Really appreciated the great recap from the roundtable!
It’s fantastic to see Coda laying out a robust vision for where AI is heading—especially the focus on iterative building, schema support, and breaking AI out of the typical “sidebar” jail.
Having spent the last several months working deeply with the private beta of the Coda MCP server via Antigravity and Claude Code, I wanted to jump in and offer a bit of a “reality check.” The roundtable summary significantly understates what is already possible today when you combine Coda MCP with a competent agentic framework. And by “competent”, I mean “additive contexts and rules”.
The notes mention that the first three MCP tools coming are things like pushing buttons, adding images, and commenting. But in the current beta environment, my agents have access to over 30 distinct tools made available in Coda MCP. I’m already orchestrating full CRUD operations across Documents, Pages, Tables, Columns, Rows, and even dynamically building and executing CFL Formulas.
Through an integration framework I’ve been developing called mcpOS, I’ve learned that I do not have to wait for Coda AI to figure out schemas or relational logic [eventually]. By combining the Coda MCP server with rigorous, deterministic workflows, orchestrating these complex structural builds has been possible for months now. In fact, mcpOS has directly anticipated and remedied several of the challenges mentioned in the roundtable including, but not limited to these:
- Schema Hallucinations and Data Pollution: The digest noted that AI sometimes struggles with Coda-specific patterns and hard-codes values. mcpOS remedies this via Schema Defense (
/coda-schema). Before writing any data, the agent dynamically audits the target table’s columns and formats, ensuring it never attempts to write to a non-existent column or insert mismatched data types. - Context Window Limits and Blindspots: Answering questions about “the rest of the docs” or workspace-level context is a known challenge. mcpOS remedies this using a Shadow Context Protocol (
/coda-shadow), which actively maps document structure (pages, canvases, formulas) to gain deterministic context, bypassing the limitations and blind spots of standard RAG approaches that are likely used inside Coda and almost certainly the cause of these issues. - Complex Logic and Duplication: The native AI is still working toward understanding complex relational data. mcpOS handles this today with Smart Upserts (
/coda-upsert). It programmatically handles the logic needed to safely add or update records without creating duplicates, acting much more like a traditional database integration than a simple chatbot.
As several of my Community articles show, a deep dependence on Coda’s internal AI features comes with some downsides. I’m more of an “agentic pureist” who believes that agentic platforms (which Coda has chosen to be) should implement safety and defenses for everyday use cases. But we cannot—and should not—expect the AI tooling to do everything because with everything that is possible comes the expense of agility.
Coda’s philosophy of eliminating the “work behind the work” is absolutely spot-on, and validates my assertions in several MCP-related articles. As predicted here, here, here, and here—Makers will soon be operating with a new definition of making Coda solutions.
It’s important that the community knows that, by leveraging Coda’s MCP through rigorous workflows that anyone can create in Markdown files with tools like Antigravity, Cursor, Windsurf, or custom agentic setups, you have the license to functionally operate months ahead of the native roadmap laid out in this digest. Here’s another prediction:
Because of the agentic abilities to utilize MCP with agent-native rules and visual reasoning of Coda itself, which can be designed and implemented seamlessly, Makers will forever have slightly more control than Coda developers.
Exciting times ahead!