Most teams now use more than one AI tool. One person drafts in ChatGPT, another lives in Claude, the developers work in Cursor or Claude Code. Every one of those tools starts the day knowing nothing about how your team works.
The knowledge exists. It is the way you price a deal, the decision you made in March and why, the template the sales team swears by, the five steps of a recurring task. It sits in Drive folders, Notion pages and people's heads, and each person re-explains it to their own AI, slightly differently, every week.
Warmhand exists to solve one narrow problem: package what a team knows, and deliver it efficiently to whichever AI each person uses. Before we build further, we want to be honest about how teams solve this today, and where those solutions are already good enough.
This is the first post in a series where we go head to head with each of these tools. This one is research. The posts after it measure.
How we did this
- Sources. We read each vendor's own documentation and pricing pages on 2 October 2026. Every claim about a product comes from its own pages.
- Gaps. Where a page would not load for us (OpenAI's help center and pricing pages among them), we say so, and we do not quote numbers we could not check.
- Bias. We are a competitor writing about competitors. We tried to describe each product the way its own team would accept as fair. If we got something wrong, tell us and we will correct it.
- Not a benchmark. Nothing in this post is a measurement yet. That is what the rest of the series is for.
What solving it actually requires
The problem has two halves, plus one complication.
Packaging. Someone decides what goes in, keeps it true, and decides who gets it. A sales playbook should reach the sales team and the new contractor, not the whole company. A pricing rule from last year should stop being handed out once it is replaced.
Delivery. When someone asks their AI a question, the right piece of that knowledge has to reach it, at a cost that does not crowd out the actual work. A 40-page handbook pasted into every conversation is delivery, technically. It is also slow, expensive, and the answer gets buried.
The complication. The people receiving it use different AI tools, and some of what a team knows is not reference material at all. "Always quote in euros for EU customers" is a rule the AI should follow, not a fact it might stumble on. "When someone asks for a renewal quote, do these five steps" is a skill that should kick in when the moment comes.
The short version
Almost every option today lands in one of two camps:
|
You choose exactly what the AI gets |
It works in any AI tool |
| Claude Projects, ChatGPT shared projects, Gemini Gems |
Yes |
No, one vendor's AI only |
| Google Drive, Notion, Confluence, Guru (through their AI connectors) |
No, the AI sees everything you can see |
Yes, broadly |
| Obsidian |
No, a shared vault is all or nothing |
Desktop and coding tools, through a local plugin |
| A text file you paste in |
Yes |
Yes, by hand, every time |
| Instruction files in a code repository |
Yes |
Coding tools only |
The tools that let you hand-pick the contents keep it inside their own AI. The tools that reach every AI give it everything the user can open. The two things that do both are a file you paste by hand and an instruction file in a git repository, which is a fine answer if everyone on the team writes code.
The options, one by one
Re-explaining, every time
Good at: costs nothing and needs no setup.
Where it stops: at the second person. The memory features in these tools are personal. Claude's help center says it plainly: "Memory is not shared with teammates."
A shared text file
Good at:
- It is free and works in every chat window on earth.
- The AI reads exactly what you gave it, nothing more.
Where it stops:
- Someone has to paste or upload it in every conversation.
- Copies drift the moment anyone edits one, and nobody knows who has which version.
- Giving different people different subsets means maintaining several files.
Instruction files in a code repository (AGENTS.md, CLAUDE.md, Cursor rules)
Good at:
- Coding tools load these automatically at the start of every session.
- They are versioned and reviewed like code.
- Rules can be scoped to parts of the codebase.
- For developer teams this is genuinely good, and we would not tell anyone to stop.
Where it stops:
- Only coding agents read them. Web and mobile chats do not.
- The audience is whoever has the repository checked out, which leaves out sales, support and the contractor.
- Everything loads whole, every session. Anthropic's own guidance for Claude Code is to "target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence."
Google Drive
Good at:
- The documents already live there, and they are always the latest version.
- Claude's Drive connector is available on every Claude plan and answers with citations that link back to the original files.
- Gemini in Drive lets you pick specific files or folders as its sources.
- ChatGPT projects accept a link to a Drive folder.
Where it stops:
- Claude's connector works across the whole Drive you can access. Its help page describes no way to limit it to one folder.
