Wire AI agents together like parts on a breadboard.
Drop a model into any step, connect it to your data and to other agents, and watch every step run. When a better model ships, swap the chip and keep the board.
Claude, GPT, Gemini, hundreds more through OpenRouter or a model on your own computer, with tools from any MCP server. Preview personal builder plans without a subscription; business access and provider charges are separate.Plans from $7.99 a month.
How one run moves through a board
-
Step 1: the trigger
Something starts the run.
A button, a schedule, a form, a chat message or another board. This one starts from a campaign brief someone pasted in.
-
Step 2: research
An agent reads your sources.
It works from the brief and the reference docs wired into it, puts a source next to every fact and lists what is still unclear.
-
Step 3: writing
A second agent writes, on a different model.
Each chip picks its own model: a cheap, fast one for sorting and a frontier one for the words that matter.
-
Step 4: review
A checker flags anything unsupported.
Claims without a source get marked. An approval step holds the run until you say yes, and every step’s input and output stays open to inspect.
Boards call other boards.
Turn any board into an agent another board can call, or group agents into a project with a manager AI that hands out the work. One research board can feed everything else you build.
Here’s the builder, not a mockup.
This is the sample board straight after a run. Every wire shows how many items went through it, and the run stops at the approval step until a person answers.
The run waits for you.
The approval step shows the drafts next to the checker’s notes. Edit the text, then approve or send it back with a note. Here the checker caught a claim the brief doesn’t support.
Open the sample boardEvery model fits the same socket.
Pick a model for each step: a quick, cheap one to sort and tag, a frontier one for the words that matter. When a new model ships, type its name into any AI part and it runs that day, then put it next to your current one in a Compare models part.
Claude and OpenAI parts
- Claude Fable 5.1Anthropic’s most capable
- Claude Opus 5.5the default for AI parts
- Claude Sonnet 5.5fast all-rounder
- Claude Haiku 4.5quickest and cheapest
- GPT-5OpenAI’s flagship
- GPT-5 minismall and cheap
Custom API part: one part, any provider
- OpenRouterhundreds of models
- Google Geminilong inputs
- Groqvery fast open models
- MistralEuropean models
- DeepSeeklow-cost reasoning
- xAI GrokxAI’s models
- Together AIopen-weight models
- Fireworksopen-weight models
- Perplexityanswers with search
On your own computer
- Ollamalocal models, no key
- LM Studiolocal models, no key
- Next month’s modeltype its name, it runs
Tools your agents can use
- Any MCP serversearch, docs, tickets
- Any APIand web pages
- Your other boardscalled as tools, on Pro
- Knowledge basesyour saved notes, on Pro
- Your key, your bill. AI calls go from your browser straight to the provider, and your keys stay on your device. We add no markup.
- No preset for it? Pick “Set up from the docs” in the Custom API part and an AI reads the provider’s docs and writes the request.
- Kept current. Model list checked 30 September 2026, and a New model alert board can tell you the morning a new one ships. See what changed
One board, whatever you want to automate.
Pick an example to see its board. Each one is a template you can open in the builder and change.
What’s on the bench.
70+ parts for triggers, APIs, data and AI. These are the ones people who build with AI tend to reach for first.
Compare models side by side
Every plan. Real answers with your own key.
Send one prompt to up to eight models at once. Each answer comes back with how long it took and what it cost, and a judge model can score them blind and name the best.
- FACE-OFF
- SPEED
- COST
- JUDGE
Agents that remember
Pro and up
AI agents with chat memory, plus knowledge bases they can save to and search, so a board gets better context every run.
- MEMORY
- KNOWLEDGE BASE
- CHAT
Projects with a manager AI
Pro and up
Pick agents from the Ecosystem and group them into a project. A manager AI hands each step to the right agent and checks the work.
- MANAGER
- 18 AGENTS
- HAND-OFFS
See inside every run
Every plan. Replay on Starter, remix on Pro.
Click any part after a run to see exactly what went in and what came out. Replay a past run on the board, or remix it into a new board.
- INSPECT
- REPLAY
- REMIX
A person in the loop
Approval on Pro and up
An approval step pauses the run until someone says yes, and a safety check reads text before it goes anywhere.
