MCP vs. the in-app AI

Spreeflo's in-app AI chat and the hosted MCP server are two surfaces over the same thing: the same tool registry, the same workspace data, the same validation, and the same audit trail. A segment built through your AI client is indistinguishable from one built in chat.

What changes is who drives the model, and — the part that matters most — when an edit takes effect.

What's the same

Reads. Both surfaces see the same campaigns, segments, contacts, forms, and performance data.

Creation. Both create real assets in the workspace — and both create them dormant: drafts and unpublished forms, not live sends.

Validation. The same rules reject the same malformed documents, with the same errors.

The audit trail. Every applied change is recorded and attributed, and shows up in the workspace's change history.

The hard limit on sending. Neither surface sends, publishes, or deletes on its own authority. In the app, a person approves those actions before they happen; over MCP they're refused outright.

What's different

	In-app AI chat	Hosted MCP server
Who drives the model	Spreeflo	Your AI client
How an edit is expressed	A proposal staged for review	A direct tool call
When it applies	After a person approves it in the app	Immediately, on the call
Confirmation prompts	In-product approve / reject	None in Spreeflo — your client decides
Reads while a change is pending	Show the pre-approval state	Reflect the change as soon as it applies
Deletes, campaign start/resume, form publish	Available, with approval	Refused
Authoring long documents	Spreeflo writes the JSON from your brief, in your brand voice	Your client authors the raw JSON itself, using `read_format_doc`
Business profile & brand voice	Loaded automatically into every run	Not applied — your client only knows what it reads through tools
Memory across sessions	Yes	No
Code sandbox, file attachments	Yes	No
Who pays for the model	Spreeflo runs the model, so the cost sits with Spreeflo	Your client runs the model, so the cost sits with your client — governed here by a call rate limit instead

Approvals and confirmation

This is the difference that matters. In the app, when the AI decides to change something, it doesn't change it — it stages a proposal, and you approve or reject it. Nothing in your workspace moves until you say so, and reads keep showing the pre-approval state until you do.

Over MCP there is no staging step. A write tool applies the moment your client calls it, and the very next read reflects it. The review step hasn't been removed — it has moved into your AI client, which decides whether to ask you before it runs a write.

Warning: Over MCP, responsibility for confirming a write is your client's, not Spreeflo's. Spreeflo annotates every tool so a well-behaved client can tell reads from writes, but annotations are hints — not a security boundary. That's exactly why the irreversible actions are refused server-side instead of being left to client settings.

Practical consequence: check your client's confirmation settings before you start, and prefer the in-app AI when you want to see a change before it lands. See Check your client's confirmation settings.

Business profile and brand voice

In-app, your business profile and brand voice are loaded into every run automatically, so copy comes out sounding like you without being asked. Over MCP, none of that is applied — your client knows only what it reads through tools.

Tip: To get brand-consistent copy over MCP, put your brand voice and operating rules in your client's own system prompt or project instructions — and have it call get_business_profile before it writes anything.

Memory

The in-app AI remembers across sessions. The MCP server does not — it's stateless, and every request starts clean. Continuity, if you want it, comes from your client's own memory or project files.

Sandbox and files

The in-app AI has a code sandbox and can work with files you attach to a chat. Neither is exposed over MCP, along with the agent's memory tools and its agent-identity settings — those are in-app chat only.

Cost model

Spreeflo runs the model for in-app AI, so that cost sits with Spreeflo. Over MCP your client owns all the reasoning and Spreeflo exposes only deterministic tools, so the model cost sits with your client instead. The governor on the MCP side is a rate limit of about 120 tool calls per minute per workspace, with no daily quota — a refusal is always retryable after a few seconds.

How format specs load

Building a sequence, segment, form, or email means submitting a full JSON document that matches Spreeflo's wire format. In chat, Spreeflo loads the right specs for you and writes the JSON from your brief.

Over MCP your client authors that JSON itself, and fetches the specs on demand with read_format_doc. It's the same reference material either way — see Working with documents for the catalog and the read-before-write rules.

At a glance

Capability	In-app AI chat	Hosted MCP server
Read workspace data	Yes	Yes
Create drafts & unpublished assets	Yes	Yes
Edit existing assets	Yes, after approval	Yes, immediately
Pause a campaign or journey	Yes	Yes
Start or resume a campaign	Yes, after approval	No
Publish a form	Yes, after approval	No
Delete anything	Yes, after approval	No
Brand voice applied automatically	Yes	No
Memory across sessions	Yes	No
Code sandbox & file attachments	Yes	No
Spreeflo runs the model	Yes	No
Audited change history	Yes	Yes

Ready to connect? Connect your AI client walks through the setup.

Need Immediate Help? Contact our support team at support@spreeflo.com or check our other guides for more information.