ClawLink
TL;DR
| Question | Answer |
|---|---|
| What does it do? | Your agent fetches today's unread Gmail through ClawLink, writes a five-bullet brief, and saves draft replies to the urgent messages in your Drafts folder. |
| Who is it for? | Anyone who opens their inbox to 30 unread emails and wants the five that matter first. |
| What do I connect? | Gmail, once, with its own sign-in. ClawLink never sees your password. |
| What does it cost? | About 2 tool calls on your plan. Starter includes 1,000 a month. |
| Does it work? | gmail_fetch_emails succeeded on 96.5% of 14,613 calls on ClawLink in the last 30 days. |
| Which agents? | OpenClaw, Hermes, Claude, Codex, ChatGPT and any MCP client connected to ClawLink. |
How do you try it?
- Set up ClawLink once. Add it to your agent from the install page, then connect Gmail.
- Send this prompt to your agent. Click the card to copy it.
It works the same in OpenClaw, Hermes, Claude and Codex.
How does the agent do it?
The agent makes 2 calls through ClawLink. You do not pick the tools: the agent finds them by searching ClawLink, reads their input schema and runs them.
| Step | Tool | What it does | Calls in 30 days |
|---|---|---|---|
| 1 | gmail_fetch_emails | Searches your mailbox with a Gmail query such as is:unread newer_than:1d and returns the messages. | 14,613 (96.5% ok) |
| 2 | gmail_create_email_draft | Saves a reply as a draft in your Gmail. It does not send. | 154 (100.0% ok) |
How often does this work on ClawLink?
We do not publish a sample run for this job: its output is somebody's own Gmail. Instead, here is how the same tools performed for every ClawLink user from 2026-09-03 to 2026-10-03. A call counts as failed only when it reached Gmail and came back with an error.
| Tool | Calls | Users | Succeeded | Average time |
|---|---|---|---|---|
gmail_fetch_emails | 14,613 | 109 | 96.5% | 3.9 s |
gmail_create_email_draft | 154 | 16 | 100.0% | 2.6 s |
How do you write the prompt?
- Say "today" or a time window. The agent turns it into a Gmail query (newer_than:1d). Without a window it may read hundreds of old unread emails.
- Define urgent. "Urgent means from a client or about money" gives better drafts than letting the agent guess.
- Keep drafts, not sends. Drafting is safe to automate. Sending on your behalf is a write the agent should ask you about first.
What can go wrong?
- Too many results at once. Asking for a very large page of emails makes slow, oversized replies. Most agents fetch 10 to 25 at a time and page through.
- Placeholder message ids. An agent sometimes invents an id like "message_id_here" for a reply. ClawLink rejects obvious placeholders before they reach Gmail; if you see that error, ask the agent to fetch the email again.
- Wrong tool names. Agents often guess names like gmail_list_messages. The real tool is GMAIL_FETCH_EMAILS; ClawLink suggests it when the guess fails.
How does this compare with other ways to do it?
| Option | How it works | Cost | Best for |
|---|---|---|---|
| ClawLink + your agent | One prompt, the brief and drafts arrive in your agent's chat and in Gmail | 2+ tool calls on your plan | A daily brief inside OpenClaw, Hermes or Claude |
| Gemini in Gmail | Summaries inside Gmail, one thread at a time | Included with some Google Workspace plans | Reading one thread inside Gmail |
| Gmail API yourself | Google Cloud project, OAuth consent, code | Free API, your time | Developers who want full control |
Frequently asked questions
Can the agent send emails?
It can, with the GMAIL_SEND_EMAIL tool, but this prompt only creates drafts. ClawLink treats sends as writes that agents should confirm with you.
Does ClawLink store my emails?
No. The email text passes through to your agent for that call. ClawLink logs that a tool ran, not your mailbox.
Does it work with Google Workspace accounts?
Yes, personal Gmail and Workspace accounts both connect with Google sign-in.
What does it cost?
Each tool call counts once against your plan: 1,000 calls a month on Starter. A brief with three drafts is about four calls.
Can it run every morning?
Yes. Ask your agent to run the prompt on a schedule; OpenClaw and Hermes both support recurring jobs.