Skip to main content

​
General

Conversion AI is connected to your workspace. It can read your contacts, audiences, emails, forms, workflows, and campaigns, and it can create or edit those assets, with your approval. A generic chatbot doesn’t know anything about your data; Conversion AI does.
No. Conversion AI cannot send a blast on its own. Blasts always land in draft and you publish them yourself. Other changes to your workspace either ask for an explicit approval before they run or are held for your end-of-turn review. See Approvals.
No. Each chat is independent. There’s no shared memory across chats today.
No. Conversion AI cannot delete contacts, audiences, emails, forms, workflows, campaigns, folders, fields, or tokens. If you ask, it’ll tell you it can’t and that you’ll need to do it from the UI.
No. Workflows are create-only — to change one, Conversion AI creates a new workflow. To replace an existing workflow’s behavior, activate the new one and deactivate (or rebuild against) the old one yourself.

​
Files and uploads

5 MB per file. Files larger than that are rejected when you try to attach them. There’s no fixed limit on the number of files per message, only the per-file size cap.
Images (JPEG, PNG, GIF, WebP), PDFs, and text-based formats (CSV, TSV, JSON, XML, YAML, plain text, code files, and similar). Conversion AI reads these directly.Other formats (like .xlsx or .zip) are dropped from the message itself, but the file is still saved to the sandbox at /vercel/sandbox/uploads/. You can still ask Conversion AI to open it with a script.See Files and sandbox for the full breakdown.
Yes. CSVs and TSVs are read as text. For large CSVs, Conversion AI typically opens them in its sandbox with Python and processes them there before answering.
Yes. Conversion AI reads PDFs directly, including text and embedded images.
Yes. Conversion AI can see and reason about images you attach. For example, “use this logo’s color palette” or “what’s wrong with this email screenshot”.
When Conversion AI generates a file, it surfaces it as a card in the chat with view and download actions. You don’t need to go anywhere else for it.

​
Sandbox and code

Yes. Conversion AI has a Linux sandbox with bash, Python 3.12+, and Node.js 20+. It uses the sandbox when a task needs real computation, like parsing a large file, generating a chart, or building a downloadable report.It also avoids writing code when a tool call would do. For example, “how many contacts match this filter?” is answered with a tool call, not a script.
Yes. The sandbox has network access. Conversion AI sticks to well-known libraries (e.g. pandas, openpyxl, requests).
No. The sandbox is isolated. Conversion AI uses tool calls to read or change anything in your workspace. The sandbox is just for files and scripts.
Within a single chat, files in the sandbox persist across messages, so Conversion AI can save something now and use it next turn. Between chats, sandboxes are separate. Don’t rely on the sandbox for long-term storage; if you want to keep a file, download it.

​
Edits and approvals

When it edits an existing email, form, audience, or campaign, those changes are proposals. They’re collected into a pending edit list and shown to you at the end of the turn to approve or reject, per asset, so you see exactly what’s changing before it’s written. See Creating and editing.
That’s expected. Proposed edits don’t apply while Conversion AI works — reads keep showing the current state until you approve at the end of the turn. They’re queued, not lost.
Yes. Multiple edits to the same asset in one turn are merged (latest value per field wins), so you review one combined change per asset rather than a stack.
Renaming a workflow or folder changes a live asset directly, so it asks first. Renaming an email, form, audience, or campaign is folded into that asset’s edit and reviewed at the end of the turn instead. See Approvals.

​
MCP and external tools

Yes. Conversion runs an MCP server. In ChatGPT, add Conversion from the app directory; in Claude, add it as a custom connector with https://mcp.conversion.ai/mcp (a directory listing is coming soon); other clients (like Cursor) connect by URL. Either way you sign in to authorize access to your workspace. See Connecting.
Almost. It exposes the same read and build actions, but not the sandbox and file tools, and it edits existing assets with raw, immediate tools instead of a reviewed edit list. See Differences from chat for the full breakdown.
Conversion’s in-chat Allow / Deny prompts only appear in chat. Over MCP, each tool tells your AI client whether it reads or writes data, and your client decides when to ask you to confirm. Review your client’s settings so you know when it will act without asking.