What does an agent send when it logs hours?
A time tracker only needs a few details to create an entry. Anything extra is unnecessary risk. In Half5, here’s what’s sent:
| Data | Sent? | Purpose |
|---|---|---|
| Description | Yes, one line | What you did, in words a client could read |
| Date, duration, start and end time | When known | The entry itself |
| Client, project, activity | When you name them | Selecting the right client |
| Repository name | From coding assistants | Identifying the client; name only, never a local path |
| Branch and commit subjects | Optional, up to 10 | Extra hint for client and project |
| File paths, code, file contents | No | Not needed for a time entry |
The description can appear on an invoice. Make sure your agent’s project instructions say to write a single client-friendly line, without file or class names.
What does the AI provider see, and what does Half5 store?
These are two separate parties. Your assistant runs at an AI provider like Anthropic, OpenAI, or Cursor. The provider processes your full conversation: your request, the agent’s tool calls, and Half5’s responses, such as client names or a list of your entries. What the provider does with that is covered by its own terms and settings.
Half5 only sees what the agent sends in a tool call, not the rest of your conversation. It stores the entry itself, a note of which agent created it with the repository name, and your repository memory.
- At the AI provider: your conversation, including what Half5 sends back.
- At Half5: the entry (description, date, duration, client, project, activity), which agent created it, the repository name, and your repository memory.
- Nowhere: your password. You sign in to Half5 in your browser, not in the assistant.
What do you consent to?
Connecting uses OAuth: the agent opens a Half5 consent screen, you sign in, and pick the organization. The screen lists exactly what the agent can do. An agent can’t get more access than what’s listed.
- View hours: your own hours, and the clients, projects, and activities you’re allowed to see in that organization.
- Log hours: create new hours that still need approval, and change or delete your own unapproved hours.
- Not: view or change invoices, colleagues, or settings.
If you see ‘Unverified app’, make sure you just started this connection yourself. If you’re not sure, click Deny.
What does Half5 remember, and how do you make it forget?
Half5 remembers for each user which repository belongs to which client, so the agent doesn’t have to keep asking. That memory is tied to you and shown under Settings → ‘Linked agents’ → ‘Repository storage’. Use ‘Forget …’ to clear one repository, or ‘Forget all’ to clear everything.
Two organization settings under ‘AI agents’ control whether Half5 also learns from others. ‘Learn from colleagues’: if a colleague already logged a repository on a client, Half5 suggests that client, but the agent always asks first. ‘Learn from the desktop app’: agent work you register in the desktop app counts toward your own memory. Admins can turn these on or off.
How does this relate to your timeline?
The MCP connection is separate from the timeline in the Half5 apps. The desktop app can add your AI sessions in Claude Code, Codex, and Cursor to your timeline, alongside apps and documents. That timeline is end-to-end encrypted: only you can read it, Half5 can’t see it, and it isn’t shared with your organization. Meetings from your connected calendar are fetched by the server and visible only to you.
Your manager and admins can’t open your timeline; no screen or report in Half5 shows it to anyone else. They only see what you register: hours, leave, and rides. Productivity numbers in Half5 come from those registered hours, not from the timeline. Why that matters is explained in Productivity without surveillance; how the timeline works is in Automatic time tracking.
What should a team agree on?
A few clear agreements prevent most issues. They’re quick to set up and apply to any time tracker with an AI connection.
- Agree on which AI assistants may be used, and under which account—business or personal.
- Agree that descriptions are a single client-friendly line, with no code, paths, or confidential details.
- Only connect agents you started yourself, and deny any unverified app you don’t recognize.
- Always send agent-logged hours through approval, just like other hours. See Logging hours with AI.
- As an admin, decide whether ‘Learn from colleagues’ and ‘Learn from the desktop app’ should be on.
- Review ‘Linked agents’ from time to time and disconnect anything you’re not using.
How Half5 does it
Half5’s MCP server explicitly tells agents never to send file paths or code, and to write the description as a single line a client could read. Entries arrive as ‘Not approved yet’. Under Settings → ‘Linked agents’ you’ll see each connection’s name, whether it’s verified, when it was last used, what it can do, and you can revoke access with ‘Disconnect’.
How log_work uses the repository to pick a client is explained in How log_work works. Half5 doesn’t claim any certifications here; what it does and doesn’t do with your data is on the privacy and trust page.
Frequently asked questions
Does Half5 read my conversation with Claude or ChatGPT?
No. Half5 only gets what the agent sends in a tool call, like the description and duration. The conversation itself stays with the AI provider.
Can my manager see what I do with my AI assistant?
No. AI sessions are in your timeline, which is end-to-end encrypted and only you can read it. Your manager sees the hours you register, including those an agent logs for you.
How do I clear what Half5 knows about my repositories?
Go to Settings → ‘Linked agents’ → ‘Repository storage’. Use ‘Forget …’ to clear one repository or ‘Forget all’ to clear everything.
What does ‘Unverified app’ on the consent screen mean?
It means Half5 can’t confirm who made the client. It might be a regular assistant, but only allow it if you just started the connection yourself.
Does an agent send my source code to Half5?
It shouldn’t, and the server specifically tells agents not to. At most, the repository name, branch, and a few commit subjects are sent as hints for the client.
