footnote
Install

Tell your AI agents why the number moved before they guess.

A person asks, in the browser

How did orders in Norway change in October 2026?

Checked This is written down and checked.

A new customs rule stopped all online sales to Norway from 4 August 2026, so orders there fell to almost nothing.1

How were orders counted in the United Kingdom in October 2026?

Conflict Two items say different things.

One says they count store credit.1One says they are card and cash payments only.2

Why did returns in Sweden jump in March 2023?

Not found Nothing is written about this yet.

Nothing is written down for Returns in Sweden in March 2023.That's not the same as there being no reason.

Checked by the owner of Sales

Online sales to Norway stopped on 4 August 2026

"Parcels from webshops outside Norway are held at the border from 4 August 2026 until further notice."

Sales, owned by Jon Abara and Sofia Brandt.
Don't conclude: fewer orders in Norway after 4 August do not mean fewer customers.

Sales. Written by Jon Abara

United Kingdom orders count store credit

≠

Sales. Written by Sofia Brandt

United Kingdom orders are card and cash payments only

No footnote

Nobody has written this down yet.

An agent asks, over MCP

How did orders in Norway change in October 2026?

This may reflect weaker marketing performance.

called get_context

"status": "matched"

"say": "Checked. 5 context items apply to Orders in Norway in October 2026. Nothing else was looked up. 1 of them applies to only part of that question."

1"title": "Online sales to Norway stopped on 4 August 2026"

"not_to_conclude": "Fewer orders in Norway after 4 August do not mean fewer customers."

How were orders counted in the United Kingdom in October 2026?

called get_context

"status": "conflict"

"say": "Conflict. Two items say different things. 10 context items apply to Orders in United Kingdom in October 2026. Nothing else was looked up. 2 of them apply to only part of that question. Some of them disagree. 1 more context item matches but isn't shared with AI agents: it's high risk and waits for a check."

1"title": "United Kingdom orders count store credit"

2"title": "United Kingdom orders are card and cash payments only"

Why did returns in Sweden jump in March 2023?

called get_context

"status": "no_match"

"say": "Not found. Nothing is written about this yet. Nothing is written down for Returns in Sweden in March 2023. Nothing else was looked up."

"context": []

Open source under Apache-2.0. Not released yet: it runs on one computer with Docker.

Before either of them reads a word of it.

Draft

Anyone can write a context item. It's short: what happened, where and from when, and how you know.

An agent can suggest one too, and it waits for the owner like anything else. A draft never reaches an agent.

Checked

It lands in the Inbox of whoever owns that area. They see whether the claims match the source, and what agents would say before and after. Then they publish it, or send it back with a note.

From now on get_context returns it with every answer it applies to, with its source.

Due for a check

When the check date passes, or a definition it relies on changes in Cube or dbt, the owner looks again: still true, or edit.

Until then the answer starts with Outdated, so the agent can say it may be old.

Not found, again

If questions about one topic keep finding nothing, the owner gets a card about it. It's how people find out what to write next.

We count the topic, not the words.

In the docs: Write context Review a suggestion

It sits next to your semantic layer, not inside it.

Cube or dbt MetricFlow still defines Orders and still returns the figure. Footnote only reads the names, so that a note can say which metric and which market it's about. After the owner publishes the note, it goes out with every answer that uses it, as a footnote.

The figure

Cube or dbt MetricFlow defines Orders and returns it. Footnote reads the names only.

The footnote

Online sales to Norway stopped on 4 August. Someone in Sales wrote it, and the owner checked it.

Orders in Norway fell to almost nothing from 4 August.1
The answer

Claude Code over MCP, or Ask in the browser.

No model picks the context. Plain rules match the question's metric, market and dates against what's written. Footnote doesn't store numbers either. Your semantic layer returns the figure, and the item points at it.

A context item on the demo: Online sales to Norway stopped on 4 August 2026, with the sentence copied from the customs notice, its domain, owners, check date and source link.A context item on the demo: Online sales to Norway stopped on 4 August 2026, with the sentence copied from the customs notice, its domain, owners, check date and source link.

A context item is short

What happened, where and from when, the sentence copied from its source, who owns it and when they last checked. That's all an agent reads.

Observability on the demo: what agents asked this week, by topic and answer line, with the share of checked answers and the questions that found nothing. Counts only.Observability on the demo: what agents asked this week, by topic and answer line, with the share of checked answers and the questions that found nothing. Counts only.

Gaps come back as counts

Observability shows what agents asked about, by topic and answer line, and which questions found nothing. It keeps counts, not the words.

In the docs: How agents use it

Every answer starts by saying what it found.

Ask is the page for people. Agents get the same first line over MCP. There are six of these lines, and code picks which one, not the model, so Checked means the same thing every time.

  • CheckedSomething applies, and its owner published it with a source and a check date.Nothing to do.
  • PartialPart of the question is covered, or the dates were assumed. The answer says which.Write the rest, or tell the owner.
  • Not foundNobody has written this down. That's not the same as there being no reason.Write it, or tell the owner.
  • ConflictTwo items apply and they disagree. Agents get both and say so.An owner keeps one, narrows one, or keeps both.
  • OutdatedThe check date has passed, or a definition the item relies on changed.The owner looks again: Still true, or Edit.
  • Not looked upAsk couldn't look just now. It doesn't mean nothing is written.Try again in a minute.

In the docs: Answer lines

Agents only see what an owner has published.

