The Figma MCP server stops reads once the allowance for your plan and seat runs out. This article lists the official caps for Starter, Professional, Organization and Enterprise, compares seat prices, and shows fixes such as checking where the file lives, running whoami, shrinking selections and caching results.
You asked your agent to build a component from a Figma frame, and the run stopped with this: "You've reached the Figma MCP tool call limit on the Starter plan. Upgrade your plan for more tool calls." No reset date. No usage counter. No hint about which of your calls ate the allowance. The limit is real, it is documented, and it is easier to beat than that message suggests. Below you will find the exact caps for every plan and seat, what each step up costs in dollars, and the fixes that clear the error, including several that cost nothing.
The Figma MCP server lets a coding agent read your designs through tools such as get_design_context, get_screenshot and get_variable_defs. Figma offers a remote server and a desktop server. The remote one needs a link to a frame or layer, while the desktop one can work from whatever you have selected. Either way, every read counts against an allowance tied to your plan and seat type. When the allowance runs out, the server refuses further reads and your client prints the limit message. Nothing is broken, and your files are untouched. The meter simply says stop.
The wording above comes from Starter plan reports on Figma's community forum. Figma staff have acknowledged in another thread that the message does not show the reset date, so you are left guessing. A different poster asked whether any page shows MCP usage, and staff moved the question to a private support ticket instead of pointing to a dashboard. Plan on tracking your own count.
Figma's rate limit page says limits apply to tools that read data from Figma. That includes the ones agents lean on most:
get_design_context returns the layout, styles and structure of a layer as code-ready context.get_screenshot captures the selected element as an image.get_metadata returns a sparse XML outline of layer IDs, names, types, positions and sizes.get_variable_defs returns the variables and styles used in a selection.search_design_system, get_code_connect_map and download_assets also read from your files.Three tools are named as exempt: add_code_connect_map, create_new_file and whoami. Those never touch the allowance, so use them freely.
💡 One prompt is not one call. A single "build this card" request can trigger
get_design_context, thenget_screenshot, thenget_variable_defs. Three reads per frame is a normal pattern, and it explains why a small allowance vanishes so fast.
Per-minute limits apply on top of the daily or monthly allowance. Here is the table as published on the official page when this article was written:
| Plan | View or Collab seat | Dev or Full seat |
|---|---|---|
| Starter | Up to 20 per month | Up to 20 per month |
| Professional | Up to 20 per month | Up to 200 per day, 10 per minute |
| Organization | Up to 20 per month | Up to 200 per day, 15 per minute |
| Enterprise | Up to 20 per month | Up to 600 per day, 20 per minute |
| Education | Not listed | Up to 200 per day, 10 per minute |
Figma adds one more line: it "reserves the right to change rate limits." Treat every number here as a snapshot and check the page before you plan around it.
Three different counters can stop you, and each one feels different:
Do the arithmetic at three reads per frame. A Starter allowance of 20 calls is about six frames a month. A Professional Dev seat's 200 calls per day is around 66 frames a day, but the cap of 10 per minute means roughly three frames a minute at best. Agents that fire several reads at once hit the per-minute wall first, which is why a session can fail at 9:05 and work again at 9:06.
If you searched for this error before, you probably saw "6 tool calls per month" for Starter. Figma staff did quote six in older forum replies, and several third-party write-ups still repeat it. The official page now lists 20, and a more recent staff reply on the forum also says up to 20 per month. If your own account hits the wall well before 20, the cause is usually the file location problem described in the fixes below, not a hidden lower cap.
Figma does not bill per MCP call. You pay for seats, and the seat decides the allowance. Figma staff have also said the MCP allowance is separate from Figma AI credits, so a full credit balance does not help when the tool call limit is hit.
Prices below are per seat per month on annual billing, taken from a 2026 price tracker. Confirm them on Figma's pricing page before you buy, since Figma adjusts them.
| Plan | Full seat | Dev seat | Collab seat |
|---|---|---|---|
| Professional | $16 | $12 | $3 |
| Organization | $55 | $25 | $5 |
| Enterprise | $90 | $35 | $5 |
The cheapest route to a daily allowance is a Professional Dev seat at about $12 a month. It cannot edit designs, but it reads them, which is exactly what an MCP workflow needs. Compare that with Starter's 20 calls a month.
A worked example makes the gap concrete. Say you need to implement 30 frames at three reads each, so 90 reads in total. On Starter, that is more than four months of allowance. On a Professional Dev seat, it fits inside half of one day's allowance, and the per-minute cap of 10 means the whole batch can finish in about nine minutes of steady calling.
Figma's own advice is a ladder:
Read the table again and a pattern shows. Professional to Organization leaves the daily cap at 200 and only lifts the per-minute cap from 10 to 15. The jump to 600 per day arrives only at Enterprise. If your pain is slow bursts, Organization helps a little. If your pain is running out by lunch, you need Enterprise or fewer wasted calls.
