Deynix 為 AI 智能體提供營銷工作區，貼文仍須本人批准

Hi dev.to, I'm Kyle. I'm a developer, and like most developers I'm bad at marketing what I build. I know I should post about releases, but every social media tool I tried wanted me to open another dashboard and stare at a blank text box.

Meanwhile I was already spending half my day talking to Claude. So I started wondering what would happen if the assistant I already use did the marketing work, and all I had to do was say yes or no.

That idea became Deynix, "the marketing layer for AI agents". It's live at deynix.com, and we're launching properly on 27 October. This post covers what it is, how it's wired up, and an honest look at where it stands.

What it is

Deynix is a social media workspace that you can drive two ways: through a normal web app, or through your AI assistant over MCP.

• It reads your website and builds a brand kit and content strategy from it.
• Drafts go into a Review queue. Nothing publishes until you approve it.
• A remote MCP server exposes around 70 tools, so Claude, ChatGPT, Cursor, VS Code or any MCP client can read your analytics, draft posts and X threads, render branded images and carousels, draft replies to your mentions, and schedule approved posts.
• X, LinkedIn, Facebook, Instagram and Bluesky work for everyone today. Threads, YouTube, TikTok and LinkedIn company pages are built and switch on as each platform's API approval comes through. ## Why MCP, and why approval-gated

I didn't want to build yet another "AI writes your posts" box. The model you already chat with is better context for your work than any prompt field I could add: it has your repo, your changelog, the conversation you just had about the release. MCP lets that assistant reach into Deynix instead of you copying text between tabs.

But I also didn't want to hand an agent the keys to my public accounts. So the approval rule is enforced on the server, not in the prompt:

• create_draft_post always lands as pending approval.
• approve_post_group is a separate call, and the server instructions tell the assistant to confirm the copy and time with you first. Clients like Claude also let you require a confirmation for every write tool.
• post_now refuses any post that hasn't been approved.
• Whatever permissions you grant, an assistant can't connect or disconnect social accounts, change billing, manage team members, change autonomy settings, or see other workspaces. The same tools run on Deynix's own API with the same validation and plan limits as the app, so anything the assistant does shows up in the UI straight away.

I'm also deliberately not calling it an "AI marketing employee". An earlier version of my site did, and it overclaimed. What runs on its own is narrow on purpose: it drafts up to a few posts a day into Review, refreshes your reply inbox (it never sends on its own), publishes what you approved, and sends a morning digest.

Connecting it from Claude

It's a remote server using Streamable HTTP and OAuth, so there's no API key to paste and nothing to install. Your client registers itself (RFC 7591 dynamic client registration), you sign in, pick a workspace and choose which permissions to grant.

Claude Code:

claude mcp add --transport http deynix https://api.deynix.com/mcp

Then run /mcp, pick deynix, choose Authenticate and approve access in the browser. Add --scope project if you want it in the repo's .mcp.json for your team.

claude.ai or Claude Desktop: Settings › Connectors › Add custom connector, name it Deynix, paste https://api.deynix.com/mcp, then Connect. Setup for ChatGPT, Cursor and VS Code is in the docs.

A short example

This is the workflow I actually use. From the repo, after tagging a release, I ask Claude Code:

Read the changelog for this release and draft an X thread and a LinkedIn post about it in Deynix. Use my brand kit.

Claude reads CHANGELOG.md, calls get_brand_kit and get_platform_algorithm_hints, then create_x_thread and create_draft_post. It replies with something like:

I've drafted a 4-post X thread and a LinkedIn post. Both are in your Review queue as pending approval. Want me to schedule them for your next best slots (Tue 09:30 and 12:15)?

If I say yes, it calls approve_post_group and the posts go out at those times. If I say "make the second tweet less salesy", it calls update_post and shows me the new copy first. I can also just open the Review tab in the app and approve them there.

Under the hood

For the curious: it's a TypeScript monorepo deployed with SST on AWS. The API, including the MCP endpoint, runs as a single router Lambda that lazy-loads endpoints, video rendering runs on a separate Lambda with ffmpeg, and the docs are an Astro Starlight site. I'm happy to write a deeper post on building a remote MCP server with OAuth on Lambda if people want it.

Honest status

It's early. Around 30 people a month visit the site, a handful have signed up (mostly friends, plus me testing), and only a few people have used the MCP server. I've built a lot and marketed almost none of it, which is a bit ironic for a marketing tool. This post is me starting to fix that.

[Kyle: one honest sentence on the rename from AdVentur.ai to Deynix, if you want to mention it.]

Pricing is $19, $49 or $129 a month, with every feature on every plan; the tiers differ only in brands, seats and credits. There's a 7-day free trial.

What I'd love from you

If you've tried letting an AI assistant help with your own marketing, where did it fall down? The writing, knowing what to post, or trusting it enough to let anything go out at all?

And if you try Deynix, tell me what breaks: deynix.com. I read every comment.