- APPROVAL
- SAFETY CHECK
- NOTES
Plug in the tools you have
Every plan
Give an agent the tools on any MCP server, reach any API, or paste an n8n workflow and it turns into parts you can edit (moving from n8n). A Code part is there when a part won’t do.
- MCP
- N8N IMPORT
- ANY API
Free AI tools, no account needed.
Count tokens, price a prompt on every model, pick a model, and find what to automate first. Or embed our live AI price table on your own site.
AI token and cost counter
No account
Paste a prompt and see its tokens and what it costs on every model, per call and per month.
- TOKENS
- COST
- 7 MODELS
Which model should I use?
No account
Pick the job and what matters most. Get a model to start with, two to compare, and the monthly cost.
- PICKER
- MONTHLY COST
Time-saved calculator
No account
How many hours and how much money an AI workflow gives back each month, after the AI bill.
- HOURS
- ROI
Automation ideas by role
No account
Marketing, support, sales, founders, agencies, developers: what to automate first, ready to open.
- 8 ROLES
- READY BOARDS
Prompt library
Copy, or run as a board
18 tested prompts for writing, support, research, data, marketing and code.
- 18 PROMPTS
- COPY
- RUN
Workflow templates
Open in a click
Every ready-made board, from a model face-off to a weekly campaign package.
- 11 TEMPLATES
- AGENTS
- ALERTS
What works today, and what doesn’t yet.
AI tools are easy to overpromise. Here is exactly where Breadboard stands.
Everything runs in your browser. Your boards, runs and keys are saved on your device, and AI calls go from your browser straight to the provider you picked.
Status October 2026
- Every AI part, on any model
- Custom API to any provider or a local model
- MCP tools in parts and agents, and model face-offs with a judge
- Replay and pinned test data
- On Pro: memory, knowledge bases, projects, remix and approval
- Answers written in advance for a made-up client, so you can try the whole flow before adding a key
- Checkout, so nothing is charged
- AI on Breadboard’s own keys (built, switched off)
- Web search inside AI parts
- A server that runs boards with the tab closed
Who it’s for.
You try every new model
You want to put a new model on a real job the day it ships, and compare it with the one you used last week.
You want to see inside your agents
No black box: every step shows its input and output, and you can replay any run.
You give agents real tools
Hand an agent the tools on any MCP server, your APIs or your other boards, and cap what a run can spend.
You keep your own keys
You’d rather pay the provider directly than buy credits through a middleman.
You run work for clients
Small agencies get a weekly campaign template with research, drafts, a claims check and an approval step, plus the Agency desk below.
Probably not a fit yet
If you need boards running on a server around the clock, or work with sensitive customer data such as health or financial records.
Plans from $7.99 a month.
Plans, starting free.
Cancel any time. Plans renew automatically each month or year until you cancel. Change plan, update your card or cancel from Settings in the builder, and you keep your plan to the end of the period you paid for.
Early access: preview personal plans before checkout opens. Create an account and preview builder features in Settings without buying a subscription. Business workspaces require approved Studio or Agency access, or an invitation. A preview does not enable funded AI; calls on your own keys may incur provider charges.
FreeTry it, build your ownAvailable now
See what it can do, no card needed.
- Run the full sample campaign, start to finish
- Build your own boards in the visual builder, with 70+ parts
- Bring your n8n workflows over with one import
- A taste of AI on us once it launches: about 26 GPT-5 mini messages a month
- Keeps 2 boards and 100 runs a month
StarterYour keys, any model
Run real AI on your own terms.
- Your own Claude or OpenAI key in any AI part, billed by them with no markup
- Any other model, or a local one, through the Custom API part
- Agents that use the tools on any MCP server
- Your AI designs boards from a sentence
- Replay runs and pin test data to build faster
- A Usage page: AI cost by board and by model, and where the month is heading
- The board doctor: finds broken wiring, keys typed into parts and AI with no spending cap, and fixes some in one click
- 2 boards on hourly schedules, while the builder is open
- 10 boards, 2,000 runs a month, 7 days of version history
- Funded AI is currently off; default allowance when enabled: six agent steps per rolling twelve hours
ProFor building agentsRecommended
Where agents get memory, teams and a person in the loop.
