A Claude second brain for your to-dos and your Claude Code plans

Gnothi's MCP server gives Claude, Cowork and Claude Code one board to keep with you: what you have to do, what each to-do is worth, the plan a coding session is on, and every change Claude made, waiting in an inbox for you.

The connector and API keys are free. You need a Claude account; a personal one, since on a Team or Enterprise account everyone in the organization would share your key.

Everything Claude can do with your Gnothi board

Your life

  • Check in when you have 20 minutes, and Claude offers the first thing that fits
  • Tidy your to-dos: Claude estimates what each is worth and how long it takes, and Gnothi's formula ranks them
  • In Cowork, with Claude in Chrome, take on a to-do like ordering a gift, and ask before it pays
  • Log a run or any habit from a chat, the same as a tap, gold included
  • Say a to-do to your watch with Do it, and Claude sees it at your next check-in
  • A morning brief and a nightly triage as scheduled tasks
  • Read and write your journal entries, on a key you grant that

Your code

  • A plan becomes a tree any Claude Code session can read and keep
  • Titles first, about 30 tokens an item, and details only as deep as needed
  • Parallel sessions claim the steps they're on with #wip
  • Every edit to a step you tagged #spec waits on you in Needs review
  • Decisions worth keeping move to Reference when a feature ships
  • Add a to-do mid-session without leaving the terminal

Your say

  • Every edit Claude makes is logged with its reason and shows as a diff
  • Approve, reject (it rolls back) or comment, one at a time or all edits at once
  • Every delete waits for your approval
  • Claude's questions wait on their items, for you to answer from your phone
  • Changed or deleted items can be restored for about 30 days
  • Each key reaches only the fields you pick, and journal access starts off

Claude as your second brain, without building an Obsidian vault

The rules for keeping a to-do list come with the connector, so there's no skill or prompt file to maintain.

Make an API key, add Gnothi as a custom connector on claude.ai and paste one prompt. Claude asks a few questions about what you're juggling, shows you the whole board as a dry run, and builds it on your yes. From then on, the same board works in the Claude app on your phone, in Cowork and in Claude Code.
To change a rule, write it in a field's notes, such as "anything with a phone call goes under Calls" or "never show me work items on weekends", and Claude follows it from then on. Tell Claude a preference in the conversation and it adds the line to that field's notes itself, then tells you in one line what changed.
Capture starts at your wrist. On Apple Watch or Wear OS, with AI on, tap Do it and say "dentist appointment 3pm Saturday", and it goes into the right field, due Saturday, where Claude sees it at your next check-in. Tap Journal instead and say a thought: the day's thoughts become one journal entry, which Claude reads only on a key you've given journal access. More on the watch
Between conversations, the board talks back. Tap #together in the app to see what you and Claude do together, #you for what's yours alone, and #urgent for what can't wait. Add a line under any item to reply, retag a #together item #claude to hand it back with Claude's recommendation accepted, or add #pinned to keep an item where you put it.

Claude with Obsidian, Notion or Gnothi

A lot of people build a Claude second brain on Obsidian, which stores notes on your device as plain-text Markdown files, and Anthropic's docs say Claude Code can run inside a notes vault the same way it runs in code. Notion's MCP server lets Claude search a workspace, create and edit pages, and manage databases.
Gnothi starts from the to-do list. Its connector ships the conventions: how to judge a to-do, where finished work goes, what Claude may change on its own and what waits for you. Code ranks the list, and every change Claude makes lands in an inbox. Each item's note is Markdown, Claude reads and writes the tree as indented lines, the way it would edit a Markdown outline, and any branch exports as Markdown or OPML on the web.
Obsidian keeps everything on your device, offers thousands of plugins and themes, and is free without limits; Notion is a whole workspace of pages and databases. Gnothi is hosted: your board lives on Gnothi's servers and syncs to the web, iPhone, Android and your watch.

