Base

How to share a team knowledge base across ChatGPT, Claude, and other AI tools

Put your team's documents in one Base and connect each person's AI assistant to it. Everyone works from the same shared knowledge in the tool they already use, whether that is Claude, ChatGPT, Grok or a coding agent, within the access their role gives them, and nobody uploads copies into chats again. Contributors add and edit documents straight from their AI, and every teammate's assistant picks up the change on its next call, usually within minutes. Base hands back every fact with the quote it came from, so the whole team can check any answer. Working alone, it is the same connection: connect your AI once and it can search every Base you own or were sent.

Updated November 18, 2026

How the team workflow works

The owner sets the Base up once. Each teammate then connects their own AI app and signs in as themselves, so every assistant works from the same documents with that person's access. The owner:

  1. Creates a Base from the home page, named for the project, and adds the documents: PDF, Word, PowerPoint, Excel, CSV, plain text, Markdown and images, up to 50 MB per file. Base reads them into an overview, facts, a timeline and the people and organizations involved, and every fact keeps the quote it came from.
  2. Adds each teammate by email with a role. A viewer reads the Base, a contributor can also add and edit documents, and an admin can also manage it. People in the owner's workspace are already on the team of every Base in it.
  3. Keeps the Base invite-only, so only the people on the list can open it, and publishes it. Invited people get an email, and the Base is already in their AI's list, with no link to paste.

Each teammate:

  1. Connects Base to the AI app they already use: Claude from its connector directory, ChatGPT from its plugin directory, Grok or any app that accepts a custom MCP connector with https://createbases.com/api/mcp, and coding agents with npx base-knowledge init. Setup for each app is on the connect page.
  2. Signs in with the email address they were invited with, using the one-time link Base sends, and presses Allow on Base's consent screen.
  3. Asks their AI about the project. It searches the Base, answers with the quotes it used, and opens the full source whenever they want it.
  4. As a contributor, has their AI save what they produce, such as a brief or meeting notes, straight into the Base, or edit a document it already holds. Every earlier version is kept.

An illustrative example: three people, three AI tools, one Base

This example is illustrative. It shows what the tools do, not a customer's project. Maya, Devon and Sam are contributors on a launch Base.

  1. Maya, in Claude, asks what the customer interviews say about signup. Claude calls explore_base, which answers with quotes from the interview notes. She has Claude draft a launch brief from them and save it to the Base with add_document.
  2. Devon, in ChatGPT, asks what the brief means for launch. ChatGPT finds the brief with explore_base, reads it whole with fetch_document, and updates the launch checklist with edit_document.
  3. Sam, in a coding agent, asks it to build the signup flow from the brief. The agent reads the brief with fetch_document, writes the code in Sam's repository, and saves its implementation notes to the Base with add_document, where Maya's and Devon's assistants can find them.

Nobody forwarded a file or pasted a summary into a chat. Each of them signed in as themselves, each assistant did what that person's role allows, and the version history shows who added what. The brief, the checklist and the notes reached everyone because they were saved to the Base.

Which AI tools work, and what each can do

Any AI app that accepts a custom MCP connector gets every Base tool: searching, reading sources, adding and editing documents, and working on threads. What an assistant may do on a Base is decided by its person's role there, not by the app. The app decides which of its plans can add a connector and how it asks before running a tool.

AI toolHow it connectsWhat it can do in BasePlan and admin limits
Claude, on the web and desktopBase's listing in the Claude connector directory, or the connector URL as a custom connectorEvery Base tool the person's role allowsCustom connectors are on every plan, and Free allows one. On Team and Enterprise an Owner adds or enables the connector for the organization, then each member connects their own account.
ChatGPTBase Knowledge in the ChatGPT plugin directory, or the connector URL as a custom MCP server under PluginsEvery Base tool the person's role allows. ChatGPT asks you to confirm write actions by default.OpenAI decides which plans can add plugins and custom MCP servers. In a workspace, its admin decides which plugins members may add.
GrokA custom connector at grok.com/connectors, with the connector URLEvery Base tool the person's role allowsOn Grok Business and Enterprise, a team admin provisions the connector first.
Claude Code, Codex, Cursor and other coding agentsnpx base-knowledge init, or the agent's own command for a remote MCP serverEvery Base tool the person's role allowsNone from Base. An agent that cannot open a browser uses a key made under Settings, Connect.
Other apps that accept a remote MCP connectorThe connector URL, with OAuth sign-inEvery Base tool the person's role allowsSet by the app

Tools that change what readers see, cut off access or post in your name are marked so your AI app asks before running them. Step-by-step setup for each app is on the connect page, and coding-agent setup, keys and the full tool list are on the agents page.

