What’s new

Every change to Breadboard Studio, newest first. When a model ships, it works in any AI part the day it’s out: type its name. We add it to the suggestions and note it here. Follow along with the feed.

Business workspaces, wiring from the keyboard, and moving from n8n

Business workspaces (Studio and Agency)

  • Create a business from your profile menu and invite people with a private link that only works for the email you sent it to. Links expire after a week.
  • Switch between your personal workspace and the business at any time. Agents made in the business are shared with everyone in it; your personal boards stay private until you copy one in.
  • Anyone in the business can add agents. Only the person who shared an agent, or the owner, can change or remove it.
  • The owner gets a usage panel showing who used the business’s AI, and can remove members. Adding members never adds AI allowance or charges; the business shares the owner’s plan. AI on Breadboard’s keys is still switched off, so for now that panel stays at zero.

Also new

  • Wire parts together without a mouse: open a part and use its Wires section to send it to another part, take input from one, pick an If’s true or false pin, or remove a wire. Screen readers also hear what each part sends to.
  • If our server doesn’t answer, the builder opens your personal boards from this device and keeps trying, instead of locking. Business sharing pauses until it’s back.
  • Moving from n8n: what an n8n workflow looks like after you paste it in, and an honest side-by-side.
  • A “Forgot password?” link on the log-in page, and the sign-up form now points at the first empty field.
  • The plans now say that schedules run while the builder is open, and the builder says what AI on our keys will allow once it opens: up to 6 AI steps every 12 hours, within each plan’s monthly limit.
  • Stronger security headers on every page, and readable cookie notices on the legal pages.

Approvals and shared work

  • Closing an approval without a decision now leaves the run Waiting instead of counting it as worked, and nothing goes down the send-back path until someone decides. Runs, the desk and reports show waiting work separately.
  • Reviews queue one at a time, with the decision buttons kept on screen on phones.
  • Continue review in Runs picks a held run back up with its earlier results, so AI steps before the approval aren’t paid for twice.
  • If a teammate changed a shared agent while you were editing, your draft is kept, sharing pauses, and you can download it or load the shared version without losing it.
  • Importing a file with several linked agents keeps them linked to each other and tells you about anything it had to disconnect.

Free AI tools, a prompt library, test cases, a Usage page and client reports

Free, no account needed

Invite a friend

  • Settings has your own invite link, with buttons to post it on X or LinkedIn or email it, and a count of how many people signed up through it.

For paid plans

  • Board doctor (Starter and up): the pulse button on a board scores it out of 100 and lists what will go wrong: parts no wire reaches, two parts with one name, expressions that read from a part that isn’t there, keys typed into a part, pinned data left on, http:// addresses, AI with no key or no spending cap, and AI that sends things out with nobody checking. A spending cap, unpinning and renaming are one click.
  • Time saved (every plan): tell a board how many minutes it saves a person each run (Workflow settings) and what an hour is worth (Settings). Home adds up every run that worked this month, in hours and dollars, next to what your plan costs.
  • Usage (every paid plan): Runs › Usage shows the month’s AI cost by board and by model, tokens, failed runs, a cost-per-day chart and where the month is heading. Each run now records which models it used.
  • Test cases (Pro): save starting data for a board and the checks its result should pass, one per line: contains:, not contains:, field: score >= 70, faster than: 20s, cheaper than: $0.02. Run them all with one click after you change a prompt or model, and see which check failed and why. “Save as a test” turns any past run into one. The board’s toolbar shows how many passed.
  • Monthly AI budget (Pro): set one on the Usage page. You get a warning at 80%, and once it’s used up, boards that use AI on your own keys don’t start until next month. Export any month as CSV.
  • Client reports (Studio): from the Agency desk, pick a client, the workflows and a period, and get a one-page report with runs, success rate, failures, typical run time and AI cost for each workflow, a runs-per-day chart and your notes. It opens as its own page to print, save as PDF or send.

Agency desk

  • A picture of what really happens: each board is a dot on a ring with a cloud of particles round it, denser and faster the more it ran this week, with coral in the share of runs that failed. Particles only travel between boards that really hand work to each other: an agent asking its team for help, a part that calls another board, or an error workflow (dashed coral). The floor glows under the busiest boards, a running board spins, and a short legend under the ring says what is what.
  • Click any dot and the camera swings that board to the front while its parts float above it in run order, each with how it did last time. A panel opens beside it: when it last ran and how long it took, runs, success rate, typical run time and AI cost this week, a 7-day chart, what starts it, which AI it uses, which boards it works with (click one to jump to it) and anything the map check found. The Steps tab lists every part with its last result.
  • A board assistant (Studio): the panel’s Assistant tab changes that board for you. Ask for “an approval step before it sends anything” or “post the result to Slack”, see exactly what it would add, change, remove and rewire, then apply it, or discard it. Every applied change can be undone, and the version before it is saved in the board’s history. It uses your own AI key, or Breadboard’s AI where your plan includes it; it never sees your keys. Without a key it can still explain the board, say why the last run failed, rename it, schedule it, add an approval step or remove a part.
  • Turn the ring by dragging (it keeps a little momentum), pinch or Ctrl+scroll to zoom, and double-click to reset. The toolbar turns, zooms, stops the slow spin, shows every board’s name and saves a picture. It all works without WebGPU too, and stays still with reduced motion.

