Documentation
Footnote is where the people who know why a number moved write it down, and the owner of that part of the business checks it. When someone asks an AI agent about the number, the answer says who checked it and when. When nothing is written down, the answer says so instead of guessing.
Try the demoOn your own computer, with Docker. The first start can take ten minutes.
Start
- Try the demoThe quick start. Install the demo on your computer with Docker, ask it a question, and connect Claude Code to it.
- Set up your workspaceThe six steps from an empty workspace to a first checked answer, and the page for each step.
- Connect an agentOne prompt for Claude, Cursor or another agent, and what approving does.
- WordsThe glossary of the ten words the screens and these pages use, each in one sentence.
Ask
- Ask a questionAsk about a number, read how the answer was made, ask again at other dates, and tell the owner what is missing.
- Answer linesThe six lines an answer starts with, and what each one asks of you.
- PrivacyWhat is kept about questions, what is never kept, and who can see a chat.
Write and review
- Write contextWhat a context item is, what to put in it, its source and its check date.
- Review a suggestionThe quick view, the review, Try it, and publishing, returning or closing.
- Keep context up to dateThe Inbox cards for a context item that is due for a check or was reported, and the ways to answer each.
- When nothing is written yetThe Inbox card a question with no answer raises, and how to write the context, check a draft or close it.
- Decide a conflictTwo items that say different things: compare them, then keep one, narrow one or keep both.
- Definitions and synonymsMetrics and terms, other names for them, and why some are read-only.
- Publish without reviewLet one agent publish low-risk new items in your domain, as your decision, and turn it off.
Set up
- Domains and ownersWho publishes what, how to add a domain, the sandbox, and inviting people.
- IntegrationsConnect where metrics are defined and where documents live, and what is read.
- See what agents askedObservability for owners and admins - what agents asked and got, saved questions in Quality, feedback, and alerts.
- Variants: a dimension, not a metric eachWhy one metric with currency as a dimension works better than a metric for each currency.
How it works
- How agents use itPublished context only, one answer line first, and what happens when nothing is written down.
Run the service
- Where to run FootnoteDocker Compose, a cloud host or container platform, or Vercel for the demo, and what each one asks of you.
- Use your own dataRun it on one computer with your own semantic layer instead of the demo, and keep a private profile apart.
- Run in productionRun the service and web images for real people, behind a TLS proxy, with company sign-in and their keys kept.
- Set up company sign-inLet people sign in to the web app with your company's OpenID Connect provider, and choose who may join.
- Agent sign-in settingsThe login server agents sign in through, how long a sign-in lasts, and how to make and rotate its keys.
- Choose a model providerThe ways to reach a model, the models tested so far, and who receives what Ask and the test runs send.
- Back up, restore and upgradeWhat to back up, how to prove a backup restores, how to restore it, and how to upgrade and roll an upgrade back.
- Watch the serviceCheck the service is healthy, read its logs, its request limits and its scheduled jobs, and serve its metrics.
- Service settingsThe settings the service reads from its environment when it starts, with their defaults.
- Administer the serviceChange settings, read the history log, import items, export an edition and use the footnote command.
- SecurityWhat is scanned on every change, what leaves the service, and where to report a security problem.
- Run the demo on Vercel and NeonThe web app as static files, the service as one function, and PostgreSQL 18 on Neon.
Connectors
- ConnectorsHow the service reads the tools you already use, what each connector brings in, and what a read may change.
- The Cube connectorReading definitions from Cube model files, in a Git repository or a folder on the server.
- Cube CloudReading metrics and dimensions from the metadata of a Cube Cloud deployment, through its APIs.
- dbt MetricFlowReading metrics and dimensions from the semantic manifest your dbt pipeline writes, wherever it puts the file.
- Static listMetrics and dimensions an admin writes in the form as a JSON list, for a trial or a team without a semantic layer.
- Markdown in Git and Google DriveThe two document connectors, which turn new and changed documents in the folders you choose into suggestions for owners.
- Connector securityWhat a connector may reach, how its credentials are kept, and how its output is treated.
- Build a connectorAdd a connector for your tool, by hand or with a coding agent, from the template to a passing kit.
Reference
- Agent toolsThe six tools an agent can call, what each takes and gives back, and how to read a reply.
- Connect an agent by handThe MCP address, each agent's own steps, access keys for bots, and the limits on what an agent may send.
- HTTP API overviewHow to call the service from a script, and where each part of the API is described.
Changes
- What's newChanges people notice, newest first.