Connect Claude over MCP: one key, one connector, one prompt

  1. In Gnothi, open Settings › Security › API keys and make a key: all your fields or a few, read only or able to change them. Journal access starts off.
  2. On claude.ai, add Gnothi as a custom connector with that key. In Claude Code, run the one-line command in the Claude Code section below.
  3. Paste the setup prompt. Claude builds your board as a dry run and applies it on your yes.
  4. For check-ins by voice or while driving, set every Gnothi tool to Always allow in the connector's settings, so no call waits for a tap.
Claude still checks changes with you before applying them, and a mistake can be restored from a backup for about 30 days.

A personal AI assistant for your to-dos

Say "check in with me, I've got 20 minutes." Claude opens with its read of your situation, such as 20 minutes at your desk on a Tuesday evening, plus where your dailies stand today, and offers the first thing that fits the time and the place. A to-do that needs a phone call doesn't come up at 10 at night, and one that needs business hours doesn't come up on a Saturday.
Questions Claude can't settle wait on the to-dos they're about, in Needs review. Answer them in the app when you have a minute, say "pong", and Claude picks up from your answers. From your phone, say "log that I ran" and it counts the same as a tap in the app, gold included.

Claude Cowork and Claude in Chrome can do some of your to-dos

In Cowork, with Claude in Chrome, Claude can take a to-do like Order a birthday gift for Mom, find options and fill a cart, then ask before it pays. Whatever a field's notes say, the connector's rules have Claude wait for your yes before it spends money, messages anyone, changes a security setting, does anything that can't be undone, or runs a command an item names. Those are instructions Claude follows, not a lock in Gnothi. A yes you write elsewhere on the board to spending money, messaging someone or anything that can't be undone waits until you confirm it with Claude in a conversation; approving Claude's question in Needs review is itself your yes, exactly as asked.
Cowork is on paid Claude plans, and it needs the Claude desktop app open for browser use.

Claude sorts your to-dos, or the Triage button does

Say "tidy my to-dos" and Claude works through the board on your Claude plan. It judges each to-do, merges duplicates, folds notes into the items they belong to, checks off what's done, and asks to delete what has nothing left to keep. The top three stay in front and the rest file under Next, in folders such as Calls and Paperwork.
Never connected Claude? The Triage button in the app runs the same rules on Gnothi's own AI. It shows the cost in credits and names every field it will read before it starts. A Triage chip on each field turns it in or out, and the changes it makes go into the same review inbox, under the name Gnothi Triage.

Payoff: how Gnothi ranks your to-dos

The AI estimates a few facts about each to-do. Code does the ranking, the same way every time.

When Claude or the Triage button judges a to-do, it writes a few short keys on it. Gnothi computes everything else when it reads the board, and stores no score.
InputWho sets itWhat it means
Value, 1 to 5Claude or Triage estimatesWhat doing it is worth: its stakes times the chance it pays off. 5 is harm now, 3 is about $100 a month or a person waiting, 1 is trivial; each step is about three times the one below. Unset counts as 2.
MinutesClaude or Triage estimatesHow long it takes: 5, 15, 30, 60, 120 or 240. Unset counts as 15.
NeedsClaude or Triage estimatesWhat it takes beyond a phone: a computer, a call, going out, or business hours.
Not beforeClaude or Triage setsThe first day it can be done, such as while you wait on a reply. Until then it's hidden, with everything under it.
Target dayClaude or Triage setsAn aim for a day that isn't a deadline.
Due date, repeat, #pinnedYouA due date you set stays until you say otherwise.
Urgency, step dates, the first list, the score, the orderGnothi's codeComputed from the above, every time the board is read.
  1. Left out first: anything checked off, anything with a not-before day still ahead, and anything whose needs the moment rules out: business hours outside Monday to Friday 9 to 5, a call or an errand outside 8am to 9pm. In a check-in with a time limit, a to-do longer than the time you have drops out of the ranked part.
  2. The first list: a to-do that's overdue or due by tomorrow, or tagged #pinned or #urgent, goes ahead of every score, earliest date first. A repeating chore worth under 3 never jumps the line this way. A project counts as due when any of its open steps is.
  3. Everything else gets a Payoff score:
Payoff = 10 × 3^(value − 1) × urgency × target lift × field turn ÷ √(max(minutes, 5) ÷ 15)
FactorRangeHow it moves
3^(value − 1)1 to 81Each step of value triples the score.
urgency1 to 21 when the due date is two weeks or more away, rising each day as it nears.
target lift1 to 1.51 two weeks before the target day, 1.25 on the day, 1.5 two weeks after. It never puts a to-do on the first list.
field turn1 to 9Each day, the two fields you've let slide most get a turn: their next few to-dos are multiplied by up to 9. Only when Claude asks what's next.
√(minutes ÷ 15)0.58 to 4A 15-minute to-do divides by 1, an hour by 2, four hours by 4.
To-do (no due date or target, no field turn)ValueMinutesPayoff
Unjudged(2)(15)30
Call the landlord about the deposit31590
The same call, but it takes an hour36045
The 15-minute call, due in 7 days315135
Dispute a charge before the window closes415270
Within one list, your own order holds; across lists, whichever list's next to-do scores highest goes first. When Claude reads the changes and questions waiting on you, they come in the same order, first list then score. Zoom into a judged to-do and its Payoff shows beside its chips, or First list when that list ranks it; hover or long-press to see what Payoff means.

