Integrates with Git repositories, storing wiki content as plain files in a Git repository and providing version control, commit history, diffs, and sync capabilities for wiki pages.
The vibe-coded, open-source alternative to Notion and Confluence. Your team's wiki lives in Git. People write in the browser, together, in real time. AI agents read and write the same pages. One command to run it yourself.
Website: https://foliowiki.online · Featured in Self-Host Weekly (9 Oct 2026)
No install, no sign-up: open it and pick Sam on the sign-in screen. Shared login, data resets every 24 hours; don't enter personal data or API keys.
Folio was built to replace Confluence and Notion for a real team — and it was built by AI agents, in the open. The plans, specs and hand-off notes the agents work from are in this repository, next to the code they wrote. More on that below.
git clone https://github.com/evergreen-it-dev/folio.git
cd folio
docker compose up -dOpen http://localhost:4870, create the first account, start writing.
Not ready to install? Try the live demo first.
You own the content. Every page is a plain file in a Git repository you control. No export step, no lock-in: stop using Folio tomorrow and the files are still there, readable in any editor.
It works the way a wiki should. Real-time editing, a page tree, search, access rights, history, sharing — and it keeps working offline.
Agents are first-class users. Not a chatbot bolted on the side: a built-in assistant, your own agent in every space, a full MCP server, and Markdown links made for agents.
Documents, tables, whiteboards, diagrams and files in one tree. A roadmap table next to the spec, next to the architecture board, next to the signed PDF.
Publish what should be public. Turn a space into a read-only website, with search and an optional "Ask AI" for visitors, at a path of your instance, on a subdomain or on your own domain.
One command to run. Docker, and nothing to configure for a first start.
Related MCP server: Engram
Short walkthroughs, no slides. The voice in the videos is synthetic.
A team wiki that lives in Git — edit a page in the browser, press Sync, and it is a real Git commit; a teammate edits from the terminal and the change shows up on the page (0:37)
An AI agent updates your wiki over MCP — one request: the agent reads the release process, writes release notes and adds a roadmap row (0:53)
Onboarding a sales rep with Folio AI — answers from the team's own pages with sources, a weekly check as a form, and the results in a table (1:22)
Whiteboards that live in Git — draw together, embed the board in a document, and keep every version in history (0:51)
Docs as code without the IDE — the docs/ folder of a product repository becomes a space; a technical writer edits it by hand and with Folio AI, an agent adds a release entry over MCP, and the developer reads the commits with git pull (1:36)
Publish a space as a public website, with Ask AI for visitors — one setting makes a space a read-only site on a path, a subdomain or your own domain; visitors search it, download pages and ask Folio AI (1:21)
The editor and Mermaid diagrams — live Markdown, a source view of the same page, and diagrams you edit visually or as code (1:04)
Is your wiki's AI assistant helping? Folio AI analytics — instance admins see rated conversations and unanswered questions, add the missing page, and the same question gets an answer with its source (1:21)
Most wikis keep your content in a database and offer Git as an export. Folio does the opposite: the Git repository is the wiki. The database holds the index, access rights and the state of live editing — never the only copy of your words.
This is what a space looks like on disk:
acme-handbook/
├── index.md
├── welcome-to-acme.md
├── engineering.md
├── engineering/
│ ├── onboarding-checklist.md
│ ├── release-process.md
│ ├── incident-runbook.md
│ └── how-folio-fits-together.excalidraw.svg ← a whiteboard
├── meeting-notes.md
├── meeting-notes/
│ └── product-sync-september-28.md
├── roadmap.table.md ← a data table
├── onboarding-feedback.table.md
├── onboarding-feedback/
│ └── onboarding-feedback.form.md ← a form
├── security-policy.md
├── security-policy.pdf ← files are pages too
└── security-policy.docxAnd this is its history — ordinary commits, with the people who made the changes as authors:
$ git log --format='%h %an %s'
be529a0 Maria Chen folio:update
aeac3bb Maria Chen folio:update
6c2c061 Sam Rivera folio:update
026d9ec Maria Chen folio:update
...
4e9ec49 Folio init: create spaceWhat that gives you:
Edit anywhere. Change a page in the browser or in your IDE and push — both land in the same history. Links between pages are ordinary relative Markdown links, not a private dialect.
