to operate pendpost?: pendpost ships an AGENTS.md contract: the rules an agent follows (the approval gate, why secrets never flow through the agent) plus a developer-portal playbook for each platform. Clients that read AGENTS.md, like Claude Code, pick it up from the project; any client can also call pendpost_health to get the same playbook data live. The dashboard Setup page adds a Copy AI prompt button that hands an agent the exact steps to connect one platform.
Do my API tokens pass through the agent?: No. Secrets are entered only by you, in a local ceremony you run yourself - the dashboard Setup page or your terminal - which writes your .env on your own machine. The agent never reads, types, or stores a token; it can set non-secret identifiers only. Your keys stay in your own .env, and pendpost never phones home.

Start hands-on. Scale to hands-off.

Run pendpost in mock mode, connect your first platform, then let the agent take the wheel at the pace you trust.

Quickstart MCP server details Try mock mode Source on GitHub