Fable 5.1, MCP tools, model face-offs, and a lot of polish

Models

  • Claude Fable 5.1, Anthropic’s most capable model, is now in the Claude part’s model suggestions. Claude Opus 5.5 stays the default.
  • Model list checked: Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5; GPT-5 and GPT-5 mini. Anything else runs through the Custom API part.

AI and agents

  • Compare models: a new part sends one prompt to up to eight models at once and gives back each answer with how long it took, marking the fastest. Start from the new Model face-off template. Suggested model names are one click away, “Find the newest models” asks your provider for its list (newest first), and after a real run it names the fastest and cheapest. Add a judge model and it scores every answer from 1 to 10 without knowing which model wrote it, and names the best.
  • New models: a new part asks your provider for its model list and passes on only the models that appeared since the last run. The New model alert template runs it every morning and tells you the day one ships. Compare models takes {{ $json.model }} in its list, so a board can face off each new model against yours.
  • MCP tool: a new part calls a tool on any MCP server, the way AI apps reach search, docs and tickets. “List the tools” shows what the server offers and fills in the arguments for the one you pick. AI agents can use an MCP server’s tools too: switch on “Tool: an MCP server” and the agent picks which to call. The new Agent with MCP tools template is a starting point. The server has to accept calls from a browser.
  • n8n agents import whole: the chat model, memory, calculator, HTTP and MCP tools wired into an n8n agent become that agent’s own settings instead of loose parts.
  • Click an AI agent after a run to see its tool calls in order: which tool, with what, and what came back. The demo agent now does maths written in plain English, like “18% of 240”.
  • Spending cap: set one per board in Workflow settings and a run stops once its estimated cost on your own keys passes it, even in the middle of an agent’s tool rounds.
  • Run details estimate what a run cost on your own key at list price, next to its token count, the Runs list shows each run’s cost, and each Compare models answer shows its cost.

Builder

  • Go to anything: press Ctrl K (⌘K on a Mac) to open any board, page or template, add any part to the board you’re on (type “mcp” or “compare”), run it, or switch themes.
  • Share a board as a link: anyone who opens it gets their own copy. Keys, pinned data and runs stay behind. Find it in Ctrl K or the board’s history menu.
  • Press ? on the board to see every keyboard shortcut, and / to search the parts list.
  • Light or dark now follows your system setting until you pick one. Settings has a new System option, and the dark-mode pop-up is gone.
  • Connections shows one-click starts for a Claude, OpenAI or other key when you have none yet, plus how to use a local model without a key.
  • The sample board wraps onto two rows so the whole run fits a laptop screen, and its note says what to try next instead of repeating the banner.
  • On a phone, the tab bar uses short names (Agents, Desk, Keys) so every tab fits on one line.
  • Home has a “Get started” checklist (run a board, add a key, compare two models, share a board) that ticks itself off as you go, and can be hidden.
  • An empty Runs page has a button to open a board and run it.
  • A part’s input and output show as a table (with a JSON switch), so a list of items is easy to scan, and Copy CSV puts every row in your clipboard for a spreadsheet.
  • Wires that run back to an earlier part loop neatly between the rows instead of cutting across the board, and the digest, price alert and lead scoring templates wrap onto two rows so they fit the screen.
  • Selecting a part keeps it in view when the side panel opens.
  • Agent cards in the Ecosystem show their kind of work by color and icon.
  • The parts list splits AI into “AI with your key” and “AI with no key”, and adding a key links to where each provider hands them out.
  • Opening the builder from a plan on the pricing page switches you to that plan.
  • Parts that need server runs are tagged Soon, with what that means on hover.
  • Fixes: the board name no longer cuts off early, the logo stays on one line, the tour counts its ideas correctly, home stats read “1 workflow” and “1 run”, and the Ecosystem page says its plan limits once, and “Add a key” links read properly for every kind of key.

Safety

  • Boards from a link, a file or a paste stay switched off until you trust them: their field previews and runs wait until you’ve seen which parts use expressions, Code or outside addresses.
  • What a part’s button brings back from an outside server, like an MCP server’s tool list or an API’s docs, can’t run code: anything holding a {{ }} expression waits until you trust it.

Website

  • A new model rack shows every model and provider you can wire in.
  • Each homepage example has an “Open this board” button that starts you on that template in the builder.
  • Real screenshots of the builder, in light and dark, and a recording of a run you can play.
  • A cost calculator compares what a board costs to run on each model with your own key.
  • Pricing says what each plan adds on top of the one below, and every plan is free until checkout opens.
  • A feature-by-feature plan comparison under the prices.
  • Feature cards say which plan each feature needs.
  • A menu on phones, so Models, Plans and the FAQ are one tap away.
  • Every page and every builder screen now passes an automated WCAG AA check in light and dark mode, after fixes to muted text, buttons and links.
  • This page, with an Atom feed.
Build your first board