- Folder-level scoping exists, but only inside Gemini or inside a ChatGPT project.
- Everyone who receives the knowledge needs Drive access to the underlying files, so sharing with an AI also means managing Drive permissions.
- Long files have to "fit within the conversation's context window", and images inside documents are not read.
- Google's own page warns that with too many sources, "Gemini might base its response on some, but not all, of the content."
Cost: Google Workspace runs $7, $14 or $22 per user per month. Gemini in Drive is listed on the $22 Business Plus plan.
Notion
Good at: Notion runs an official, hosted MCP server: the standard way AI tools connect to outside data. Claude, Cursor, Claude Code and others can read and write pages through it, live.
Where it stops:
- The AI can "read and update content that you can access in the selected Notion workspace." The documentation describes no way to say "only these pages, for this job."
- Notion's agent and enterprise search features sit on the Business plan and up.
Cost: Free, Plus at $10, Business at $20 per member per month, Enterprise on request.
Obsidian
Good at:
- Your notes are plain files on your own disk, which coding agents read directly.
- The app is free, including for work.
- Shared vaults merge edits from several people.
Where it stops:
- Whole vault only. In a shared vault, "All collaborators receive the same permissions as the vault owner," "Fine-grained permissions are not supported yet," and a vault holds at most 20 collaborators. Sharing part of your notes means a separate vault.
- Technical setup. Connecting an AI means a community plugin running on your own machine plus an MCP server and a configuration file, which rules out most non-technical teammates and every web or mobile chat.
Cost: the app is free. Sync is $4 per user per month billed annually ($5 monthly), and Publish is $8 per site.
Claude Projects
Good at:
- Upload files and instructions once, chat against them repeatedly.
- On paid plans Claude switches on retrieval automatically when a project outgrows the context window.
- Sharing has view and edit roles.
Where it stops:
- Sharing needs a Team or Enterprise plan, and you can only add people "from your organization."
- It is Claude only. Anthropic's project pages say nothing about project knowledge reaching Claude Code, the API, or any other tool.
- Free accounts can create up to five projects but cannot share them.
Cost: Team seats are $20 per month billed annually ($25 monthly), for teams of 2 to 150.
ChatGPT shared projects
This is the strongest alternative we found, and it deserves to be said clearly.
Good at:
- "Project sharing is available on ChatGPT Free, Go, Plus, Pro, Business, Enterprise, and Edu."
- You can invite people, invite a group, or share a link that any logged-in ChatGPT user can join.
- A project holds files, pasted text, saved answers and links to Drive folders and Slack channels.
- It keeps its own memory, separate from each member's personal memory.
Where it stops:
- ChatGPT only. Your Claude and Cursor users get nothing from it.
- File limits. A project holds 5 files on Free, 25 on Plus, and 40 on Pro and the business plans.
- Managed workspaces. There, invitations "may be limited to members of your workspace."
- Built to collect, not to curate. OpenAI itself describes a shared project as a "live context hub" where everyone's chats accumulate. That suits ongoing group work, and it is a different thing from a small set of material someone vouched for.
Cost: included in every ChatGPT plan. We could not load OpenAI's pricing page to quote plan prices.
Gemini Gems
Good at:
- A Gem bundles instructions and files, and you can share it.
- Files from Drive stay current: "Gemini will use the most recent version of the file."
Where it stops: Gemini only.
Google also says that "starting in November 2026, Gems will go away on personal Google accounts and automatically transition to skills," so this one is moving.
Knowledge tools with AI built in: Guru, Confluence, Asana
What they share: each grounds its AI in the content it already holds, and each opens that content to outside AI tools through an MCP server.
Where it stops: those servers answer with everything the asking user can see.
- Guru: "Each query only returns knowledge that the authenticated user can access."
- Atlassian: "MCP clients can perform actions on all connected products (such as Jira, Confluence, Bitbucket) with your existing permissions."
- Asana: an AI Teammate's access is "always bounded by the permissions of the person who triggers it."
That is the right security model for a company wiki. It is not a curated bundle for one job or one audience.
Guru deserves credit:
- A card can have a named verifier and a verification interval, so someone vouches for it on a schedule.
- People who follow a card are notified when it changes.
- That is exactly what durable knowledge needs in order to stay true, and it is on our list.