What happens when someone changes a document

Base starts reading a new or changed document the moment it arrives, from the browser or from anyone's AI, and it is usually searchable within a few minutes. While it is being read, Base tells each assistant so, which keeps an assistant from reporting something as missing when it is on its way.

Every edit is a new version, and nothing is lost: the earlier version is kept, and the overview, facts and timeline are rebuilt from the new text. The rebuilt knowledge goes live in one step, so no one ever reads a half-updated Base.

Each assistant gets the current version every time it calls Base, and when a person opens the Base, their AI leads with what changed since they last looked. An answer already given in a chat stays as it was written; ask the assistant to check the Base again and it fetches the latest.

Edits land as targeted changes to the current version, and an edit made from an older version can be refused, so teammates editing the same document do not overwrite each other's work.

Who can read, edit, contribute, or revoke access

WhoWhat they can do
ViewerReads the Base in the browser or through their AI, within the owner's settings for downloads, AI access and comments. On Pro a viewer can be limited to particular documents. A viewer never adds or edits documents.
ContributorEverything a viewer can do, plus adding documents and new versions, editing documents, and working on task threads. A contributor always sees the whole Base, and the viewer settings do not limit them.
AdminEverything a contributor can do, plus managing the Base: its people, settings, sharing and documents. Archiving the Base and billing stay with the owner.
The owner's workspaceEveryone in the workspace that owns the Base can manage it, and every other Base in the workspace.
Anyone with the linkOn a link-mode Base, reads it as a viewer without signing in. Someone on the list who signs in still gets their own role.

Connecting an AI is separate from access. Each person connects their own AI app and signs in as themselves, and their assistant reaches only the Bases that person can reach, with that person's role. One person's connection never gives anyone else access.

Changes to access take effect right away. The owner or an admin can remove a person or change their role, pause the link, set an expiry date, or turn AI access off for viewers, and because Base checks every request from a browser or an AI against the current settings, the change applies from that person's next request. To disconnect your own AI, remove Base from that app's connector settings. Keys made for agents are revoked under Settings, Connect.

Revoking access cannot recall what someone already copied out, downloaded, or left in an assistant's memory, which is why Base lets you turn downloads off for viewers and limit them to particular documents from the start.

Invite-only keeps team material private. Link mode is for material you want anyone with the link to read, and search engines and AI crawlers can find it.

Why teams use Base for shared AI knowledge

Connecting a drive or a wiki to each assistant gives every AI the files. Base gives every AI the same understanding of them, and keeps that understanding right and improving on its own:

  • Facts with their sources. Base reads each document once into facts, a timeline and the people and organizations involved. Every fact keeps the verbatim quote it came from, and a fact whose quote is not in the source is dropped, so any teammate can check any answer.
  • Conflicting and outdated figures caught for you. When two documents give different figures for the same thing, Base flags both, each with its quote. When a later document updates a figure, Base marks the earlier one superseded instead of leaving each assistant to pick one.
  • It improves itself toward your goal. Each Base has an intention: who it serves and what they decide. When an answer comes back weak, Base fixes what it safely can, such as linking a term to its definition, checks that the fix did not make another answer worse, and asks the owner only what a document could settle and the goal needs. Every change is logged and can be undone. Pro runs unlimited goal checks; Free includes one per Base.
  • It gets faster the more the team uses it. Base reuses the documents that answered earlier questions: on a 31-question evaluation it answered 1.9 times faster with 44 percent fewer tokens.
  • Skills from the team's own work. Base finds work the team repeats in its history, writes it up as skills every teammate's AI follows, and rewrites them when the work changes.
  • One permission model for every AI. Roles and per-document access apply to every fact and summary, so a summary never reveals a document its reader cannot open.
  • Room for people outside your company. A client, investor or contractor joins with an email address, not a seat in your workspace tools, and uses whatever AI they already have.
  • One price for the whole team. Pro is $20 a month per workspace, not per seat, and nobody you invite pays anything.

