Coda MCP is now available in public beta!

[Ruggy-Joesten]

Your AI tools can now connect directly to your Coda Docs

Hi everyone,

Today, we’re excited to share that Coda MCP (Model Context Protocol) is now available in public beta.

This introduces a new way to work with Coda alongside AI tools like Claude and ChatGPT, allowing them to read from and write to your docs using plain-language prompts.

Instead of copying information between tools or making manual updates, you can ask your AI assistant to work directly inside your docs and tables.

What is Coda MCP?

MCP is a connection layer that enables AI assistants to interact with Coda in the same way a person would.

One way to think about it: it gives your AI assistant a proper seat at the table, so it can understand your docs and help you take action with the right context.

Another analogy we’ve been using internally: it’s like a USB stick you plug into an AI assistant—it tells the assistant how to use Coda without needing a custom integration.

Connect Coda to your AI tools

You can now connect Coda directly to tools like:

• ChatGPT

• Claude

• Cursor

• Codex

through a simple OAuth connection. That means:

No API keys. No complicated setup. Just “Connect with Coda” and you’re ready to go.

Your AI assistant will only see and interact with the documents you already have access to, and it will act on behalf of the account connected to it.

What you can do with MCP

Here are a few workflows people are already using it for:

1. Turn rough notes into structured docs
   Share a Coda doc link and ask your assistant to turn messy notes into a structured brief, proposal, or PRD — written directly into the page.

2. Build and populate tables
   Describe the table you want — columns, types, example rows — and MCP can create it in your doc and fill in the first entries.

3. Analyze tables and pull out insights
   Paste a doc link and ask something like: “What are the top 3 themes in this table? Add them as a summary at the top of the page.” Your assistant reads the data and updates the doc with the results.

4. Turn meeting transcripts into structured updates
   Paste a Zoom transcript, ask your assistant to extract decisions, action items, and owners, then log them in a Coda meeting tracker.

5. Generate status reports from project trackers
   Ask your assistant to summarize project trackers, flag risks, and draft a weekly update across multiple workstreams.

What this unlocks

• Your AI assistant can now work with the actual data in your Coda Docs, rather than guessing based on whatever context you’ve pasted in.

• Less copying. Less switching between tools. More getting things done inside your docs.

You can connect Coda MCP to your AI tools and start experimenting today.

Getting started with Coda MCP

If you give it a try, share what you build here in the community: workflows, experiments, examples, whatever you’ve got. Seeing how people use new capabilities like this is always the best part.

And if you have ANY feedback at all on things acting up, please let us know by filling out this doc.

[Gielton_Pereira]

Interesting, does it also work with Gemini?

[Nick_Sims1]

Ruggy,

This is great. I’m using Claude Code with Coda through the API right now which works well. Are there any limitations with the MCP or advantages of one method or the other?

Nick

| Ruggy-Joesten Community Manager at Superhuman
April 1 |

• | - |

Your AI tools can now connect directly to your Coda Docs

Hi everyone,

Today, we’re excited to share that Coda MCP (Model Context Protocol) is now available in public beta.

This introduces a new way to work with Coda alongside AI tools like Claude and ChatGPT, allowing them to read from and write to your docs using plain-language prompts.

Instead of copying information between tools or making manual updates, you can ask your AI assistant to work directly inside your docs and tables.

What is Coda MCP?

MCP is a connection layer that enables AI assistants to interact with Coda in the same way a person would.

One way to think about it: it gives your AI assistant a proper seat at the table, so it can understand your docs and help you take action with the right context.

Another analogy we’ve been using internally: it’s like a USB stick you plug into an AI assistant—it tells the assistant how to use Coda without needing a custom integration.

Connect Coda to your AI tools

You can now connect Coda directly to tools like:

• ChatGPT

• Claude

• Cursor

• Codex

through a simple OAuth connection. That means:

No API keys. No complicated setup. Just “Connect with Coda” and you’re ready to go.

Your AI assistant will only see and interact with the documents you already have access to, and it will act on behalf of the account connected to it.

What you can do with MCP

Here are a few workflows people are already using it for:

1. Turn rough notes into structured docs
   Share a Coda doc link and ask your assistant to turn messy notes into a structured brief, proposal, or PRD — written directly into the page.

2. Build and populate tables
   Describe the table you want — columns, types, example rows — and MCP can create it in your doc and fill in the first entries.