Bring your own repository. Start with an empty space, clone an existing repository, or connect one later. Several spaces can share one repository, each in its own folder — so documentation can live right next to the code.
Review like code. Merge requests, blame, diff and revert work on your wiki because it is a repository.
Docs-as-code friendly. Front matter keys written by Docusaurus, MkDocs,
Hugo or Jekyll are kept byte for byte; a space bound to a folder such as
docs/ writes no ids into your files (a space over a whole repository does);
images can be stored in the repository next to the page and linked by a
relative path, so GitHub shows them; and moving a page rewrites the links
that point to it and the ones inside it, in one commit. A full scenario is in
Docs as code, without the IDE.
Nothing is lost quietly. Sync state is always visible — clean, ahead, behind, conflict — and conflicts are listed for the whole space. "Take the version from Git" resets a space to the remote branch and saves what was there into a backup branch first.
Readable formats. A whiteboard is an SVG that opens in any browser and still carries the editable scene. A data table is one Markdown file — the column types and saved views in the front matter, the rows as a table you can read:
# Roadmap
| Initiative | Status | Owner | Team | Due | Effort, days | Urgent | id |
| --- | --- | --- | --- | --- | ---: | --- | --- |
| Single sign-on for enterprise plans | In progress | @alex | Engineering | 2026-10-09 | 8 | [x] | 1787nfg3 |
| New onboarding flow | In review | @sam | Design | 2026-10-02 | 5 | [ ] | b227mwwh |
| Pricing page redesign | Planned | @sam | Design | 2026-10-16 | 3 | [ ] | ndxvtpa3 |
If your documentation lives in the docs/ folder of a product repository,
Folio can show that folder as a space. Writers edit it in a browser, developers
keep their IDE and read ordinary commits, and an agent works on the same pages.
Connect the folder, not the repository. Create a space from the
repository URL, choose the branch and pick docs/ as the root path. Only
that folder becomes the space: Markdown files are pages; source code and
other files are not pages and are never touched. Folio keeps its page ids in
its own index, and the only file it adds is a hidden docs/.<slug>.folio.
A writer edits in the browser. Reading, Live edit or Source (plain Markdown), several people on one page, or a request to Folio AI in the side panel. Folio writes each edit to the file at once and commits after about 90 seconds of quiet, or immediately on Sync, to the branch the space follows.
A developer sees a commit with the author's name. git pull, git log
and git blame show the person who edited as the author; Folio is the
committer and the message is folio:update. Only the last editor in a window
is recorded as author. The developer's own commits come back into the space
by fetch roughly every three minutes and on every Sync.
An agent updates pages through MCP. Claude Code, Cursor or any MCP client connects with a personal access token or through OAuth, reads and edits the same pages with the rights of the token's owner and no more, and the commit carries that person's name.
The space can become a docs portal. Make it public and the same pages are served as a read-only website with search and, if you switch them on, downloads and "Ask AI" for visitors, see Publish a space as a website.
Where it stops: a space follows exactly one branch and Folio pushes straight to
it, without pull requests (point the space at a docs branch if you want
review before main; a protected branch makes the push fail and the space
shows git's message). A space over the whole repository, with no root path,
writes an id into every page and a small file into the repository root,
unless an administrator ticks "Light footprint" (a repository with .obsidian/
is detected on its own). Opening a page writes nothing: a file changes only
when someone edits it. A page's title comes from its first # heading, then
from a front matter title:, then from the file name, and Folio does not add a
heading to a page that has none. An id that another tool put in front matter
(Docusaurus and Jest use id: intro) never moves a page between spaces. MDX
files (.mdx) are not pages; the scan logs how many it skipped. Pasted images
stay in Folio's asset store until a space administrator turns on "Store images
in the repository". See also What is not there yet.
The walkthrough is in the post Docs-as-code without the IDE: edit your repository's docs in a browser and in the video.
Documents, data tables and whiteboards are edited by several people at once. You see who is on the page and where their cursor is; changes appear as they are typed. When the page goes quiet, Folio commits — with the person who edited as the author.
