Docs / For owners
Connect Claude
Connect your assistant with a key or by signing in, and let it hear replies while you are away.
Checked against the product on 8 Nov 2026
What this page states, and how each is checked
- Apps such as claude.ai connect by signing in to Two Cans, with no key to paste.Automated test
- Each app you connect by signing in shows as its own connected app, and they share one inbox.Automated test
- Setup offers Claude Code, Claude Desktop, Cursor and Other (any MCP client).Setting in the code
- Two Cans says it is tested with Claude Code.On-screen text
- Claude Code is connected with one claude mcp add command that carries the /api/mcp address and your key.Setting in the code
- Claude Code sessions that were already open do not see new tools, so you open a new session.On-screen text
- The check question asks your assistant to list your Two Cans conversations.Setting in the code
- Setup shows the key as connected once your assistant has used it.Setting in the code
- While it waits, setup checks for the first call every 2 seconds.Setting in the code
- Setup shows the key once, already inside what you paste.On-screen text
- Keys start with ak_, and a revoked key stops working at once, along with its watch command.Automated test
- The keys list shows each key only as ak_, dots and its stored tail.Setting in the code
- The stored tail of a key is its last 4 characters, never the whole key.Automated test
- Setup has a Make a new key button.On-screen text
- The server offers exactly 16 tools, from start_conversation to keep_listening.End-to-end test
- Through request_approval, money, deleting and work outside the project always wait for you.Automated test
- The wait_for_replies tool waits up to 50 seconds per call; longer waits use the watch command.Automated test
- The background watch command waits up to 840 seconds (14 minutes).Automated test
- A working watch first prints "Waiting for replies" and ends with a line starting WAKE when there is something to read.Automated test
- Your assistant is never handed a command to run; keep_listening hands one over only when your assistant asks for the older listen command.Automated test
- The older listen command is one line that can only wait, with a listen token and never your key.Automated test
- When only you can unblock your assistant, need_owner moves the conversation to Needs you and emails you.Automated test
- keep_listening is for Claude Code only; other apps use wait_for_replies.Setting in the code
- keep_listening waits inside the tool call itself, so there is no command to run and nothing secret is handed to your assistant.Automated test
- Each keep_listening call listens for up to 13 minutes, then your assistant calls it again.Setting in the code
- keep_listening returns as soon as someone replies, and says what to do next.Automated test
- Keeping a chat awake longer does not restart listening; the call already running keeps going.Automated test
- Listening lasts an hour after the last message in the chat, or until the time you asked for if that is later.Automated test
- One session listens on a chat at a time; another session's command leaves that chat out.Automated test
- A newer keep_listening call from the same session takes over from the older one.Automated test
- Your assistant keeps a chat awake longer only with your own words, 7 days at most.Automated test
- The inbox has a Keep awake button.On-screen text
- When the listening time is over listening ends, and a guest who writes later goes to the owner.Automated test
- If listening stops, a guest left waiting goes to Needs you once it has been quiet for 90 s.Automated test
- A guest who writes while nothing is listening puts the chat in Needs you, you are emailed, and the guest is told the message is saved.Automated test
- A server restart ends an open wait_for_replies call with a line telling your assistant to wait 30 s and call it again.Automated test
- With a session label, the watch covers only the conversations that session started.Automated test
- An error that looks like an old tool list tells your assistant to reconnect through /mcp in Claude Code, or restart the session.Automated test
- Under Run it yourself (advanced) in Settings, a channel bridge that wakes Claude Code the moment a reply lands is offered as a preview.Setting in the code
Automated test: a test in the product must pass. On-screen text: the words must exist in the product. Setting in the code: the number is read from the code. The docs are published only when every check passes.
Tested with Claude only
Two Cans adds tools to your own Claude session, so your assistant can talk with guests. Guided setup makes a key, gives you one command to paste into Claude Code, and checks that it works.
Run guided setup
- Under Which agent do you use?, pick Claude Code.
- Select Create my key. The key is shown once, already inside the command.
- Paste the command into a terminal, then select I pasted it.
- Open a new Claude Code session. Sessions that were already open do not see new tools.
- Under Check it works, ask Claude to list your Two Cans conversations. Setup checks every 2 seconds and shows when the key has been used.
claude mcp add --transport http --scope user two-cans https://twocans.io/api/mcp --header "Authorization: Bearer ak_your_key"
Setup also lists Claude Desktop, Cursor and Other, but Two Cans is tested with Claude Code only.
Connect an app by signing in
Apps such as claude.ai connect by signing in to Two Cans, with no key to paste.
Each app you connect this way shows as its own connected app, and they share one inbox.
Look after your key
Keys start with ak_. Lost one before pasting it? Select Make a new key. Your keys in Settings shows each key only by its last 4 characters; Revoke stops it at once. See Security.
What your assistant can do
The server offers 16 tools:
- Talk: start a conversation (
start_conversation), send (send_message), read (read_conversation), list (list_conversations), change whose turn it is or end one (update_conversation). - Listen: keep listening in Claude Code (
keep_listening), or wait for replies (wait_for_replies). - Ask for you: when only you can unblock it (
need_owner), the conversation moves to Needs you and you get an email. - Ask first: ask before acting (
request_approval). Money, deleting and work outside the project always wait for you. - Show progress: working signals (
set_activity), boards (manage_items), tasks (manage_tasks). - Share: pages to look at (
publish_preview,manage_preview), emailing the link (email_link). - Interview: run an interview (
manage_interview).
How your assistant hears replies
In Claude Code, your assistant calls keep_listening. The call itself waits on Two Cans and returns as soon as someone replies, so there is no command to run and nothing secret in your terminal. Claude Code runs a long call in the background, so your assistant can keep working, and it works in auto mode with no extra settings.
- Each call listens for up to 13 minutes, then your assistant calls it again.
- It listens for an hour after the last message in the chat.
- One session listens on a chat at a time. A newer call from the same session takes over from the older one.
- To keep listening longer, tell your assistant, for example "keep this awake until Friday", or select Keep awake in your inbox. The most is 7 days. Listening that is already running keeps going. See Your inbox.
- When the time is up, listening ends. A guest who writes after that goes to your Needs you.
- If listening stops early, for example because the computer slept or the session closed, a waiting guest goes to Needs you once it has been quiet for 90 s. You get an email, and the guest is told their message is saved.
Other apps, such as claude.ai, use wait_for_replies instead:
- The
wait_for_repliestool waits up to 50 seconds per call. - Your assistant is never handed a command to run. The one exception is the older listen command, which your assistant gets only if it asks for it; it can only wait, and Claude Code's auto mode may block it unless you allow it in your settings.
- If you run the wait yourself, the command under Run it yourself (advanced) in Settings waits up to 840 seconds (14 minutes). It prints "Waiting for replies" first and ends with a line starting WAKE when there is something to read.
After each wake, Claude reads, answers and listens again. After a server restart an open wait_for_replies call ends with a line saying to wait 30 s and call it again. If several sessions share one key, give each a session label so its waits cover only its own conversations. Emails and presence shows how to tell it is listening.
Reconnect after an update
After an update, an open session can keep the old tool list, and Claude gets an error saying so. Type /mcp in Claude Code and reconnect the server, or start a new session.
Channel bridge (preview)
A channel bridge that wakes Claude Code without a watch sits under Run it yourself (advanced) in Settings. It is a preview.
Then open Your inbox.