Cozy vs Heptabase for AI agent access

Updated · By the Cozy team

Heptabase is a visual thinking tool built around whiteboards and cards, with a remote MCP on paid plans. Cozy (cozy.cloud) is a Markdown wiki and file space with one hosted MCP endpoint per account, scoped agent access and no whiteboards. Choose Heptabase to think visually. Choose Cozy to give agents a governed, text-first wiki.

Cozy here means the wiki at cozy.cloud, operated by Coral Sourcing LLC. It is not affiliated with Cozy Cloud (cozy.io) or Cozi. We make Cozy, so read this with that in mind. Where Heptabase is better, we say so.

The short version

Heptabase is built for visual research. You lay cards out on whiteboards, link them, highlight PDFs and use an AI tutor. Its product page lists apps for web, Mac, Windows, Linux, iOS and Android, plus offline access and real-time collaboration.

Cozy is a private wiki and cloud file space for each person and their AI agents. Every account gets "My wiki", and you can create more private workspaces. It has a web app and a Mac app. It has no whiteboards, no mobile apps and no AI tutor. Cozy runs no model itself: you bring your own agent.

Side by side

AreaHeptabaseCozy
Core modelWhiteboards, cards, tag databasesMarkdown wiki and files
Agent accessRemote MCP, OAuth sign-inOne hosted MCP per account with OAuth, optional per-workspace endpoints with bearer
MCP plan gatePro, Premium, Premium+ or active trialNot documented here
SearchKeyword and semanticLexical only
Whiteboard toolsYesNone
Agent scopingWhole Space, no per-card exclusionsScopes and folder-prefix grants
AppsWeb, desktop, iOS, AndroidWeb and Mac
Native storageLocal database plus syncMarkdown, text, JSON, CSV files

How agents connect

Heptabase documents a remote MCP at https://api.heptabase.com/mcp with OAuth sign-in. According to its MCP help article, it needs a Pro, Premium or Premium+ subscription or an active trial, even for read-only connections. It offers keyword and semantic search, read, create and edit, and whiteboard tools. Some operations, such as whiteboard editing, tag-database querying, backlinks and renames, need the desktop or web app open and signed in.

Cozy gives each account one hosted, stateless Streamable HTTP endpoint at https://cozy.cloud/mcp, covering all your workspaces. Each workspace also has https://cozy.cloud/wiki-mcp/WORKSPACE_UUID for bearer credentials. It supports MCP protocol versions 2025-03-26, 2025-06-18 and 2025-11-25. Tools include cozy_context, cozy_graph, cozy_search, cozy_read_many, cozy_write, cozy_edit, cozy_move and cozy_restore. See the MCP server page for the full list.

In Codex:

codex mcp add cozy --url https://cozy.cloud/mcp

Codex opens a browser to sign in.

Setup guides exist for Claude Code and Codex.

Controlling what an agent can touch

Heptabase's help center says the AI agent searches the whole Space by default, with no per-card or per-whiteboard exclusions. It also says that telling the agent not to read a card is not enforced. The details are in its AI access article.

Cozy limits access at the grant. OAuth agents get no more than the scopes granted and your own access, and can be disconnected under Connectors → Agents. Bearer credentials, for scripts, are created in Agents, shown once, stored as SHA-256 digests, revocable, and expire within 90 days. Scopes are read, wiki-write and source-write, with optional folder-prefix grants, and never exceed your own access. Agents cannot see connector credentials or permanently purge.

Edits need an expectedRevision, so concurrent edits return a conflict instead of overwriting. Deleted files are tombstoned, and the last 50 revisions stay readable and restorable.

Neither product is end-to-end encrypted. Heptabase says its data on AWS is encrypted in transit and at rest, but not end to end, because search and AI read content server-side. Cozy also runs server-side on Cloudflare. Privacy is not the difference here. See privacy.

Data format and export

Heptabase exports notes as Markdown plus an All-Data.json backup from Settings, Backup. Its portability article warns that other apps may not keep whiteboard layouts, relationships or metadata.

In Cozy, Markdown, text, JSON and CSV files are the native storage. The graph is derived on every save from conventions in the files, such as frontmatter types, wikilinks and relation:: lines, so it travels with the text. Read more on typed links and the Markdown wiki. Migration is covered in the Markdown import guide.

Where Heptabase is stronger

Where Cozy fits

Choose which

Pick Heptabase if you think in diagrams, read many PDFs, or need mobile apps. Pick Cozy if the main reader and writer of your notes is an agent and you want to bound what it can do. Some people use both: Heptabase for visual research, then push finished notes into Cozy through the ingest API. For other tools, see the comparison index.

Frequently asked questions

Does Heptabase have an MCP server?

Yes. Heptabase documents a remote MCP at api.heptabase.com/mcp with OAuth sign-in, and it requires a Pro, Premium or Premium+ subscription or an active trial. Cozy also provides a hosted MCP, one URL per account; see the MCP server page.

Can I limit which Heptabase notes an AI agent reads?

Heptabase's help center says there are no per-card or per-whiteboard exclusions for its AI agent. In Cozy, credentials carry read, wiki-write or source-write scopes and optional folder-prefix grants.

Does Cozy have whiteboards?

No. Cozy is a text-first Markdown wiki with a typed link graph. If visual whiteboards are central to how you work, Heptabase is the better fit.

Can I move Heptabase notes into Cozy?

Heptabase exports notes as Markdown, and Cozy accepts Markdown, text, JSON and CSV files. Whiteboard layouts may not survive the export. See the Markdown import guide.

Give your agents a memory worth keeping

Cozy gives every account a private wiki that Claude, Codex, ChatGPT and other MCP clients can read and maintain.

Start your wiki