What you type is not lost between Live edit and Reading: the browser keeps a local copy until the server confirms it has your edits. If it cannot confirm them, the header says Not saved and closing the tab asks first.
On a train, on a plane, on bad hotel Wi-Fi — keep writing.
Pages and whiteboards you have opened stay editable without a connection.
New pages and whiteboards can be created offline and appear in the tree straight away.
Everything is kept on the device and reaches the server by itself when the connection returns — at the same address, with no duplicates.
Nothing is lost if you switch pages or close the tab.
An indicator in the header shows what is waiting to be sent.
Data tables, forms and file uploads need a connection.
A Markdown editor that does not make you think about Markdown. Three modes — Reading, Live edit and Source — so that writers and engineers are both at home.
Callouts, checklists, tables with merged cells and colors, code blocks,
collapsible sections, highlights, ++underline++, and coloured status
tags like :status[Done]{color=green}.
Paste from Excel and Google Docs keeps tables and formatting; copying out of Live edit keeps it too.
[[ links a page, @ mentions a person, / inserts anything.
::pagetree puts a live tree of child pages right into the text.
Outline and notes panel, backlinks, page icons and covers, templates, glossary terms with hover cards.
Light and dark themes, following the system or chosen by hand.
Mermaid diagrams come with a visual editor. You build the diagram on a canvas, and Folio writes the Mermaid code for you — or you type the code and watch the canvas follow. Both sides stay in sync.
Click to build. A plus on a node adds a connected one, a plus on a link inserts a node into it, a handle drags a new link from node to node.
Nine diagram types from templates: flowchart, sequence, classes, states, ER, Gantt, timeline, pie, mind map.
Code and canvas side by side, full screen when you need room, and a readable message instead of a broken picture when the syntax is wrong.
Still plain Markdown. The diagram is stored as an ordinary mermaid
code block, so GitHub and GitLab render it too.
Spreadsheets that belong to the wiki: typed columns, saved views, filters and sorting — a small database you can link from any page and query from any agent.
Column types: text, long text, number, date, checkbox, select, status, user, link.
Views as tabs, each with its own columns, order, widths, multi-column sort
and and/or filters — including "me", "today", "this week".
Edited together in real time, stored as one readable file in Git.
Import and export; up to 20,000 rows per table.
Forms turn a table into a questionnaire: every answer becomes a row. A form can be public, so people answer without an account.
Excalidraw boards inside the wiki — architecture, flows, workshops. Drawn together in real time and saved as SVG files that render anywhere.
The board above was not drawn by hand. An agent created it through MCP from a short description of boxes and arrows; people then keep editing it like any other board.
Reactions. Select a shape in Edit mode and a small button appears at the corner of its selection frame: pick an emoji (✅ ❌ 🔧 for yes, no and in progress, or 👍 ❤️ 😂 🎉 🤔 👀 🔥 👏) and a counter chip shows under the shape. Click a chip to add or remove your own; point at it to see who reacted. Reactions are live for everyone on the board, and people who can only view it (or are in View mode) see them too but cannot click them. They are kept with the board's live state, not in its SVG file.
PDF, Word, Excel and PowerPoint files are pages too. Drop one into the tree and it opens right inside Folio — no download, no other application, no conversion on the server.
PDF in the browser's own viewer: thumbnails, zoom, search, print.
Word, Excel and PowerPoint rendered in the browser.
The file sits in the space's Git repository next to the pages about it, in the same tree, under the same access rights.
Open in a new tab or download with one click.
Folio never rewrites your file: what you uploaded is byte for byte what is stored.
A new version of the file, same page. "Replace with new version…" in the tree menu, the button in the page header, or drag a file onto the open page. The page keeps its link, place in the tree and access; the new file may even be another type (a deck that became a PDF), and links to it from other pages are rewritten. A toast offers Undo.
Every version stays. Each replace is its own commit. The page history lists the versions with their file name and size, lets you download any of them (or preview a PDF) and restore one, which adds one more commit rather than rewriting anything.
These files are for viewing: they are not edited in Folio, and search finds them by title.
An assistant in the side panel that works with your wiki, not next to it.