Where Notion, Glean, Guru and the other options fall short

Each option below gives a team something. None of them documents what Base does with the knowledge itself: a source quote behind every fact with a rule that drops the rest, figures marked superseded when a later document updates them, and gaps found and fixed against the Base's goal. Every entry below comes from the vendor's own documentation, listed under Sources.

OptionKeeping the knowledge rightImproving with usePrice and access
BaseEvery fact keeps its verbatim source quote, and a fact whose quote is not in the source is dropped. Disagreeing figures are flagged with both quotes, and a figure a later document updates is marked superseded, on every new document or version.Fixes what it safely can after a weak answer, rechecks that the fix hurt nothing, and asks the owner only what serves the Base's goal. 1.9 times faster with 44 percent fewer tokens on a 31-question evaluation. Turns repeated work into skills and rewrites them when the work changes.$20 a month per workspace, not per seat. People outside your company join by email, pay nothing, and use any AI.
NotionAnswers link to the pages they used. Pages stay current when their owners re-verify them after verification expires, and Notion's guide says Q&A "can still discover" outdated pages. No documented check for figures that disagree between pages.Notion AI's memories of the pages you use often are private to each person. Custom agents run on instructions people write. No documented step that finds and fixes gaps in the knowledge.Priced per member, and Notion AI needs Business or Enterprise. Guests reach only the pages they are invited to and cannot ask custom agents questions. AI apps act with the user's full Notion permissions, and only Enterprise admins can limit which AI apps connect.
GuruAnswers cite the cards they used. Cards stay current through verification, by experts on a schedule or automated checks, and near-duplicate cards are flagged weekly once you have over 100. Guru's own guide says conflicting cards give "inconsistent results" until someone consolidates them.Shows topics that go unanswered and routes flagged answers to an expert, who makes the fix. Learns from experts' edits which sources to surface. Skills are written out step by step by a person.Every person invited to the workspace is a billed user, whatever their role, and pricing is by quote. Outside readers need Guru access, even to view an embedded card.
GleanAnswers cite sources after the statements they support, but answers from the model's own knowledge carry no citations. Owners verify documents or mark them deprecated, and deprecated documents may still appear in results. No documented check for figures that disagree between documents.Trains on your company's language to improve search, and reuses tool sequences rather than the documents that answered. People create skills with Glean's help, and updates from feedback wait for approval.Licensed per user, per month, with MCP use billed on top. AI apps sign people in through the company's SSO, and an admin has to enable it first.
HjarniStores notes your own AI writes; Hjarni "does not run any AI model on user data." Instructions to quote sources "do not guarantee" the AI follows them, in Hjarni's words.Every save is versioned and can be reverted. Templates and team instructions are written by people.Teams pay per seat ($200 a seat a year), every member needs their own Hjarni account, and in a team only the owner can share a folder. The free plan holds 25 notes.
ObsidianNo built-in AI: none of its core plugins is an AI, so reading and answering come from community plugins.Templates are written by hand.Every collaborator on a shared vault needs a paid Sync subscription, priced per user, gets the owner's permissions, and a shared vault tops out at 20 collaborators.
Claude ProjectsClaude searches the uploaded files once they outgrow its context window. No documented check for figures that disagree between files.A memory for each project. Skills come from a conversation where you explain your approach, and are updated when you ask.Shared only with people in your organization, on Team or Enterprise plans priced per seat. A shared project cannot hold Google Drive files.
ChatGPT ProjectsChats draw on the project's files and earlier chats. No documented check for figures that disagree between files.Project-only memory on Business and Enterprise. Skills are built when you ask ChatGPT to build one.Collaborators work inside ChatGPT, and only project owners can invite or remove them.
Google Drive with GeminiGemini numbers each sentence with the file it came from, and Google warns that it "might miss a source it used" and "can make up a source." Comparing files is something you ask it to do.Saves your conversation history. Workspace Studio flows are built from what you describe.Gemini in Drive needs Business Standard or higher. Each person connects their own Google account, and on Claude Team or Enterprise an Owner must add the Drive connector first.
A Markdown repository on GitHubPlain files that the connected AI reads as they are.Every change is a commit someone makes.Every collaborator needs a GitHub account, outside collaborators on an organization's private repository take a paid seat, and a personal repository cannot give anyone read-only access.
Built-in assistant memoryDetails one person shares with their own assistant, not your documents.Personalizes replies for that one person.Claude's memory belongs to one person, owners cannot view it, and on Team and Enterprise it stays off until each member turns it on.

