Figma launches Specs beta for shared AI coding context

Figma announced Specs on October 15, 2026, with an open beta rolling out to paid plans. The feature brings written product intent into the same workspace as visual exploration and technical detail. It is a context-sharing release, not evidence that coding agents can ship correct software without review. citeturn2view1

What Specs adds to the handoff

Teams can import a Markdown file and connect the document with frames, diagrams and other canvas references. Figma says a coding agent connected through its Model Context Protocol server can receive a Spec link and read the associated context. That makes the written rationale available alongside the visual material, rather than relying only on an image of the intended interface. citeturn2view0

The help documentation also describes a figma-spec skill for agents to read, create and update standalone or embedded Specs. This supports maintaining the plan during implementation, not just sending a document once. citeturn3view0

Availability is gradual, and agent work has costs

Figma's guide lists Professional, Education, Organization and Enterprise eligibility and requires edit access. It says rollout will take place over the next few weeks, so an announcement does not mean every account has access immediately. Writing and editing a Spec manually does not consume AI credits. citeturn3view0

Using the Figma agent to create or edit Specs does consume credits. Consumption varies with task complexity and the context processed; Figma's published examples are estimates, not fixed per-document prices. A pilot should measure actual usage rather than assume the beta makes agent work free. citeturn2view3

MCP access is a separate consideration. Supported clients must appear in Figma's catalog, read limits depend on plan and seat, and users can access only content they already have permission to view or edit. citeturn3view1

Context access is not execution reliability

Figma's developer documentation explicitly distinguishes the server from the coding assistant. MCP supplies structured design input and a code starting point; the assistant produces the final code. The server does not automatically understand a team's conventions or fix implementation problems. citeturn3view2

Our interpretation: Specs is most interesting as a way to make decisions inspectable by both people and agents. Whether that improves delivery remains a workflow question to test, not a guaranteed outcome of the launch.

A useful first evaluation

Choose one small interface change. Record the intended behavior, edge cases and acceptance criteria, then connect the relevant design references. Ask the agent to explain those requirements before implementing them.

Compare the result with your existing handoff: count missed requirements, review corrections and credit consumption. Keep a human approval step before merging. Expand only if the shared context produces a measurable improvement, not merely a more convincing demonstration.