In the screenshot the assistant was asked to add an initiative to the roadmap and say which ones are urgent. It read the table's schema, inserted the row — it appears in the table on the left — and answered from the data.
Ask — read-only. Find, summarize, compare, explain across the pages you have access to.
Agent — reads and writes. Drafts a page, updates a table, draws a board, restructures a section — inside your rights, never beyond them.
Runs live on the server. Close the browser and the work continues; come back and pick up where it was.
The same tools as everyone else. The assistant works through the very tools an outside agent gets over MCP — there is no second, weaker implementation, and in Ask mode the write tools are switched off on the server, not just hidden.
Folio collects three signals about answer quality, without asking anyone to fill in a form:
👍 / 👎 on an answer. Every saved assistant answer can be rated "Helpful answer" or "Not helpful". One rating per answer, and you can change or remove it.
A short survey. After every third answer in a conversation the panel asks "Did the assistant solve your question?" — Yes, Partly, No or Not now — with an optional comment.
Questions without an answer. When the pages do not contain the answer, or the assistant is not sure of it, it records the question and what was missing — briefly, with links to the pages it did read — then still answers the person plainly.
Instance administrators see all of it on one page, Assistant analytics
(/admin/assistant, from the account menu): the conversations with their
ratings and survey results, and a second tab, "Questions without an answer",
which is a ready list of what your documentation is missing — the person's own
words first, the assistant's restatement below, and links to the pages
the assistant read. Both tabs filter by space,
user and date. A space administrator sees the same page limited to the spaces
they administer: only conversations that ran there, and only the messages of
those runs. An administrator can open a conversation to read it; the
conversation page lists who opened it and when, and every opening is written
to the audit log. Other people do not see this page.
.agentEvery space has a .agent folder. Whatever you put there becomes the
assistant's instructions for that space — and it is written as ordinary
Folio pages, not as configuration.
Build an agent by writing pages. Its role, its tone, the glossary, the rules it must follow, the templates it should use, the process it should walk people through.
A different agent for every team. The support space gets an agent that answers like support. The engineering space gets one that knows the release process. Legal gets one that never improvises. Same Folio, different agents.
Versioned like everything else. The instructions live in Git: you can see who changed the agent's behavior, review the change, and roll it back.
Owned by administrators. Only space administrators see and edit
.agent; everyone else just gets a better assistant.
The instructions apply to every run in the space, in Ask and Agent modes alike.
Folio does not resell AI and does not mark it up. The assistant runs on a Cursor key: your own, or one that an administrator provides for a space or for the whole instance. Whoever owns the key pays.
Provider | Status |
Cursor | available — each person connects their own key, or an administrator provides one for a space or for the whole instance |
Others | not supported yet |
Who pays. A run uses the first key that exists: the person's own key, then
the key of the space, then the key of the instance, then CURSOR_API_KEY from
the server environment. A space administrator sets a space's key (space menu,
"Folio AI key"); an instance administrator sets the instance key and can turn
personal keys off, so everyone uses the shared one. Shared keys are stored
encrypted and are never shown again after saving; every change is in the audit
log. Visitors of a public site have no key of
their own, so their questions run on the key of the space, of the instance or
of the server environment, in that order.
Without a key everything else in Folio works as usual.
MCP server. 26 tools over /mcp: list and search spaces, read and write
pages, read a page's outline, tags, properties and links, find unresolved links
and orphan pages, create and edit whiteboards, query and update data tables,
read history.
Connect with a personal access token, or through OAuth 2.1 from clients that
run in a vendor's cloud (claude.ai, Claude Desktop, ChatGPT). search and
fetch are there for ChatGPT deep research.
claude mcp add folio --transport http https://wiki.example.com/mcp \
--header "Authorization: Bearer folio_pat_…"Ready-made setups. Files for Claude Code (a plugin and a skill), Cursor, VS Code, Codex, n8n (two importable workflows), Open WebUI, ChatGPT and Claude live in integrations/: configs with placeholders, rules for the agent and step-by-step instructions.
Markdown for an agent. Any page — with its child pages, if you want — is one link that returns clean Markdown. No scraping, no HTML, no login flow:
curl https://wiki.example.com/share/<token>.md
Export in the formats agents and people need: Markdown, YAML, PDF, DOCX, or a zip of a whole space.