- Everything in Starter, plus:
- Agents with memory and knowledge bases they can search
- Up to 2 projects, each with a manager AI that directs its agents
- Ask for approval: runs wait for a person
- Runs that recover on their own, with retries and error workflows
- Test cases that check a board still gives the right answers after you change a prompt or model
- A monthly AI budget that stops runs on your keys, and a usage export
- Remix any past run into a new board
- The weekly campaign package template
- 50 boards, 10,000 runs a month, schedules every 15 minutes while the builder is open, personal workspace
- Funded AI is off; live access defaults to six agent steps per rolling twelve hours
StudioA team running many boards
Run every board from one desk.
- Everything in Pro, plus:
- The Agency desk: every client on one page, with a health check, run caps and a morning digest
- Client reports to print or send: runs, success rate and AI cost for each workflow
- Up to 10 agent projects, each with its own manager AI
- Max effort on every model for your hardest jobs
- One existing business AI allowance shared by all members; seats do not add allowance
- Unlimited boards, 50,000 runs a month, schedules every minute while the builder is open
- Shared agents and owner usage reporting with approved Studio business access
AgencyFor bigger teams
A plan built around how much you’ll use.
- Everything in Studio, with limits sized to your volume
- Unlimited agent projects
- An AI allowance and a price agreed together
- Help setting up your first workflows
- The always-on server when it launches
Compare every plan, feature by feature
Swipe sideways to see every plan.
| Starter$7.99/mo | Pro$19.99/mo | Studio$99.99/mo | AgencyCustom | |
|---|---|---|---|---|
| Using AI | ||||
| AI with your own key, any model | Included | Included | Included | Included |
| Custom API part: any provider or a local model | Included | Included | Included | Included |
| Compare models side by side | Included | Included | Included | Included |
| MCP tools, in parts and agents | Included | Included | Included | Included |
| Your AI designs boards from a sentence | Included | Included | Included | Included |
| Funded AI, currently off | Six-step rolling default | Six-step rolling default | One shared existing allowance | One shared existing allowance |
| Highest effort on our AI | Medium | High | Max | Max |
| Building | ||||
| Boards | 10 | 50 | Unlimited | Unlimited |
| Runs a month | 2,000 | 10,000 | 50,000 | Sized to you |
| Boards on a schedule (they run while the builder is open) | 2, hourly | Unlimited, every 15 minutes | Unlimited, every minute | Unlimited, every minute |
| Version history | 7 days | 90 days | Forever | Forever |
| Replay runs and pinned test data | Included | Included | Included | Included |
| Usage page: AI cost by board and model | Included | Included | Included | Included |
| Test cases for boards | Not included | Included | Included | Included |
| Monthly AI budget and usage export | Not included | Included | Included | Included |
| Retries and error workflows | Not included | Included | Included | Included |
| Remix a past run into a new board | Not included | Included | Included | Included |
| Agents | ||||
| Agents with memory and knowledge bases | Not included | Included | Included | Included |
| Projects with a manager AI | Not included | 2 | 10 | Unlimited |
| Ask for approval | Not included | Included | Included | Included |
| Weekly campaign package template | Not included | Included | Included | Included |
| Team | ||||
| Business workspaces, after owner review | Not included | Not included | Shared agents and usage | Shared agents and usage |
| Agency desk | Not included | Not included | Included | Included |
| Client reports to print or send | Not included | Not included | Included | Included |
| Always-on server, when it launches | Not included | Not included | Not included | Included |
What will your own key cost?
Your AI provider bills you directly for each AI step. Set how much you run and compare models.
Estimates use each provider’s list price and a typical AI step of about 3,000 tokens in and 1,500 out, over 30 days. Claude prices checked 29 September 2026; OpenAI prices are from an older list, so check them before you rely on them. Your real cost depends on your prompts, and each board can have a spending cap that stops a run once it passes.
Funded AI is off. Live funded access defaults to six agent steps per rolling twelve hours; multi-agent work uses several steps. Studio/Agency business membership shares the owner’s existing allowance without creating extra seat allowance.
How the AI allowance works
Once funded AI is switched on, each account gets at most six agent steps in any rolling twelve hours, and the cost limits below apply on top. A message is one typical AI step or chat reply, about 3,000 tokens in and 1,500 out. A board with three AI steps uses three a run, and a message to a project’s manager uses one for the manager plus one for each agent it asks. Each plan’s AI is shared out over 5-hour windows and weeks so it lasts the month, and the builder shows how much you’ve used as a percent. AI through your own key never counts toward it. Allowances may change.