Working alone: why connect instead of uploading to each chat

  1. Create a Base from the home page and add the documents.
  2. In Claude, open Customize, then Connectors, then Browse connectors, find Base and press Connect.
  3. In ChatGPT, open Plugins, find Base Knowledge, press + and connect it when ChatGPT asks.
  4. In Grok or any app that accepts a custom MCP connector, add https://createbases.com/api/mcp and leave every OAuth field blank.
  5. In Claude Code, run claude mcp add --transport http createbases https://createbases.com/api/mcp && claude mcp login createbases. For other coding agents, npx base-knowledge init sets up the connector and the skills.
  6. Sign in with the one-time link Base emails you, then press Allow on the consent screen.
  7. Ask your AI about the documents.
  • The documents are read once, into facts with quotes, a timeline and the people and organizations involved. Your AI searches that instead of rereading every file in every conversation.
  • Every Base you own or were sent is in one place, from every AI app you connect.
  • When you update a document, your AI's next search uses the new version once Base has read it.
  • A connector can be removed. A file pasted into a chat stays in that chat.

Questions

Can people using different AI assistants work from the same Base?

Yes. Claude, ChatGPT, Grok and coding agents all connect to the same Base and work from the same documents and the same cited facts. Each assistant words its answer its own way, and every one points back to the same sources.

Why not use Notion, Glean or Claude Projects instead?

Because none of them documents keeping the knowledge itself right. Notion and Glean rely on owners re-verifying pages, Guru's own guide says conflicting cards give inconsistent results until someone consolidates them, and none of the three documents finding and fixing its own gaps against a goal. Base ties every fact to its source quote, flags conflicting figures and marks superseded ones, improves itself toward the Base's goal, and gets faster the more the team uses it. It also costs $20 a month per workspace rather than per seat, while Notion is priced per member, Glean per user, and Claude Projects are shared only inside a Team or Enterprise organization.

Does every teammate need their own account and connection?

Yes, and that is what keeps a shared Base safe. Each person signs in with the email address the owner added, using a one-time link Base emails them, and authorizes their own AI app. Nobody shares a login, every assistant acts with its own person's role, and the version history shows who did what.

Can teammates edit documents from their AI?

Yes, as contributors or admins, from any connected AI. Their assistant can add documents and new versions, and edit text, Markdown, CSV, Word and Excel files in place, with Word edits recorded as tracked changes. A PDF or slide deck is updated by adding a new version. Viewers read.

How soon does a teammate's AI see an update?

Usually within a few minutes. Base reads the change as soon as it arrives, and every teammate's assistant uses it on its next call. While a document is still being read, Base tells the assistant so.

Are chat histories or assistant memories shared through Base?

No. Conversations and assistant memories stay private in each person's own AI app. Base sees only what an assistant sends it, such as a question or a document, and records those calls in the Base's activity. What the team shares is what someone deliberately saves to the Base: documents, comments and tasks.

What does it cost?

Connecting and reading are free for everyone. Base is free for up to 2 Bases with 300 credits and no card. Pro is $20 a month for unlimited Bases. Credits come from the workspace that owns the Base: when it reads new documents, whoever added them, and when the owner lets it work harder on a question.

Does this work with a coding agent?

Yes. Run npx base-knowledge init to set up Claude Code, Codex, Cursor and similar agents, or add the connector at https://createbases.com/api/mcp yourself. The agents page lists the tools and limits.

Can my AI create a Base for me?

Yes. Connected, an AI can create a Base, add documents and publish it, then hand you the share link.

Sources checked on November 18, 2026

Try it with your own documents

Free for your first Bases, no card. Every new account starts with a Base that explains Base, so there is a cited answer to inspect right away.