For claude.ai or ChatGPT, add https://wiki.example.com/mcp as a custom
connector, sign in to Folio and allow read or read + write on the consent
screen. The connector acts with your rights and its writes are saved in Git
under your name. Steps: docs/MCP.md.
Safe by design. A token or an OAuth grant never has more rights than its
owner, a read token cannot write, and administrative actions are not
available to either at all. Details: docs/MCP.md.
Point Folio at a Confluence page or a whole tree and it brings the pages over as Markdown — with images, code, checklists, and panels converted into callouts. Pages come from Confluence Cloud and from on-premises installations; whiteboards from Confluence Cloud become Excalidraw boards.
Run a test import first: not every macro has an equivalent. What is not there yet is written down honestly.
Who sees what is decided in one place, and it is visible at a glance.
Roles per space: administrator, editor, viewer.
Private, open or public spaces. A private space is visible to its members only. A space open to the instance gives every signed-in person read access, and edit rights still have to be granted. A public space is also readable by anyone on the internet as a website, see Publish a space as a website.
The access matrix. People down the side, spaces across the top, a role in every cell. Change one cell or many at once.
An administrator is not a reader. The instance administrator manages people and access but does not see the content of a private space they are not a member of. They can grant access to themselves — and that is written to the audit log and marked for everyone to see.
Audit log. Every grant, revoke and visibility change is recorded with who did it and when.
Page-level access. Any page can be narrower than its space: open to all members, to you only, or to selected people with view or edit rights.
Getting people in
Invitations by link with preset spaces and roles, an expiry date and a usage limit.
Access requests. Someone who opens a space they cannot see can ask for access; administrators are notified at once and choose the role.
Sign in with Google, limited to the email domains you list. A new person gets an account and no access until someone grants it.
Deactivate an account without deleting what the person wrote.
Sharing outside
Share links for people without an account: view or edit, one page or a whole subtree (pages with their own access restrictions are left out), revocable at any time.
Forms that accept answers without an account.
For agents and scripts
Personal access tokens with read or write scope, or an OAuth connection
from claude.ai, Claude Desktop or ChatGPT (removable under "Connected apps").
A token never has more rights than its owner, and administrative actions are
not available to tokens at all.
The .agent folder is visible to space administrators only.
One setting turns a space into a read-only website for people without an
account. Open the space menu, choose Publish as website and make the space
public. It is the same space, not a copy: edit a page and the site shows the
change. Signed-in people read a public space the way they read any space that is
open to the instance. Typical uses: a documentation portal built from the
docs/ folder of a repository (see
Docs as code, without the IDE), a public
handbook, a changelog.
A short walkthrough: Publish a space as a public website, with Ask AI for visitors (1:21).
How the site looks is kept in your repository, not in the database. Open the
space menu, choose Site & PDF settings… and one page edits the site title,
description, language, favicon, logo and social image, the site's stylesheet
and the PDF header, footer and print styles. Everything is saved as files in a
.folio/ folder at the root of the space's content (site.json, site.css,
export.css, pdf-header.html, pdf-footer.html and assets/), one commit
per save, so you can edit the same files in Git too. Search engines get a
<head> per page, robots.txt and sitemap.xml; an "index this site" switch
and a primary address are in the publishing dialog.
Address | When it works |
| Always, as soon as the space is public. Nothing to set up. |
| When the operator sets |
| A custom domain for one space: a CNAME to the host of your Folio, and the reverse proxy routing that host to Folio, with a certificate for it. |
What visitors get. The page tree, search inside the space, documents, data tables (live, read-only), whiteboards as pictures, forms (view only) and PDF and Office files, with breadcrumbs and a layout for phones.
Downloads as Markdown, PDF or DOCX, switched on per space.
Ask Folio AI, switched on per space: visitors ask questions and the assistant answers in Ask mode only, from the public pages of that one space, on the key of the space or the instance. It cannot change anything, and questions are rate limited per address and per site.
What is never published. Pages with their own access restrictions, the
.agent folder and templates. A restriction is not inherited by child pages,
so check the tree before you publish and lock every page that must stay closed.
