Skip to main content
The AI-native / MCP surface is verified on staging and rolling toward production. The tool catalog and the scope allowlist keep growing — this page tracks what’s live, what’s reserved, and what’s still dashboard-only.
Sapot Chat ships the agent surface deliberately: a capability is only reachable once its tools and its server-side guardrails are in place. Scopes that don’t have live tools yet are reserved — you can see them, but they refuse until their tools land. Everything below is forward-looking.

​
Reserved scopes

The scope catalog is intentionally larger than the live tool set. One scope is a reserved placeholder today: Because enforcement is a default-deny allowlist, a reserved scope grants nothing on its own. If a key carries automation:manage and tries to reach an unmapped endpoint, the server refuses with This endpoint is not available to agent access tokens — not a scope error, because no tool is wired up behind it. Granting it early has no effect and is safe; it simply lights up when automation tools ship.
The * (full) scope is not a bypass. It satisfies the scope check for every mapped endpoint, but it still can only reach endpoints that have live tools — reserved capabilities stay dark for * too.

​
What still lives in the dashboard

These actions are out of scope for an agent by design, not missing features. A human does them in the app.

Sending a blast

Agents draft campaigns and preview audience; the send stays a human dashboard click. There is no autonomous-send scope.

Building a flow

There is no create-from-intent, export, or import tool over MCP. Design flow agents in the Flow Designer; the agent operates and publishes them.

Channel & inbox setup

Connecting a WhatsApp / Facebook / Instagram channel (OAuth) and creating inboxes is UI-only — no MCP tools for it.

AI knowledge generation

Generate-from-text/file, bulk operations, and document import are omitted (LLM cost + blast radius). Agents do plain CRUD + approve.
The campaign and flow boundaries are permanent product decisions — see the Campaigns and Flow Agents pages for how to work within them. The knowledge, channel, and setup gaps are candidates for future tools.

​
Deferred controls

There is no per-minute, per-token rate limit yet. Today, the only cap is the per-plan MCP-action quota — the total number of successful agent actions your billing period allows. Exhaust it and calls return 429 until the next renewal (see the Authentication & Scopes page).
A blanket per-token throttle would regress existing personal- and bot-token integrations, and the plan quota already bounds total usage — so a dedicated per-token rate limit is a later, deploy-time step rather than a runtime feature. Design for the quota now; treat any future rate limiting as additive.

​
On the roadmap

Forward-looking, in rough order:
1

Intent-level flow creation

A create-from-intent path where an agent describes a goal and the server-side Designer compiles the graph — returning a summary and a preview link, never the graph itself. Keeps the IP boundary intact.
2

Automation tools

Wiring automation:manage to real endpoints (rules, macros, canned responses) so the reserved scope becomes live.
3

Per-token rate limiting & audit

A per-minute throttle alongside the plan quota, plus per-key usage and audit views in the Console.
4

Wider catalog

More thin tool wrappers over the existing REST surface as each one clears its safety review.

Authentication & Scopes

How scopes are enforced today, including the default-deny model and the plan quota.