Cancelling and support: cancel any time from Settings in the builder, under Manage billing, and your plan runs to the end of the period you paid for. For anything else, email breadboardstudio@gmail.com and a person will reply.
Cancelling and support: email breadboardstudio@gmail.com and a person will reply. Checkout isn’t open yet, so nothing has been charged and there is nothing to cancel. Before checkout opens, this page will say how quickly we answer and how to cancel.
Running a lot of boards? One desk watches them all.
The Studio and Agency plans include the Agency desk: one page in the builder that reads every board and run, and tells you what needs a person this morning.
Client map
Every board, what it is for, which boards it hands work to and how it last went. The map check flags links to deleted boards, deleted keys, AI steps with no key, missing descriptions and duplicate names, with the fix for each.
- BOARDS
- HAND-OFFS
- MAP CHECK
Routines with caps
Every scheduled board in one table: on or off, next run, last result and runs today. Each routine has a daily cap you set, the workspace has an hourly cap, and a routine whose last three scheduled runs failed pauses itself.
- DAILY CAP
- HOURLY CAP
- AUTO-PAUSE
Morning digest
A plain summary of the last day: runs, failures, approvals waiting on a person, what is due today and what the map check found. It is written from your run history, with no AI and nothing sent.
- FAILURES
- WAITING
- DUE TODAY
Readings, not records
The desk never becomes the place where a fact lives. Each panel says when it read your workspace and marks itself stale when newer runs exist. The boards stay the one source of truth.
- AS OF
- STALE MARKERS
- READ AGAIN
Coming later: today routines run only while the builder is open in a tab. An always-on server that keeps them running and emails the digest is planned. It will draft and never send, so client work still stops for a person.
Optional: help setting up your first real workflow.
Setup help is optional; you can preview the personal builder without it. If your team would rather have help, we can run a four-week assisted pilot on one real deliverable. It is priced after a written scope, provisionally $750 to $1,500.
Week 1: map it together
Pick one deliverable you repeat, like one client’s weekly social package. We walk through how your team does it today and set up the board with you.
Weeks 2 and 3: run it for real
Real briefs go in and drafts come out on your AI key. Your team reviews and approves as usual, and we fix whatever slows them down.
Week 4: measure and decide
We compare staff time and revision rounds with your normal process. If it isn’t clearly better, you stop there.
Assisted setup is the only thing here that involves a person from Breadboard. It is not required, and it is not a sign-up for a plan.
Ask about setup helpQuestions.
Models and AI
Which AI models can I use?
The built-in AI parts suggest Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5, and GPT-5 and GPT-5 mini. The Custom API part reaches anything else: OpenRouter (hundreds of models with one key), Google Gemini, Groq, Mistral, DeepSeek, xAI, Together AI, Fireworks, Perplexity, or a model on your own computer. You use your own key, on Starter and up, and pay the provider directly.
How is this different from a chat app?
A chat answers one message at a time. A board runs the same steps every time, from a button, a schedule or a form, with the model you pick for each step. It can call APIs, read your notes, branch, wait for a person and hand work to other boards, and every run is kept so you can see what went in and out of each step.
What happens when a new model comes out?
Type its name into the Model field of any AI part. The field takes any name your provider accepts, so a new model works the day it ships. To see how it stacks up, the Compare models part sends one prompt to up to eight models at once and shows each answer with how long it took and what it cost. Its “Find the newest models” button asks your provider what it offers, newest first, and the New model alert template checks that list every morning and tells you the day a model appears. We also add it to the suggestions and note it on the What’s new page.
Can it use MCP servers?
Yes. The MCP tool part calls a tool on any MCP server that speaks Streamable HTTP. Paste the server’s address, press “List the tools”, pick one, and its arguments are filled in for you to edit. An AI agent can also get every tool on a server and decide which to call. The server has to accept calls from a browser; ones that only run on your computer over stdio can’t be reached yet.
Can I run a local model?
Yes. In the Custom API part, pick Ollama or LM Studio on this computer. No key is needed. Your local server has to accept requests from this site; for Ollama, add breadboardstudio.com to OLLAMA_ORIGINS.