Every change Claude makes, in one inbox you approve

Let Claude work on its own, then review what it did.

Every edit Claude or Triage makes is logged with who made it and why. The Review agent changes button on your board carries a count of what's waiting on you, and it filters the board to three sets: Needs review, what waits on you; Unseen, logged changes you haven't looked at; and All.
Open an item and its changes read as one conversation. A renamed item shows a word diff, a changed note a line diff, and a move shows the old path and the new one. When an edit moved one of a to-do's Payoff inputs, such as its minutes or its due date, the change shows its Payoff before and after, such as Payoff 45 → 90. Approve it, reject it and it rolls back, or comment and Claude reads your reply next time. Rejecting an edit that later edits built on rolls those back too, after asking.

Deletes wait for you

Every delete Claude sends is a request that leaves the item in place until you approve it, even one you asked for in the conversation. The only exceptions are a mirror, or an item Claude added in the same write.

#spec marks what you decided

Claude can't create a #spec item, and adding or removing the tag waits on you. Every edit Claude makes to a #spec item, a move or check-off included, waits on you in Needs review, and a reject rolls it back.

Approve all edits at once

When the edits look right, approve every one waiting with one confirm. Anything that also holds a delete request or a question stays for you to open on its own.

Underneath it all, nightly backups keep about 30 days, so an item changed or deleted by mistake can be restored. A deleted field can't.

Claude Code planning and memory in a task tree

A plan inside a Claude Code session lasts as long as the session. In Gnothi, it's a tree any session can read and keep.

One command adds Gnothi's MCP server to Claude Code, on an API key that reaches only the fields you pick.
claude mcp add --transport http gnothi https://api.gnothiai.com/mcp --header "Authorization: Bearer gnk_..." --scope user
Ask Claude to plan a feature, say CSV export, as a tree: the feature on top, its steps under it, each step's details in its note. As steps get done, Claude checks them off and records the progress in the parent's note, so the tree shows what's done and what's left. In any session, tell Claude "add a to-do to review Sam's pull request" and it's on your list without leaving the terminal.

CLAUDE.md, plan mode and a task tree: what each one keeps

Each Claude Code session begins with a fresh context window. CLAUDE.md carries the instructions you write into every session, and auto memory carries notes Claude writes itself. Plan mode has Claude read files and propose a plan, making no edits until you approve.
A Gnothi tree keeps the plan's state: which steps exist, which are done, which session is on which, what you answered, and what was decided. Your CLAUDE.md can point at it in one line, and a plan you approved in plan mode can become the tree.
CLAUDE.md lives in your repository, versioned with your code and loaded with no network call; plan mode is built into Claude Code.

Context engineering: titles first, details on demand

Claude reads the plan's titles first, about 30 tokens an item, and opens a step only as far as it needs. With AI on, a note over 40 words gets a brief and one over 500 also gets a summary, so Claude can read the gist before the whole note. When the context compacts, or you come back tomorrow, the plan is still there: Claude writes where each step stands into its note, and the next session picks up from it.

