Agents

Agents on Echoia

Give an AI client access to one workspace — read its accounts, draft and schedule posts, triage comments — with limits that make handing over the keys reasonable.

Two pieces that solve different problems

The MCP server gives an agent the ability to act — six tools it can call. The skill gives it the knowledge of when to act — which tool fits, what to check first, what never to do without asking. Most setups want both.

Neither invents new permissions. An agent is exactly as capable as the token behind it, which is the point.

What an agent can do

CapabilityNeedsPublic effect
List accounts, posts, stats and commentsreadNone
Create drafts and scheduled postswriteNot until publish time
Publish a post immediatelypublishImmediate
Reply to a commentengageImmediate

Start read + write

A token with only read and write lets an agent do the useful part — research, draft, schedule — while everything it produces waits in the app for a human. That is a good place to stay until you trust the setup.

The guardrails

  • One workspace, always. No tool takes a workspace argument. An agent cannot see or touch another one.
  • Scopes are checked per tool call, not once at connection. A token without engage gets a clear refusal naming the missing scope.
  • Tools are hidden, not just refused. A token without engage is not offered reply_to_comment at all, so the agent does not plan around something it cannot do.
  • Rate limits are tighter for public actions — 20 replies a minute against 120 requests overall. A loop stops before it becomes a spam incident.
  • Quotas apply identically. An agent cannot spend more of your monthly allowance than you could yourself.

What an agent cannot do

  • Read or write in another workspace.
  • Obtain platform credentials — no endpoint returns them.
  • Connect or disconnect social accounts.
  • Change billing, invite people, or create further tokens.
  • Reply to a comment this workspace never received.

Revoke, don't reconfigure

If an agent behaves in a way you did not intend, revoke its token in Settings → Developers. It stops on the next request. Adjusting a prompt is not a security control.
Was this page helpful?