Your agent connects to Coldbluff and asks on your behalf, signed in as you. What it starts runs on your account and spends from your window and your week, on whichever plan your account holds. It reads from the same pages you do, judged on the same criteria, and it gets the same answers.
Fig. 1 · Open your agent with the ask typed
Paste this into your agent, or open Claude or ChatGPT with it already in the box. If the agent has no Coldbluff connector yet, it sends you back to the steps below.
Use the Coldbluff connector (the MCP server at https://coldbluff.ai/api/mcp) to get a read of the idea I am about to give you. Start a read, answer what it asks from what I have told you, confirm the story, then show me who has the problem in their words, who already sells them something for it, and what it costs to reach them. Every line comes with its source; open the sources instead of trusting the summary. If you have no Coldbluff connector available, the connection steps are at https://coldbluff.ai/connect (as text at https://coldbluff.ai/connect.md): tell me, and stop there.
Fig. 2 · The address
One address for every client. The server speaks Streamable HTTP, the sign-in goes through your Coldbluff account, and there is no key to paste.
https://coldbluff.ai/api/mcp
Fig. 3 · One action per client
Each row is the one thing to do in that client, checked by hand on 5 October 2026. The vendors move these menus, so if a row no longer matches, tell us.
Claude
In Claude on the web or in Claude Desktop, open Customize, then Connectors, then Add, then Add custom connector. Paste the address, add it, and press Connect: that opens the sign-in. Free holds one connector; Pro and Max hold more. On a Team or Enterprise plan an owner adds it once under Organization settings, Connectors, Add, Custom, and each member presses Connect on it. No link does this for you yet.
Claude Code
One line in the terminal.
claude mcp add --transport http coldbluff https://coldbluff.ai/api/mcp
ChatGPT and dots
Open ChatGPT Plugins, press the plus, then Create custom MCP server. Name it, paste the address under Connection as a public endpoint, read the warning and press I understand and want to continue, then Create as a plugin. Install it, start a conversation, type @ in the prompt box and pick it. A dot, the always-on agent OpenAI shipped on 29 September 2026, uses the same plugin. Account and workspace policies decide who may create one. The first call opens the sign-in.
Cursor
VS Code
Add to VS Code, or add to VS Code Insiders.
Any other client
In a client that reads a JSON file, the entry looks like this.
{ "mcpServers": { "coldbluff": { "url": "https://coldbluff.ai/api/mcp" } } }
Fig. 4 · The sign-in
The first call opens a sign-in page in your browser. Sign in with your Coldbluff account, and the client keeps the token it gets back. The server uses OAuth, so your password never passes through the agent, and the agent acts as you: its reads are your reads, on your plan, against your limits. Every client you have connected is listed in your settings, where one control disconnects it. The client asks you to sign in again the next time it calls.
Fig. 5 · What it can ask
The tools, in the order your agent uses them. Each line is the one the server gives the agent, read from the server's own list as this page is built, so the table never names a tool the server does not have.
| Tool | What the server says it does |
|---|---|
| get_account_status | The account this agent acts as: email, plan, the rolling five-hour window and the week with what is spent and when each frees up. Read-only, free, safe to repeat; call it first, or whenever a tool says the window is full. |
| start_read | Start a read from the idea in the founder's words, or from the address of their site. Writes the story back (the six links: problem, customer, what they use today, the product, the price, and the market: where these people are and the language they write in), filled where the text implies it and asked where it does not, and waits up to 45 seconds for it. One cheap model call from the account's window; the read then shows in the founder's Coldbluff sidebar. Needs no prior call. |
| answer | Answer one open question of the story: the field and the founder's answer in their words ("I do not know" is an answer; a price needs a number with its currency and period). Returns the story as it now reads and the questions still open. Free, no model call, except that the last answer rewrites the story with everything known (one cheap call, waited for). Refused once the story is confirmed: a change is tweak_idea. |
| confirm_story | The founder read the story back and said it reads right. Confirms it (it is never edited afterwards; a change is tweak_idea) and starts page two: who has the problem in their words, and who already sells to them. Runs on every plan, spends from the window, takes minutes: call read with waitSeconds afterwards, never confirm_story again. Needs the story written and its questions answered. |
| run | Start pieces of a confirmed read: page two (people, competitors) on every plan, page three (buyer, disconfirming, distribution, evals, customers, board, roadmap) on Pro. With no kinds it starts every piece the plan allows that has not run. Spends from the window: a batch runs what fits and reports the rest as skipped. Returns at once; a read takes minutes, so call read with waitSeconds afterwards and never run again while a run is open. |
| read | The read so far: the story with its slots and open questions, every piece that has run (status, version, output with provenance on every claim), the sources the pieces cite (url, verbatim, venue, author, date; a quote carries its evidence id and its labels, the text is in sources), the open run with each of its pieces, its state and the last line it reported, the usage window and the pieces not yet run (shape "read"). kinds narrows the pieces returned. waitSeconds (up to 45) waits for the open run to finish, so poll with it rather than in a loop; when the run is still open at the end of the wait, the answer is shape "progress": the run with the seconds elapsed and the last line of each piece, the pieces in kinds that finished in this run with their sources, and next, without the story, the catalogue or the window, which only change when the run closes. Read-only, free, safe to repeat. |
| export | The whole read as one markdown document, provenance on every line and the sources numbered at the end, with a filename. On Pro. Read-only, no model call, safe to repeat; call it when read shows the pieces you want in it. |
| tweak_idea | Change a confirmed idea in one sentence ("sell it to agencies instead", "same thing at $9 a month"). Forks a new read next to the old one, keeps what the change does not touch at no cost, and returns the new ideaId with what runs again; the new story starts at once. One model call from the window. Call read on the new id with waitSeconds 45, then answer and confirm_story as for a new read. |
| list_reads | The reads on this account, newest first: id, title, when it started, how many of its pieces are done and whether a run is open. Read-only, free, safe to repeat. Call read with an id for one of them. |
| docs | One Coldbluff docs page as markdown, by its slug (for example "how-a-read-runs"). Read-only, free, safe to repeat; call it when the founder asks how a read works or what a tool does. The list of slugs is at https://coldbluff.ai/api/docs.json. |
| send_feedback | Tell the people who build Coldbluff what went wrong or what is missing: a tool that could not do what the founder asked, a refusal that did not say what to do next, a read that read wrong, or something the founder wants that no tool does. kind is bug, missing_capability, friction or idea; name the tool in toolName when there is one and the read in ideaId when it is about a read. It spends nothing and changes nothing, and nobody replies here. Ten a day. |
Every tool also takes a context, one sentence on why your agent is calling. Nobody reads it but us, to see what agents try to do with the server. It never changes an answer.
What comes back is the same read a person gets on the site, as data: the story with its open questions, the quotes with their URLs and dates, the printed prices, and the state of every line, found, your assumption, or nothing found. Every answer ends with the one call to make next, and a refusal says why and what to do instead.
The change your agent makes with tweak_idea works the way Iterating describes. It starts a new read next to the old one, keeps what the sentence doesn't touch, and runs page one of the new read straight away.
Every docs page is open to your agent. The list of slugs, with each page's title, description and links, is at the address below. Each page also has a markdown copy at the same path with .md at the end.
https://coldbluff.ai/api/docs.json
Fig. 6 · The same limits, and the same no
When a limit is full, the tool answers with when it resets, the way the site does. How your agent's use counts against your plan is on the pricing page, and the window is explained in Free and the limits. Your agent's reads are judged on the criteria How we judge shows, and nothing about the connection changes an answer.