Skip to main content
The MCP tools are safe by construction, but an agent still drives them best when it understands the rules before it starts. The Sapot Agent Skill is a companion skill you install alongside the MCP connection so your agent gets the mental model up front — instead of learning it by hitting refusals. It is a single skill file (mcp-server/skill/SKILL.md, skill name sapot-agent) that mirrors the tool catalog and teaches the agent how Sapot Chat governs it.
This is the MCP skill (sapot-agent) that teaches an agent to operate your account through the MCP server.

​
What it teaches

Connection contract

Auth is a scoped agent key sent as Authorization: Bearer, plus the X-Sapot-Account-Id header — the account is fixed by the connection, never a tool argument.

Default-deny scopes

A tool whose scope wasn’t granted returns an auth error naming the missing scope. The skill teaches the agent that this is expected, not a bug to retry.

The two boundaries

Campaigns are draft-only and flow agents are intent-level. The skill frames both as enforced boundaries so the agent doesn’t try to work around them.

429 = quota

A too_many_requests / “Agent action quota exceeded” means the plan’s MCP-action allowance is spent for the period — stop and tell the user.

Tools by scope

A grouped index of all tools under each scope, so the agent picks the right tool for the capability it has.

Recipes + footguns

End-to-end playbooks (triage, draft-a-blast, keep knowledge fresh, publish a flow, weekly readout) plus the sharp edges to avoid.

​
When it triggers

The agent loads the skill whenever a request involves working with your Sapot Chat account through the MCP tools — triaging or replying to conversations, managing contacts, reading reports, curating AI Knowledge, managing WhatsApp templates, operating flow agents, or drafting campaigns/blasts. In short: any time the user asks it to touch the inbox, contacts, campaigns, knowledge, or flows.

​
Install with Claude Code

1

Connect the MCP server

Register the Sapot Chat MCP server with your agent:
Replace https://mcp.sapot.ai/mcp with your environment’s MCP URL, YOUR_AGENT_KEY with your scoped key, and YOUR_ACCOUNT_ID with your numeric account id.
2

Place the skill alongside it

Copy mcp-server/skill/SKILL.md into your Claude Code skills directory (e.g. ~/.claude/skills/sapot-agent/SKILL.md), or bundle it in a plugin. Keep the skill and the MCP connection together so both are present in the same session.
3

Confirm it loads

Start the agent and ask it to work with your account (for example, “list my open conversations”). With the skill installed it should reference scopes and boundaries in how it plans the work — not just fire tool calls blind.
Grant the fewest scopes the task needs. The skill will tell the user which scope to add when a capability is missing, rather than looping on a refusal.

​
What it changes

With the skill installed, the agent avoids the failure modes that a cold MCP client falls into: The net effect is fewer wrong retries, correct draft-then-human-send behaviour, and reports that actually return data.

​
Keep it in sync

The skill mirrors the tool catalog in mcp-server/app/tools.py. As new scopes and tools ship, update the tools-by-scope list and recipes in SKILL.md so the agent keeps reaching for the right tool. Treat the skill as part of the tool surface, not as static docs.

Overview

What the AI-native MCP surface is and how it’s kept safe.

Quickstart

Connect an agent and make a first tool call.

Authentication & Scopes

The default-deny scope model in detail.