Parallel sessions and subagents on one plan

Run two Claude Code sessions at once and each claims the step it starts with a #wip tag, so the other picks a different step. Gnothi stamps each claim with the session that holds it and an expiry. While the claim is live, only that session can release it; once it expires, a claim left by a session that died can be taken. Sessions that share one key send a short agent label so their claims can be told apart.
Subagents inherit their session's MCP tools, so a subagent can read the tree and report back on a step without filling the main session's context.

Spec-driven development with #spec, alone or with Spec Kit

Tag a step #spec to mark what you decided. Claude can't create a #spec item, adding or removing the tag waits on you, and every edit Claude makes to the step, a move or check-off included, waits on you in Needs review, where a reject rolls it back. When Claude needs your call on a step, it asks on the item; answer in the app, on your phone if you like, and the next session reads your answer. A bug Claude finds off to the side can become its own item, so the step in hand stays on track.
GitHub's Spec Kit is a toolkit for spec-driven development, run as commands in your agent's chat such as /speckit-specify, /speckit-plan and /speckit-tasks. Use it to write the spec, and put the tasks it produces in a Gnothi tree, where sessions claim them and you approve changes to the parts tagged #spec. Spec Kit is open source and keeps its files with your project.

Decisions that outlive the feature

When a feature ships, what's worth keeping, like why you chose one library over another, moves to your Reference field, and later steps point back to it. Off the code side, Reference holds the same kind of thing for your life: contacts, standing preferences, and why a choice went the way it did.

A morning brief and a nightly triage

Claude can run two scheduled tasks on this connector, in Cowork, or on the web or phone with a Pro or Max plan. The morning brief reads your day each morning and changes nothing. The nightly triage sorts your to-dos overnight, so the brief and your next check-in open on a sorted board. It deletes nothing itself, only requests deletes for you to approve, and it acts on nothing outside Gnothi. Anything it's unsure of, it leaves as a question on the item.

What Claude can reach

Each connection runs on its own API key: all your fields or a few, read only or able to change them. Journal access starts off on every new key; turn it on and Claude can read, search and write entries under tags you pick. Use a personal Claude account, since on a Team or Enterprise account everyone in the organization would share your key.

When a Markdown file is enough

If a plan lasts one session and you want it in git beside the code, a Markdown file in the repo does the job: it's versioned, it's free, and Claude reads it with no network call. A Gnothi tree lives on Gnothi's servers, outside your git history, and needs an account. It earns its keep when the plan runs across many sessions, when parallel sessions work it at once, when you want to approve changes from your phone, or when the decisions should outlast the feature.

Common questions

Does Gnothi work with Claude Code and Claude Cowork?

Yes. One connector works in claude.ai, the Claude app, Cowork and Claude Code; Claude Code connects with one claude mcp add command.

Does Claude decide the order of my to-dos?

No. Claude estimates each to-do's value, minutes and needs; Gnothi's Payoff formula ranks them from those estimates and your due dates, the same way every time.

Can Claude delete my things?

Not without you. Every delete Claude sends waits in Needs review until you approve it; the only exceptions are a mirror, or an item it added in the same write. An item deleted or changed by mistake can be restored for about 30 days, but a deleted field can't.

What does it cost?

The connector and API keys are free, and Claude's own work runs on your Claude plan. The in-app Triage button and AI briefs use Gnothi credits, and every account starts with a balance.

Can I get my data out?

Any branch exports as Markdown or OPML on the web, and Settings › Your data › Export data downloads a full JSON export.

Is it an Obsidian alternative?

For to-dos and plans, it can be. Obsidian stores notes as Markdown files on your device and has thousands of plugins; Gnothi is a hosted board with to-do rules, ranking and a change inbox built for Claude. Many people will keep both.

Connect Claude to one list

The connector and API keys are free. Make a key under Settings › Security › API keys, add the connector, and let Claude build your board with you.

Get started free