Can agents browse the web?
Not on their own yet. An agent can use the tools on a search MCP server, an HTTP request part can fetch any page or API and hand it to an agent, and Perplexity through the Custom API part answers with search.
Is the sample real AI?
No. The sample board is an interactive demo for a made-up client, and each AI step shows an answer written in advance. You can edit it and press Run, but nothing calls an AI. Add your own key in Connections to run the same board live.
Do I need to code?
No. Each workflow is a board of parts you can open, read and change, and you can describe a job in a sentence to get a starting board. There is a Code part for anyone who wants it, and you can paste in n8n workflows.
Your data and safety
Where do my boards and keys live?
Personal boards and Connection keys stay in your browser. Studio and Agency business workspaces store shared agent configuration on our server for current members. Personal agents are shared only when you copy them. Business-funded AI calls use the owner’s existing allowance through our server; own-key calls go directly to your provider.
Is it safe to open a board someone shared?
Yes. A board from a link, a file or a paste stays switched off until you choose to trust it: nothing on it runs, not even its previews, and the trust screen lists its expressions, Code parts and the addresses it calls. Shared boards never carry anyone’s keys, and the builder page only runs its own scripts.
Can I take my boards with me?
Yes. Share any board as a link that opens a copy in someone else’s builder (your keys stay behind), download it as a JSON file, or back up every board, run and setting from Settings (keys are left out). Files you download can be imported again, and so can n8n workflows.
What should I not put in it yet?
Please don’t add sensitive customer data such as health, financial or personal records.
Plans and help
What does it cost?
Free is available now: the sample and the builder. Plans are Starter at $7.99, Pro at $19.99 and Studio at $99.99 a month, and Agency is a custom plan agreed with you. Paying yearly costs 10 times the monthly price, so 12 months for the price of 10. Before checkout opens, Settings lets you preview personal builder features without buying a subscription. Business workspaces require approved Studio or Agency access, or an invitation; choosing a preview does not grant that access or enable funded AI. Providers may charge for calls on your own keys. Setup help is optional and priced separately.
Why pay if I bring my own key?
Your plan pays for the builder, not the AI. Your provider bills you directly for AI through your key, with no markup from us. Starter adds the Custom API part for any provider or a local model, replay and pinned test data, and boards on a schedule. Pro adds agents with memory, projects with a manager AI, approval steps, retries and remix. The builder shows which plan a part needs before you use it.
What do paid plans add besides bigger limits?
Tools for running AI you rely on. Every plan has a Usage page with AI cost by board and by model and where the month is heading. Pro adds test cases, which run a board with saved inputs and check the answers still hold after you change a prompt or model, plus a monthly AI budget that stops runs on your keys once it’s used up, and a usage export. Studio adds the Agency desk with its board assistant, and client reports you can print or send.
How much AI does each plan include?
AI on Breadboard’s keys remains switched off. Enabled funded AI defaults to six agent steps per rolling twelve hours, with existing month, week and usage-window funding limits. Every manager, agent and additional provider round uses a step. Business members share one existing owner allowance; inviting more people does not add allowance. Samples and own-key calls are excluded from funded usage.
Do I need an account?
Yes. An account opens the personal builder and sample board. You can preview plan features, including the Studio Agency desk, without a card while checkout isn’t open. Business workspaces require approved access or an invitation.
What is the Agency desk?
A page in the builder, included in the Studio and Agency plans, that reads every board and run: a client map with a check for anything broken, routines with daily and hourly caps that pause after three failures in a row, and a morning digest written without AI. The always-on server that will keep routines running with the tab closed is coming later.
How do I get help or cancel?
Email breadboardstudio@gmail.com and a person will reply. Checkout isn’t open yet, so nothing has been charged and there is nothing to cancel. Before checkout opens, this site will say how quickly we answer and how to cancel.
What is setup help?
An optional four-week assisted pilot. We set up one real workflow with you, run it for four weeks, and compare the time it takes with how you work today. It is priced after a written scope, provisionally $750 to $1,500.
Your first board is one sentence away.
Describe a job in a sentence, start from a template or open the sample board.
Want a hand with something bigger? Leave your email and we’ll reply about setup help.
We only use this to reply about setup help. Privacy
We will email you about setup help.