Coding agents
One prompt for an agent with a terminal, one address for Claude.ai or ChatGPT over MCP. No token to make either way.
Send this prompt to Claude Code
Fetch https://sumibako.com/skill.md in full and follow it. If your web tool summarizes pages, use curl.It reads the skill and uses it for this session. Nothing is installed.
No account yet? You make one when you approve the first request. It is free, and nothing here is behind the paywall.
Use Claude or ChatGPT in a browser?
Each has its own guide, naming every click.
“Write up the migration plan and give me a link I can send the team.”
You
Step 1
In your own words. The skill tells the agent when a link is what you want.
Connect this machine
WXYZ-4821
Step 2
The first run opens a page in your browser, or prints a link to it: check the code matches, connect. A connector asks you to sign in instead.
sumibako.com/p/Migration-plan-k2f8
Step 3
The page is filed under Agent notes in your vault, and the agent hands you the address to send on.
Second run, same page, same link. A plan revised five times is one page, not five.
The same thing done by hand, for a specific client or for one that speaks MCP instead of running commands. Two steps either way. A terminal needs nothing pasted and reads the file off disk, so a 30KB plan costs nothing to send; Claude.ai and ChatGPT connect over MCP, sign you in and ask what to allow.
Which agent are you connecting?
It teaches every agent on this computer what Sumibako is, in every project. It copies one Markdown file and runs none of our code.
$ npx skills add sumibako/agents -g -yAsk for a plan or a note to be saved to Sumibako, or point it at a Markdown file in your project. The first time, your browser opens a page to connect this computer: check the code matches, approve, and the agent carries on and gives you the link. From then on there is nothing to approve, and a revised file updates the same page, so a plan revised five times is one page and one link.
$ npx sumibako publish docs/plan.md --publicGives Claude Code the tools directly instead of a command. Add it once, then run /mcp in Claude Code, pick sumibako, choose Authenticate and approve in the browser; there is no token. The command line is still the better path here, because it reads the file off disk rather than sending it through the model.
$ claude mcp add --transport http --scope user sumibako https://sumibako.com/api/mcpThe skill is what teaches an agent the rules here, including the part about not creating duplicates. It is one Markdown file, and you can read it before your agent does at sumibako.com/skill.md.
Six tools, the same over MCP and the command line. Read and write are always granted. Publishing is the one you choose when you make a token; a connector you sign in to may publish only pages it wrote.
write_pageWriteFiles Markdown as a page, creating it or updating one it has seen before.
read_pageReadReads a page back as Markdown, which is the form to edit and send back.
edit_pageWriteAdds to a page, or replaces one exact piece of text, without re-sending the rest.
search_pagesReadFinds pages by their words, or lists the newest when asked for nothing.
publish_pagePublishPuts a page on the web and returns the link, behind a passcode when asked, or takes it back down.
get_accountReadPlan, limits, usage, and what this connection is allowed to do.
Delete anything
There is no delete in the API at all, so a loop cannot destroy your writing. Unpublishing takes a link off the web and leaves the page where it is.
Put your notes in front of strangers
Published pages carry noindex, and turning that off is a decision you make yourself, per page, in the app. An agent can produce a link to send to someone; it cannot make one findable by a search engine. And a page it put behind a passcode stays behind it every time it publishes again - only you, or an agent you explicitly asked, can take the lock off.
Reach anyone else's vault
A token is scoped to the account that made it and nothing else.
Publish, if you said no
Turn the permission off when you make the token and it can still read, write and edit, privately.
The full surface, for when you want it. None of it is needed to finish the setup above.
A page you no longer have on disk is still editable: open --markdown prints it as Markdown, and that is the same text --find matches against.
The same capabilities over plain HTTP, for a script or a CI job that would rather not shell out. Send the token as a bearer header. Errors come back as JSON with a stable code, so a client can tell a plan limit from a bad request without reading the sentence.
curl -X POST https://sumibako.com/api/agent/v1/pages \
-H "Authorization: Bearer $SUMIBAKO_TOKEN" \
-H "Content-Type: application/json" \
-d '{"externalId":"docs/plan.md","markdown":"# Plan\n\nFirst draft.","publish":true}'A hosted connector signs you in rather than taking a token. Paste the address, press Connect, and you are shown what the connector may do before anything is granted.