3. Analyze tables and pull out insights
   Paste a doc link and ask something like: “What are the top 3 themes in this table? Add them as a summary at the top of the page.” Your assistant reads the data and updates the doc with the results.

4. Turn meeting transcripts into structured updates
   Paste a Zoom transcript, ask your assistant to extract decisions, action items, and owners, then log them in a Coda meeting tracker.

5. Generate status reports from project trackers
   Ask your assistant to summarize project trackers, flag risks, and draft a weekly update across multiple workstreams.

What this unlocks

• Your AI assistant can now work with the actual data in your Coda Docs, rather than guessing based on whatever context you’ve pasted in.

• Less copying. Less switching between tools. More getting things done inside your docs.

You can connect Coda MCP to your AI tools and start experimenting today.

Getting started with Coda MCP

If you give it a try, share what you build here in the community: workflows, experiments, examples, whatever you’ve got. Seeing how people use new capabilities like this is always the best part.

[Rick_Wolnitzek]

Very cool. I hope connecting to Perplexity is on the horizon.

RW

[Nikola_Lajic]

Is there a way to limit which documents the MCP has access to and what kind of access (read/write)?

I’m not comfortable with giving blank access to all of the docs in my org I have access to to any AI. There are some docs that are sensitive, and I want them hidden from the MCP, and to some I want to make sure there are no accidental changes (readonly).

[Ruggy-Joesten]

Great question!

If you’re in the Google ecosystem, there are a couple of ways to connect Coda’s MCP, though Gemini’s web portal doesn’t currently support custom connectors.

The two options available right now are:

• Google Antigravity
• Gemini CLI

Let us know if either of those avenues work for you.

[Bharat_Batra1]

Hey @Rick_Wolnitzek - I haven’t tested it yet, but it looks like Perplexity now supports Custom Connectors (see docs). The API key option seems promising! You can generate an MCP token from your coda account page using this direct link.

Let us know if it works!

[Bharat_Batra1]

Hey @Nick_Sims1 , it’s awesome to hear you’re already using the Coda API with Claude.

That being said, there are significant advantages to using MCP over the API for chat-based use cases.

1. Additional Capabilities: as of today the MCP offers expanded capabilities like creating tables and reading/writing comments. You can find the full list of capabilities here
2. Agent ergonomics: the MCP doesn’t use the API internally, and was instead built from the ground up to be used by agents. This means it is more markdown friendly, context conscious (still working on this), and streamlined in terms of input and output parameters (each tool relies on a consistent URI pattern) to reduce errors.
3. Built-in guidance: The MCP tool descriptions and responses both allow us to teach the agent how it should use Coda. With the API, agents would need to parse through the internet to find the right information for formulas, column types, etc. But inside the MCP this information is built in - improving both accuracy and speed.

However, there are use cases where the API may be more suitable. For deterministic scripts where you don’t need agentic reasoning at runtime, the API tends to be a safer approach as its endpoints are stable, versioned, and not subject to sudden change. The industry as a whole is still figuring what versioning and stability mean for MCPs, and we’re also still looking into this.

Let me know if you have further questions or thoughts!

[Bharat_Batra1]

Hey Nikola! We plan to add granular permissions to the MCP soon. As of today, you can provide read/write/read+write access using Personal Authentication tokens. Soon, we plan to introduce scoping by specific workspaces, folders, and more. Stay tuned, and let us know if a particular level of scoping will help you out most

[Marinda_Carelsen]

Interesting. I got another message from coda today that suggested I connect my cursor via the api. I have already done it. I wish I read this one first so that I could’ve used OAuth. Any difference between the two options?

[Bharat_Batra1]

MCP functionality-wise, there is no difference between using OAuth or the personal access token.

Once connected, they both function just the same.

However, there are some differences in terms of security. For instance, 3 legged OAuth is a safer protocol because it means the token exchange is handled behind the scenes and tokens cannot easily be shared. On the flip side, there are cases where you actually need the token to be copyable so you can insert it into a script or piece of code. So the personal authentication tokens are intended for those cases.

[Rick_Wolnitzek]

Thanks for responding, Bharat. I don’t think Perplexity is ready yet.

Rick Wolnitzek
Architekwiki

Architekwiki | Practical Resources for Architects

[Theresa_Hunt]

Hello,

Could you help me understand what the risk is here if we have sensitive data within our CODA docs?