Every item belongs to one area, like Sales or Paid media, and the owners of that area decide what goes out. Drafts, suggestions and comments stay inside.

Drafts stay drafts

The database role that answers agents can't read the versions table at all. That's a permission, not a setting someone can flip by accident.

A review before it goes live

The review holds the claims up against the source. It also replays the questions agents asked against the draft, so the owner sees what will change before it does.

Nothing filled in

When nothing is written, the answer says Not found and stops. It doesn't reach for a likely story.

What not to conclude

An item can say what not to read into it. The Norway one says fewer orders there doesn't mean fewer customers, and the agent gets that line with the item.

When a definition changes

If a definition an item relies on changes or disappears in Cube or dbt, the item is flagged for its owner, and answers say Outdated until they've looked.

Topics, not words

When an agent asks something, we keep what it was about and how often. We don't keep the words. Your own chats in Ask stay in your browser.

One thing to know: in this version, everyone who is signed in, and every agent, can read everything that's published. Per-person limits aren't built yet.

In the docs: Domains and owners Privacy

Six MCP tools, and no model on our side.

Your agent connects over MCP and signs in as you. Footnote doesn't run a model for agents, so the same question gets the same items until an owner publishes a change.

  • get_contextThe items that apply to a question about a metric, a market or some dates, with the answer line first.
  • find_contextLook up what's written about a few words, using the names people actually say, like GMV.
  • read_contextOne item in full, with its source and its owner.
  • get_relatedFollow the links from an item, a metric or a domain. A link is not proof of a cause.
  • reportasks firstTell the owner an item looks wrong or unclear, after the person says yes.
  • suggest_contextasks firstSend the owner a suggestion. Nothing goes live until they've checked it.

Connecting takes one command, then one approval in the browser. An admin can switch any tool off for any agent. We tell agents that item text is data, not instructions, and in our test runs they ignored the instructions we planted in items.

In the docs: Connect an agent Agent tools

One line in your AGENTS.md

When you connect an agent, the prompt the app gives you adds one line to the project's instructions file: CLAUDE.md if you have one, otherwise AGENTS.md. That's all you write.

AGENTS.mdthe line you add

# Footnote
Before you answer a question about company data, call get_context first and use what it returns.

get_context replysent every time

  1. State the documented fact with its source and dates.
  2. Say "might be related" only when the dates overlap and the metric is in scope, and say the context does not prove a cause.
  3. Never say "caused by" unless the item states the effect.
  4. When an item has not_to_conclude, never conclude what it rules out.
  5. Suggest a check that would test the effect.

How to explain the why isn't left to that file. Footnote sends the rules with every reply, so each agent gets the same ones, and they change with the server rather than in every repository.

What it reads, and what reads it.

Some of this works today. Some of it is built and still waiting for someone to try it on a real account. Each row says which.

Definitions

Metric and dimension names from your semantic layer. Read only, and it works today.

  • Cubeworks today
  • dbt MetricFlowworks today
  • A list you write yourselfworks today

Documents

Notes that arrive as suggestions for an owner. Built, not tried on a real account yet.

  • Markdown in Gitbuilt, not tried on a real account
  • Google Drivebuilt, not tried on a real account
  • Notiondesigned, not built

Agents

Anything that speaks MCP, signed in as the person. We've tried Claude Code.

  • Claude Codetried
  • Cursorinstall link, not tried
  • VS Codeinstall link, not tried
  • ChatGPT and Codexwritten steps, not tried
  • Any other MCP clientover HTTP, not tried

People

Ask, in the browser, for anyone who is signed in.

works todaybuilt or linked, not trieddesigned, not built

In the docs: Connectors Build a connector

Try it on your own computer.

You need Docker. The first start builds Footnote from the code you cloned and loads the demo company: its people and their notes, a few suggestions waiting for review, and some weeks of questions from agents. It can take ten minutes.

  1. Start it

    Clone the repository and start it. It builds on your computer.

    $ git clone https://github.com/Fredehagelund92/footnote.git && cd footnote
    $ docker compose up

  2. Open it

    Sign in as one of the demo people if it asks, open Ask, and try the Norway question.

    $ open http://localhost:8080

  3. Connect Claude Code

    Add the server, run /mcp, pick it and choose Authenticate. Your browser opens. Pick a demo person and press Allow.

    $ claude mcp add --transport http footnote http://localhost:3000/mcp

Runs on

One computer with Docker. Everything it keeps is in PostgreSQL.

  • Docker
  • PostgreSQL

A model, if you want one

For Ask's sentences and the answer tests. It's off unless you start Footnote with MODELS=live and a model key, and agents never need it. A real model costs money and receives the context Ask reads.

  • Anthropicthe default: Footnote is built and tested on Claude
  • OpenAI
  • Vercel AI Gateway
  • Any server that speaks either APIsuch as OpenRouter, LiteLLM, or Ollama on your own computer

Other models than Claude work, and their answers are less complete. Choose a model provider

Your tool

A connector is one folder in the repository. The guide is written so a coding agent can follow it.

  • On the demo nothing leaves your computer, and no AI service is called. Ask uses a scripted reader that lists what's written.
  • Footnote reads your definitions from a Cube model folder or from the semantic_manifest.json that dbt parse writes, and never writes back.
  • It has not had its first release yet. It's for trying the idea on one computer with the demo, not yet for your company's data, and it will change before 1.0. We don't have a measured number yet for how much better agents answer with Footnote.
InstallRead the docsOpen source under Apache-2.0. Not released yet.