Connect Cobrain to Claude Code: Dream cleanup and note recovery

Identify the product, plugin and update

Developed by OverPress Media, Cobrain is a third-party service independent of model vendors. It stores background information, decisions and notes as Markdown for authorized compatible AI clients to read or update through MCP. Changes to a shared note are also visible to other authorized clients.

Official release notes list Dream and Trash under product 1.14 on October 3, 2026, and targeted edits and note encryption under 1.15. The Claude plugin reached 1.0.6 the same day. The product originally launched on August 26; this is a feature update. ChatGPT plugin 1.1's @ mentions and sidebar panel remain under review, so documentation does not establish availability for every account.

Install in Claude Code, then authorize the connection

Prepare Claude Code with existing access and a Cobrain account. The author says a free Cobrain account can connect; this does not include Claude Code's own model subscription or usage. Create a practice folder for first use.

Enter /plugin marketplace add Overpress-Media/cobrain-plugin-for-claude and then /plugin install cobrain@cobrain in the Claude Code conversation input. Both are client slash commands, not shell commands for a system terminal.

The plugin's remote address is https://cobrain.space/mcp. First use opens a Cobrain sign-in and authorization page. Check the account and read/write permissions before approving. It uses OAuth, so you do not need to paste a model API key into Cobrain. If installation succeeds but tools are absent, find cobrain in /mcp, finish authentication and start a new session.

Confirm the connection with a practice note

Create a folder such as Trial project and add fictional information such as Presentation review is tentatively scheduled for Wednesday. Ask Claude to find and read the note, then explicitly append Color scheme still to be discussed. Return to Cobrain and check that the original sentence remains and the addition and editor are correct. This is a suggested acceptance example, not a test performed for this article.

If reads work but writes do not, check the signed-in account and branch permissions. Shared read-only content does not become writable because an AI accesses it. If tools are entirely missing, check the connector. If a tool returns an authentication error, follow the author's disconnect/reconnect guidance using the same email. Preserve the error before repeatedly issuing the same write request.

Choose a time range or folder before a manual scan

In Cobrain Brain settings, select a lookback of 1–15 days and click Dream now. It searches for candidate issues in notes changed during that period. To inspect older material, right-click a particular folder and choose Dream now. This path ignores note age but processes at most 300 notes; use subfolders for larger collections.

The preliminary scan does not call AI or change notes. It shows results and a prompt to pass to AI on the Dream page. Similar notes may serve different purposes, and broken links may point to material not yet created. Do not treat the results as a deletion list.

For the first scan, use only the practice folder and check target paths, dates and matching notes. If unrelated project content appears, narrow the scope before continuing. Do not use the whole team knowledge base for a first cleanup. Dream is included in the free plan, but AI reading, analysis and rewriting still consume the connected AI's allowance.

Handle duplicates and targeted edits separately after involving AI

After reviewing the scan, pass the Dream page's prompt to the connected Claude client. The author's default process lets AI first handle items it considers safe, then ask about potentially lossy changes. Add an instruction such as: List the notes, reasoning and proposed change for each item first; do not merge or delete yet, and wait for my numbered confirmation. This instruction is not system-enforced per-item approval, so still inspect actual tool calls and results.

If two records say Wednesday review and Thursday review, determine whether the latter supersedes the former. Merging merely because text is similar can lose the decision timeline. Confirm a few items at a time and reread affected notes before the next group. A reply saying cleanup is complete is not full acceptance.

For a single-sentence edit, the plugin directs AI to read_file first, then supply exact old_text and new_text to patch_file. The old text must occur exactly once. A missing or repeated match is rejected, and the tool does not create a new note. Reread and select a longer unique original fragment rather than replacing the whole note to bypass rejection. Use append_file for additions.

Let the owner restore deleted notes instead of asking AI to reconstruct them

delete_file moves a note, attachment or folder into that brain's Trash. The brain owner can restore it within 30 days; AI cannot view or restore Trash. After an accidental deletion, stop the related cleanup and have the owner check names and paths in the website's Trash, then restore and inspect the content.

Do not ask AI to reconstruct a same-named note from conversation memory, which can create an incomplete substitute. Emptying Trash removes its contents immediately, ending reliance on the 30-day window. For incorrect edits without deletion, check note version history. The free plan keeps up to two versions for three days, a different retention period from Trash.

What sharing and encryption each protect

Cobrain note encryption is not end-to-end encryption. The service application decrypts bodies for users and authorized AI. Metadata such as note names, folder names and tags remains plaintext, and attachments use separate storage-provider encryption. How a connected AI handles received plaintext also depends on that platform's terms and account settings.

The official privacy statement also says backups created before body encryption completed on October 3 still contain plaintext notes and will expire on October 12. Encrypted does not mean all historical copies have already been converted.

Use fictional data for the first connection, confirm recipients and accounts, then add real work context. To stop an AI's future access, revoke authorization on Cobrain's Connections page. Revocation blocks subsequent tool calls; it does not establish that previously read content has been removed from that AI's chats or records.