[Daniel_Stieber]

Have you fixed the issue with doc locking and filters or does this mean you just completely opened up every doc every user has access to?

[Ruggy-Joesten]

@Theresa_Hunt @Daniel_Stieber will try to answer both your questions below:

1. For the sensitive data risk question @Theresa_Hunt: the best approach today is to use Coda’s permissions model alongside Cross-doc and Synced Pages to control what data is accessible to anyone, including agents acting on your behalf. More granular agent-specific scopes are on the roadmap, which will give you finer control over what an agent can access during setup, but your permissions structure is the foundational layer regardless.

2. For the doc locking and filters question @Daniel_Stieber: MCP access is bounded by existing user permissions, so nothing is exposed beyond what a user already has access to. That said, we hear you that user-level access is a coarser control than many of you want for agent workflows specifically. More granular scopes during agent setup are coming, which will let you limit what an agent can reach even within a user’s permitted docs. In the meantime, Cross-doc and Synced Pages can help you architect around exposure if you want to get proactive about it.

Hope this helps a bit!

Also, would love each of you to drop your feedback here for the rest of the team to see. Any and all early insights as y’all start tinkering with the tool will help us make improvements as fast as possible.

[Daniel_Stieber]

Thanks for your answer. I remember the improved doc search was one of the most discussed features and as far as I remember, the community had a clear wish: being able to limit what users can see.

Coda MCP is now doc search on steroids.

Both miss out on essential filtering. E.g. a (non-maker) user can in the UI not show hidden pages, but he can access all information via the MCP now, recreate it in another doc.

Cross-doc and shared pages miss out on many coda basic features and makes doc building incredibly complicated and even unusable in many scenarios. What many of us want since years is a simple reliable filtering, locking & hiding within a doc. Now we are further away than ever…

Have feedbacked this in the closed beta already.

[Nick_Short1]

I’m trying to better understand how permissions behave with the new Coda MCP integration, especially after seeing that the current connection appears to grant both read and write access by default (without a way to toggle them separately).

In our workspace, we have a mix of access levels:

• Some users are doc makers/editors
• Others have view-only access to specific documents
• Some documents are intentionally restricted to smaller groups

I want to clarify a few things:

1. If a user connects their account to MCP, does it strictly inherit their existing Coda permissions (i.e., they can only access and modify docs they already have access to)?
2. If someone has view-only access to a document in Coda, does MCP still enforce read-only behavior, even though the connection itself requests write access?
3. Does enabling MCP introduce any new capabilities beyond the normal UI—for example, making it easier to query or modify content across multiple docs at once?
4. As a workspace admin or doc owner, is there currently any way to:
   • Restrict or disable MCP access?
   • Limit MCP to specific documents or folders?
   • Audit or monitor MCP-driven activity?
5. Finally, just to confirm: if users already have access to certain documents in the workspace, does connecting MCP simply make that access programmatic, or does it expand what they can reach in any way?

I want to make sure that enabling MCP doesn’t unintentionally broaden access or allow unintended write operations, especially for users who are supposed to be view-only.

Would appreciate any clarification from the team or others who’ve tested this in real scenarios.

[Daniele_B]

Hi guys,

In Cursor 3.0 (Linux) I can’t get to authenticate the MCP.

It keeps saying “needs authentication” → I click Connect → the browser opens → I approve the connection → I’m sent back to Cursor

But the authentication is not working.

What can I do?

Thanks!

[Marcio_Assencio_Araujo]

First, congratulations to the Coda team — the MCP itself is excellent. The quality is high, the experience is impressive, and what it enables inside docs feels genuinely powerful.

I had a very positive experience using it with Claude: it worked out of the box. I connected it, shared a doc URL, and it was immediately usable.

My issue is specifically with the ChatGPT experience. In practice, I found that the connection only becomes usable there when Developer Mode is enabled, which makes the workflow much less friendly for normal day-to-day use.

So my question is: are you already in conversations with OpenAI about improving this part of the ChatGPT experience? And is there any expected timeline for making the connection more seamless and user-friendly there?

To be clear, this does not feel like a Coda MCP quality issue — the product itself is very good. It feels more like a platform integration limitation on the ChatGPT side, but it would be great to know whether there is a path or timeline to improve that experience.

[Chris_DeAngelis]

Mine keeps saying it cannot access Coda even though MCP App set up and connected in ChatGPT and I am in developer mode. Do I need to add some permission on Coda side?