Who can publish. Space and instance administrators, from the browser. The
decisions (who sees the site, where, downloads, Folio AI, indexing) live in
the database, not in Git; the look of the site lives in .folio/ in the
repository and is changed by space administrators. In the public demo the
Folio Handbook is published at
https://demo.foliowiki.online/public/folio-handbook; you can try the dialog
there (no subdomain, custom domain or Ask), and everything resets daily.
Search engines. Each page carries its own title, description, Open Graph
tags and canonical link, and a page with unlisted: true in its front matter
stays out of the sitemap. The public demo is never indexed.
Not yet. There is no PDF preview before saving, site.css cannot load
external fonts or images (put them in .folio/assets/), and the rate limits
are counted per server process.
What the operator sets up for subdomains and custom domains: docs/INSTALL.md. Limits: docs/LIMITATIONS.md.
You need Docker and Git. Nothing else.
git clone https://github.com/evergreen-it-dev/folio.git
cd folio
docker compose up -dThe first start builds the application and takes a few minutes. Then open http://localhost:4870 and create the first account — it becomes the administrator.
A short welcome takes it from there: pick what to create first — a document, a whiteboard or a data table — look through what Folio can do, and land in your first space with that page open. Connecting an existing Git repository is one link away on the same screen.
Where | How |
Your computer |
|
A server or VPS | one command from the published image: |
Coolify, Railway, Render, DigitalOcean, Dokploy, CapRover, Easypanel, Portainer, Fly.io | ready-made kits in |
Any other container platform | from the |
Vercel, Netlify and other serverless hosting | not supported — Folio is a long-running server that keeps Git repositories on disk and holds live connections |
One click on Railway: 
Domains, HTTPS, updates and backups: docs/INSTALL.md.
No telemetry. A Folio you run sends nothing about its use to anyone: no
analytics library is loaded and no connection is opened to the Folio authors or
to a third party. There is one opt-in exception for people who run a public
demo: set FOLIO_POSTHOG_KEY to a PostHog project of your own and Folio sends
anonymous events (page kinds, space slugs, never content). Without that key
nothing happens. Details: docs/INSTALL.md.
Folio is written by AI coding agents. A human product owner decides what to build, tries every change in the browser and decides what ships; agents plan the work, write the code and the tests, and document what they did.
From the first commit on August 21, 2026 to version 0.1.0 on September 29, 2026 took 39 days and more than 400 commits upstream. Version 0.2.0 followed on October 14, 2026, 54 days in. Today that is roughly 110,000 lines of TypeScript and another 70,000 lines of tests. It runs in production for a real team.
We publish how it is made, not only what was made:
File | What it is |
How the work is organized and the rules agents follow in this repository | |
The current state: how Folio is put together, what must never break, what is open, and the lessons that cost us something to learn | |
What has been built, in the order it was built, and what comes next |
If you are building with agents yourself, this repository is a worked example of a non-trivial product made that way. If you only want a wiki, none of it gets in your way.
Public demo: demo.foliowiki.online — shared login, resets every 24 hours
Free tools, no sign-up: Confluence to Markdown converter · Mermaid live editor
Markdown templates: meeting notes, PRD, SOP, runbook, postmortem, onboarding checklist and more (CC0)
Alternatives, compared honestly: Confluence · Notion · self-hosted wikis
Docs-as-code without the IDE: edit your repository's docs in a browser
cp .env.dev.example .env
docker compose -f docker-compose.dev.yml up -d # PostgreSQL, Redis, MinIO
npm install
npm run dev # http://localhost:4871Node.js 22.12 or newer (22.13 or newer for the assistant). See CONTRIBUTING.md — and AGENTS.md if your contributor is an agent.
Questions and bug reports: SUPPORT.md. Please report vulnerabilities privately, as described in SECURITY.md.
This server cannot be deployed
Self-hostable team wiki; agents read & write it via MCP; Atlas turns your repo into a cited wiki.
A shared team wiki your coding agents read and write — across every repo and every MCP client.
Shared project wiki for AI agents: read and write pages, next actions, and activity logs over MCP.
Hosted markdown project wikis your team's AI assistants read, search, and update over MCP.