What it may do is narrower than what a token can do. It reads and writes pages, and it can put a page it wrote on the web so it can hand you a link. It cannot delete anything, and it cannot publish a page you wrote yourself - that stays yours, whatever you approve. A token you create can do all three, because you tick the boxes.
Each connection gets its own row under Settings, Coding agents. Revoking it there stops it on its next request. Removing the connector at the other end is a separate action, and worth doing too - revoking here makes it stop working, not disappear from the app you added it to.
Put it behind a passcode. Ask your agent for a passcode, or run publish with --passcode, and you get the link and a passcode of 6 characters to send with it. The link alone opens a lock screen that shows nothing of the page, not even its title, and a browser that has typed the code stays unlocked for 30 days. You can change the passcode from the Publish menu in the app, which locks out everyone holding the old one.
Anything that can run a command or speak MCP. Claude Code, Codex and Cursor have a terminal, so the CLI is the better path there - the file goes straight from disk instead of through the model. Claude.ai and ChatGPT have no terminal, so they connect over MCP and sign in rather than holding a token. A CI job can use either the CLI or the HTTP API directly.
No, as long as it uses the default behaviour. Every page is filed under a stable key, which the CLI sets to the file's path in the repository, so filing the same document again updates the same page. That is the single most important behaviour here: an agent revising a plan five times in a session produces one page and one link, not five.
Yes. It reads the page back as Markdown, changes the part that is wrong, and writes it back - or appends a line without touching the rest. A replacement has to match exactly once; if the text it is looking for appears twice, the request is refused rather than guessed at, because guessing rewrites the wrong paragraph and nobody finds out.
Yes, from an agent with a terminal. A picture, recording or PDF that the Markdown embeds is uploaded along with it and shown on the page, and sumibako attach adds files to a page that already exists. They count against your plan's storage exactly as an upload in the app does, and a file already in your vault is not uploaded a second time, so publishing a revised plan does not store its screenshots again. An embed only reaches files inside the repository the Markdown is in; anything else has to be named with attach. An agent connected over MCP cannot send files, because a tool call carries text.
The previous version is kept for 14 days, and the last 20 revisions of a page are available under Version history in the document menu. Restoring snapshots the current state first, so a restore is itself undoable.
It copies one Markdown file into the skills directory of every agent it finds installed - Claude Code, Cursor, Codex, Copilot and around forty others. -g puts it at user level, so it works in every project, and -y answers the installer's questions for you. It is the ecosystem's own installer, not ours, and it runs no code from us at any point. The file teaches your agent what Sumibako is and the rules for using it well, so you never have to name the tool in a request. You can read it before installing: it is one file in a public repository.
Yes. Paste the prompt at the top of this page and your agent fetches the skill from sumibako.com/skill.md and follows it for that session, with nothing copied into its configuration. The difference from npx skills add is only permanence: add copies one file into every agent on the machine and is done, the prompt leaves nothing behind and has to be pasted again next time.
Almost never. The first time an agent runs a command on your computer, your browser opens a page to connect it, you approve, and the agent carries on - nothing typed, nothing pasted, which is also why the agent can start the setup itself instead of waiting for you. An agent with no terminal signs in the same way: paste the address into its connector form and approve what it asks for. What still wants a real token is a caller with nobody at a screen, which means CI and scripts, where it goes in an environment variable.
No. A token is not a session. It can only reach the pages endpoints, it carries only the permissions you gave it, it cannot delete, and it cannot change your account or billing. Revoke it in Settings and it stops working on its next request.
Both take an agent's Markdown perfectly well. Neither can hand it back as a link. Notion has no API or tool that turns on Share to web, so somebody opens Notion and clicks through Share and Publish for every page; Obsidian has no hosted page at all unless you buy Publish for a whole site. Here publishing is one flag on the same command, and the address it prints keeps working when the agent revises the file.
Under a page called Agent notes, made the first time an agent writes, with a folder inside it for each repository, newest pages first. That keeps a plan-per-session out of the middle of your own writing, and the same file name in two repositories stays two pages. Both are ordinary pages once they exist, so rename them, move them, or drag things out of them; new arrivals still find them, because they are tracked by id rather than by name.
No, and the free plan is not a trial of this. It holds 1,000 pages and 1,000 published ones, which is more agent-written pages than a person files in a year. Pro removes both ceilings and raises storage, and that is the whole difference.