Retool Agents vs. Claude + MCP — limitations, roadmap, and where is the platform heading?

💬 Feature Requests
, , , ,
1

Context and motivation

Hi community,

I'm building an internal agent for my entire company and Retool was the natural choice — great development speed, solid data connectivity, workflow tooling is excellent. But as the project evolves, I'm hitting several limitations that make me question whether Retool Agents is the right platform for company-wide production deployment, or whether it's more of a prototyping tool.

Limitations I've run into

  1. Agent outputs are limited to Markdown — no interactive components, charts, or filterable tables. Retool has beautiful UI components in regular apps, but agents can't use them.

  2. Files — the agent can neither accept file uploads from users nor return files for download. For company-wide use, this is a significant blocker.

  3. Memory and context — the agent doesn't retain tool call outputs between turns, repeatedly calls the same tools, which wastes workflow runs and slows down response times.

  4. Workflow run limits — 5,000 runs/month for the entire organization (Team plan) is very tight for company-wide deployment. Every tool call = 1 run. With 10 users that's 500 runs per person per month — roughly 16 per day.

  5. MCP as output — it would be ideal if Retool Workflows could natively be exposed as an MCP server. Customers use Retool as their data and process hub, but want to call it from external LLM clients (Claude, ChatGPT). Currently the only workaround is a Cloudflare/ngrok wrapper, which adds unnecessary complexity.

Questions for the team and community

a) Is richer agent output on the roadmap — i.e. rendering Retool UI components as the result of an agent turn, not just MD text?

b) Are you planning native file upload and file output support in Retool Agents?

c) How do you handle agent memory in production — do you explicitly save tool results to Retool DB and pass them as context? Is this the intended pattern, or is a native solution being prepared?

d) Are higher or unlimited workflow run limits planned, or a different pricing model for agentic use cases where runs are not the result of direct user actions but internal agent looping?

e) Native MCP server from Retool Workflows — is this on the roadmap? (I can see the org management MCP server beta was launched, but that solves a different use case — I'm asking about exposing custom business logic.)

Alternative I'm considering

I'm in parallel testing an architecture where Claude.ai is the client/agent and Retool logic is exposed via custom MCP servers (Make, Raynet CRM). Results are promising — outputs are significantly richer, memory works, no run limits. But I lose the tooling development speed that Retool provides.

I'd love to understand where Retool is heading so I can decide where to invest my time.

Thank you & looking forward to the discussion

Thanks for taking the time to read this — I realize it's a longer post, but I wanted to give proper context rather than just fire off a list of complaints. Retool has genuinely saved me a lot of development time and I'd love to continue building on it.

I'd really appreciate any response — whether it's from the Retool team with hints about the roadmap, or from community members who've solved some of these challenges in creative ways. Even a "we've heard this, it's on our list" would help me make a better decision about where to invest my energy.

Thanks again!

David

1 Like
3

Hi @Davidkm

Workflows exposed as MCP has been requested a few times, also by me.

For a couple of clients I created this that runs on Cloudflare and gives Claude access to a Retool instance and once you try it, you can’t get back to Retool agents anymore (sorry to say that).

Hope this help.

Best

4

Hi @abusedmedia , thanks for sharing this — it was actually the first thing that came to my mind too, and the repo looks really clean.

The problem I keep running into though is the 5,000 workflow runs/month limit on the Team plan. With this approach every tool call Claude makes still triggers a Retool Workflow run — so wrapping it in an MCP server doesn't change the economics at all. For company-wide deployment with multiple users having agentic conversations, that quota evaporates fast.

Did you run into this with your clients? And if so, how did you solve it — Enterprise plan, self-hosted, or did you route around Workflows entirely?

David

5

My clients are not heavy users, 5k runs/month is enough for them.

They might buy workflow runs, in case they need more, they are very happy of the result, it’s like magic.