The tool call limit is not the only bill. Every get_design_context response lands in the model's context window, and a large frame produces a large response. Figma's docs warn that big selections slow the tools down, cause errors and can return cut-off answers. Pick the model to match the job: a fast model such as GPT 5.6 Luna for quick passes, and a coding model such as Claude Sonnet 5 when you generate the final component.
Start with the symptom, because each pattern points to a different cause:
| What you see | Likely cause | Fix |
|---|---|---|
| Error on the first call of the month | File sits in Drafts or a Starter team | Move the file to the paid team |
| Error after roughly 20 calls on Starter | Monthly cap | Upgrade, or wait for the billing cycle reset |
| Error mid burst, gone a minute later | Per-minute cap | Serialize the calls |
| Error in the afternoon on a Dev seat | Daily cap | Cache results, shrink requests |
whoami shows a View or Collab seat | Wrong seat | Ask an admin for a Dev or Full seat |
This is the cause behind most "but I pay for a Dev seat" threads. Limits follow the team that owns the file, not your personal seat. A Figma staff member explained that files in Drafts, or in a team still on Starter, get Starter limits even when whoami reports a Professional Dev seat. The fix from staff: confirm the file sits in the paid team. If it sits in Drafts or a Starter team, save a local copy and re-import it into the Professional team, then point your client at the new copy.
whoami is exempt and returns your identity, plans and seat types. Run it before anything else when the limits look wrong. If it shows a View or Collab seat, you found the problem in one free call. If it shows the seat you pay for, go back to the file location check. One catch: whoami is a remote server tool, so it is not available on the desktop server.
Figma's advice is to break screens into smaller parts, like components or logical chunks. A card, a header or a sidebar per request gives faster, more predictable output than a full screen. It also protects your allowance, because a request that fails or truncates on a huge frame spends reads and gives you nothing back. Copy the link for the exact piece you want instead of the whole page.
On Professional and above, one get_metadata call to see the layer outline can save retries, because it lets you pick child layers by ID. On Starter, skip it and paste the exact frame link, since every read is precious.
The per-minute caps are 10, 15 or 20 depending on your plan, which works out to one read every six seconds on Professional. Agents that run several tools at once blow past that. When one forum user complained about being blocked after very few calls, Figma staff replied that no exceptions are available and advised slowing the requests down. Put a rule in your agent instructions: call one Figma tool at a time, and never retry a failed Figma call more than once in a row.
Save the output you already paid for. Write the get_design_context result for a finished frame into a folder in your repo, and keep one copy of the variables from get_variable_defs instead of asking for them on every frame. Next session, the agent reads the file and skips the call. Cached notes also keep working while you are capped, which matters more than it sounds on a Starter plan.
The Figma interface still works when MCP does not. Export the frame as a PNG yourself, copy the spacing, colors and type values from Dev Mode, and hand both to a vision-capable model on PicassoIA. Gemini 3.5 Flash reads images and writes code quickly. Qwen3.7-Plus interprets images and generates text. Kimi K2.5 reads text and images in one chat. For the code itself, Claude Sonnet 5 or GPT 5.6 Sol can handle the heavier coding tasks.
You lose the structured data the MCP returns, so describe tricky pieces in words. A rough first pass still keeps the work moving.
While you are at it, map components once. add_code_connect_map is exempt, so linking Figma components to your real code components costs nothing against the allowance, and it tends to cut re-runs later because the output points at components you already have.
On Starter, a Figma staff reply says the limit resets monthly, aligned with your billing cycle. For paid seats, the official page lists daily and per-minute caps but does not spell out when the daily counter resets. If you need an exact date for your account, open a support ticket, since that is where staff send such requests.
Figma is not special. Every service behind an MCP server has a meter, and the fixes rhyme. PicassoIA's API and MCP connection, for example, use a concurrency cap instead of a call quota: 5 concurrent predictions per account, shared across every token and MCP connection on that account. Jobs are asynchronous, so you create a prediction and then poll for the result. The models behind it are PicassoIA Image, PicassoIA Image Editor Pro, PicassoIA Video and Seedance 2.5 Lite.
| Meter | What it limits | Best response |
|---|---|---|
| Monthly cap | Total reads per month | Cache results, shrink requests |
| Daily cap | Reads per day | Plan work in batches |
| Per-minute cap | Speed of calls | Run one tool at a time |
| Concurrency cap | Parallel jobs | Queue the rest and poll |
Whatever the meter, spend calls on work only a call can do, cache the results, and stop the agent from firing blindly.
A rate limit pushes you to plan before you click, and that habit pays off in visual work too. If you are waiting for a reset anyway, use the time to create the pictures your design needs. Open PicassoIA Image, type a prompt for a hero photo, a product shot or a background, then refine it with PicassoIA Image Editor Pro. When a still is not enough, animate the idea with PicassoIA Video, which makes clips from text or an image.
Start with one prompt, change the lighting and framing, and compare versions side by side. The full model list lives at picassoia.com/en/all-models. Try it on your next mockup and see how much you can produce before your Figma allowance even resets.
