Stream MCP lets you bring your notes and interactions into AI apps like Claude and ChatGPT, and bring context from those apps back into Stream.
MCP (Model Context Protocol) is an open standard that lets AI apps communicate with other tools.
What you need
- Stream on iOS
- An AI app that supports MCP, like Claude or ChatGPT
Set it up
- In Stream, go to Settings → MCP
- Pick Claude, ChatGPT or Other Agents
- Tap Copy Prompt
- Open a new chat in your AI app and paste the prompt
- Follow the steps, then sign in to Stream and approve
Once connected, your AI app may offer to make its own folder in Stream for chat summaries, and give you a short instruction to save in its settings so it keeps using Stream in future chats.
Try this
Chart your progress
Log workouts or meals with Stream, then have your AI app chart them.
Read my Stream folder called [Workouts] and chart my progress over the last month. Point out what's improving and what I can try next..
Draft from an idea
Turn something you said on a walk into a first draft.
Find my recent note about [idea] and draft a [one-page doc] from them. Keep my wording where you can.
Start your morning
Set up a routine in Claude or ChatGPT that works through your to-do list every morning.
Set up a daily routine: every weekday at 8am, check my Stream to-do list, pick the three most important items, and start on them. Tell me what you did and what needs me.
Catch Stream up
Save a summary of your day to Stream, then keep talking it through on your commute.
Summarize what happened today, including this chat and my calendar, and save it to Stream so I can pick it up on my commute.
FAQ
What can my AI app do in Stream?
Search and read your notes, lists, reminders and past conversations. Save new notes, make folders and lists, and move notes into folders. Add, edit, check off or delete list items, and set reminders. It can't change the text of a note you already have, and it can't delete a note, list or folder. It only looks in Stream after you approve it.
Will my AI app ask before saving to Stream?
Each AI app has its own setting for using MCP. Some support “Ask Once” or “Ask Always” commands.
Why can't my AI app save notes?
Disconnect Stream and connect it again.
Why is my note too long?
A note can be up to 1,000 words, including its title, and a list item up to 200. Ask your AI app to split it.
How do I disconnect?
Remove Stream from your AI app's connector settings.
How long does the connection last?
90 days, as long as you use it. If you don't use it for 30 days, you'll need to approve Stream again in your AI app.
Why doesn't Stream show up in my chat?
Start a new chat and try again.
Why can't I find a new note in search?
It's saved. Search can take a few seconds to catch up.
Still stuck? See Get support.
Related
- Set up Stream Webhook: send every interaction to your own tools as it's saved.
- What is Universal Dictation
For developers
Connection
| Setting | Value |
|---|---|
| Server URL | https://api.sandbar.com/mcp |
| Transport | Streamable HTTP, stateless, JSON responses |
| Sign-in | OAuth with dynamic client registration. The user signs in to Stream and approves |
| Scopes |
stream (read) and stream.write (write). New connections get both |
| Access token | 1 hour |
| Refresh token | 30 days, renewed with use. The connection ends 90 days after the user approves |
Tools
| Tool | Type | What it does |
|---|---|---|
get_guide |
Read | Returns Stream's guide for agents: how notes, lists and folders work and how to write each. Also available as the resource stream://guide
|
search |
Read | Searches notes, lists, reminders, conversations and interactions by meaning and keyword. Returns matching notes and lists in full. Up to 20 results per category |
get_recent_interactions |
Read | What the user has said to Stream lately, newest first. Last 7 days by default, up to 100 |
list_artifacts |
Read | Browses notes, lists and reminders, most recently updated first. Filter by type or folder. Up to 500 per page; page with before and nextBefore
|
get_artifact |
Read | Opens one note, list or reminder in full |
list_folders |
Read | Lists the user's folders and their instructions |
create_note |
Write | Creates a note, optionally in a folder |
create_folder |
Write | Creates a folder. If the name is taken, returns the existing one |
create_list |
Write | Creates an empty list. If the name is taken, returns the existing one |
set_note_folder |
Changes existing | Moves a note into a folder, or out of one |
update_folder |
Changes existing | Renames a folder or replaces its instruction |
update_list |
Changes existing | Renames a list or replaces its instruction. Items are untouched |
update_list_item |
Changes existing | Adds, edits, completes, reopens or deletes a list item, and sets or clears its reminder |
Each tool carries MCP annotations. Read tools are marked read-only. Changes existing tools are marked destructive, so most AI apps ask before running them. No tool deletes a note, list or folder, or edits the text of an existing note.
Limits
| Limit | Value |
|---|---|
| Note | 1,000 words, including the title |
| List item | 200 words |
| Folder and list names | 30 characters. Instructions 140 |
| Folders | 25 active per user |
| Search | 1 per second per user, bursts of 8 |
| Writes | 1 per second per user, bursts of 5, shared by all write tools |
| Reminder times | Unix seconds, not milliseconds |
Writing safely
-
set_note_folder,update_listandupdate_list_itemtakeexpected_last_update_ts. Pass thelastUpdateUnixTsfrom your last read. If the note or list changed since, you get a conflict with the current version. Read it again and retry. - A new note can come back with
searchIndexStatus: "pending". It's saved but not searchable yet. Don't create it again. - Writes are recorded as "Created via Stream MCP" or "Updated via Stream MCP".
Errors
Tool errors come back as a normal result with a JSON body like {"error": "..."}, written so the agent can relay it.
| Error | What to do |
|---|---|
| Connection is read-only | Remove Stream and connect again to get write access |
| Rate limited | Wait the number of seconds in the message, then retry |
| Capture changed since you read it | Use the artifact in the response and retry with its lastUpdateUnixTs
|
| Too large | Shorten it or split it |
| Not found | Check the id from search, list_artifacts or list_folders
|
What won't change
- Tool names stay stable. Renamed tools keep answering to their old names.
- New tools and fields may be added. Ignore ones you don't recognize.