Cost: Guru does not publish prices; its pricing page asks you to talk to sales. Atlassian's MCP usage draws on Rovo credits from your organization's pool.
Prices, side by side
These are what we could verify on the vendors' own pricing pages on 2 October 2026, per person per month unless stated.
| Product |
What sharing with your team's AI costs |
| Claude |
Team plan, $20 annual / $25 monthly per seat |
| ChatGPT |
Project sharing on every plan, including Free (plan prices not verified) |
| Notion |
Free / $10 / $20 per member; agent features on Business ($20) |
| Google Workspace |
$7 / $14 / $22 per user; Gemini in Drive on Business Plus ($22) |
| Obsidian |
App free; Sync $4 annual / $5 monthly per user |
| Guru |
Not published |
| Warmhand |
Not published yet |
It is not only documents
Most of these tools store and retrieve files. That covers reference material well. It does not cover the two kinds of team knowledge that matter most to an AI:
- Rules: things the AI should always do in a given area ("quote in euros for EU customers").
- Skills: a procedure that should kick in when the moment comes ("when asked for a renewal quote, do these five steps").
A rule written in a Notion page is just text the AI might find if it happens to search for it. The products that do treat instructions as instructions keep them inside one vendor: project instructions in Claude or ChatGPT, CLAUDE.md for Claude Code, organization skills in Claude. Each gives you one block per project, shared inside one company.
What Warmhand does differently
A Warmhand pack is a small, deliberately chosen bundle. It can hold:
- memories,
- pages,
- uploaded documents,
- finished artifacts,
- rules,
- skills.
A team usually holds a few packs, each named for what it contains.
One pack works across tools. Claude, ChatGPT, Claude Code and Cursor connect to the same pack, over the same open standard (MCP).
You choose who gets what. You share a single pack, a whole group of packs, or a few chosen packs from a group, with anyone by email, inside or outside your company. Removing someone takes effect immediately.
Recipients do not need access to your files. A pack holds the copy you chose when you added it. The person who had access decided to share it, once. Nobody has to keep Drive or Notion permissions in sync with an AI tool.
Rules and skills are things the pack carries. When an AI connects, Warmhand's first response lists the rules in force and the skills that may fire, each skill with the situation it is for. Reference material stays out of the way until a question needs it.
Every item says where it came from. Each piece carries who added it and when, so an answer built on it can name the colleague who vouched for it.
Packs change on purpose. A pack changes when someone deliberately changes it, not every time a source document is edited, so what your team's AI relies on is the version someone chose.
You can switch a pack off for one tool. A pack can be on in Claude and off in ChatGPT, for example.
What we are building next
Going through every alternative showed us what teams will expect from a pack. These are on our list now:
- Change notifications. When a pack changes, the people who hold it, and their AI, should know what is new. Guru's follow-a-card notifications are the bar.
- Section search for team packs. We rebuilt search for our documentation packs to rank sections and return passages (the benchmark post). Team packs with long handbooks and documents get the same treatment next.
- Edit access for teammates. Inviting someone to help maintain a pack, not only to use it.
- Verification with a due date. A named person vouches for an item, and it comes back for review on a schedule, so a pack stays true as well as small.
What this post cannot tell you
- Documentation, not behaviour. Documentation says what a product is designed to do. It does not say how well the AI answers once your knowledge is in it.
- Prices and plans change fast. Google's Gems change in November 2026, for one.
- One researcher's reading. We read a lot of pages, and still could not load some.
Next in the series
Each upcoming post puts one of these claims to a measurement, using the same method as our documentation benchmark: questions written blind, a baseline frozen before any change, a held-out set run once.
- Wrong-document answers. The same team material in Drive or Notion, connected to Claude, against the same material in a pack. How often does the answer draw on a document that had nothing to do with the question?
- Long documents. The same handbooks in a Claude project, a ChatGPT shared project and a Warmhand pack: how often the answer arrives, and what it costs.
- Rules and skills. Are they actually followed? Project instructions against pack rules, across Claude, ChatGPT and Cursor.
- Getting a teammate set up. How long it takes a non-technical person to get a colleague's knowledge into their own AI, with each option.
If you share what your team knows with AI some other way, or think we got a product wrong, reply and tell us. We would rather be corrected now than measure the wrong thing.
