The short answer: if you already use Buffer or Metricool, start with the server your scheduler already publishes. If you want an agent that drafts from stored product context rather than from a prompt you retype each time, Cadencz is built for that. If you need the server running on your own infrastructure, Postiz is open source and self-hostable.
Cadencz publishes this guide and is included in it. That conflict is stated plainly above. The useful part of this comparison is not the tool counts, which are easy to inflate, but the five refusals in the next section. An MCP server holds write access to the accounts that are often a company's most public asset, and the questions worth asking are about what it cannot be talked into doing.
The comparison table
| Server | Best for | How it runs | Check this first |
|---|---|---|---|
| Cadencz | An agent that drafts and schedules from product context it can cite back to you | Hosted; Bearer token minted in Settings, Pro plan and above | Publishing requires confirm: true as a separate argument, and owner-only tools are hidden from other seats rather than merely refused |
| Buffer | Teams already on Buffer who want their existing queue reachable from an agent | Hosted; see Buffer's MCP page | Which plan tiers include the server, and which write actions it exposes |
| Metricool | Analytics-led work where the agent reads reporting as often as it posts | Hosted; see Metricool's MCP page | Whether the account scope covers every brand you manage |
| Postiz | Self-hosting and open source, with the server on infrastructure you control | Self-hosted alongside your own Postiz instance | You own the runtime, the upgrades, and the credential storage that a hosted server would own for you |
Five questions that separate an agent-ready server from an API wrapper
- Does publishing take a second, explicit confirmation, or can an agent infer it from an ambiguous instruction?
- Does the token carry the role of the person who minted it, or is it an admin key that ignores your seat permissions?
- Are tools the caller may not use hidden from the tool list, or merely refused when called? A tool an agent can see is a tool it will try.
- Can the server spend money, cancel a plan, or mint credentials? A leaked token should not be able to reach your billing.
- Do refusals come back as structured errors an agent can act on, or as prose it has to guess at?
How Cadencz answers those five
The Cadencz server exposes sixteen tools: eight that read, six that write, and two that are owner-only. The owner-only pair does not appear in the tool list at all for other seats, so an agent holding a scheduler's token has no way to learn those tools exist. The same predicate governs both the listing and the call, because a cosmetic listing check and a load-bearing call check are exactly the pair that drifts apart.
publish_now is the one irreversible tool, and it refuses unless confirm is passed as true in a separate argument. Nothing publishes without a human, and there is no setting that changes it. A token acts as the person who minted it, role included, so removing a teammate ends their token with their seat, and a downgrade stops existing tokens on the next call rather than at mint time.
Nothing that moves money exists in the tool list. No checkout, no cancel, no billing portal, nothing that mints credentials. connect_channel_link returns a URL where a human connects a platform; it never accepts credentials.
Choose a different server when its specialty matches the job
- Choose Buffer's server when your queue, channels, and team habits already live in Buffer and moving them costs more than the integration is worth.
- Choose Metricool's server when the agent's main job is reading analytics and reporting across several brands rather than drafting.
- Choose Postiz when open source, self-hosting, or data residency is a requirement, and you have somewhere to run it.
- Choose a general automation bridge when you need one agent to touch many unrelated systems and social publishing is only a small part of it.
- Choose no MCP server at all when a human writes and schedules every post anyway. The integration earns its keep only when the agent is already in the room where the work happens.
Connecting Cadencz takes one line
Any client that sends a static Authorization header works the same way, including Claude Code, Claude Desktop, Codex and Cursor. Tokens are minted in Settings under MCP by an account owner, and the raw string is shown exactly once because only its hash is stored. There is no support path that can hand a token back to you, which is the same property that means a leaked one can be revoked without touching anything else.
Add the server
claude mcp add --transport http cadencz https://api.cadencz.com/mcp --header "Authorization: Bearer cdz_mcp_..."