Skip to content
Thrum

See what your agents did

Free and Pro

Thrum doesn’t stop your agents from acting. Instead, it keeps a record on your machine of what they did, so you can look back at any time.

What it looks like

Section titled “What it looks like”

Your coordinator is the agent you talk to, which starts the other agents, hands out work and reports back to you. Ask it, and it searches the record for you. This is a shortened example. Your coordinator’s wording will differ.

You:         What Thrum commands did orch_demo run today?
Coordinator: The Thrum commands orch_demo ran today, in order (all
             2026-10-31 UTC):
             23:24:45 thrum inbox --unread
             23:24:48 thrum message get … (the plan handover)
             23:25:01 thrum tmux create crl-check --name impl_crl …
                      (launched its implementer)
             23:25:43 thrum send --to @impl_crl (the epic dispatch)
             …
             23:26:35 thrum send --to @coord_ca_for_labs (final report)

Two logs

Section titled “Two logs”
  • The command log records every thrum command run on this machine, by you or by any agent, on every AI tool: the command, who ran it, in which folder, when, how long it took and whether it worked. It doesn’t record the command’s output, or anything an agent did outside Thrum, such as running Git or editing a file.
  • The hook log records the tool calls your agents make on Claude Code, Codex, Cursor, OpenCode, Copilot and Muse, such as shell commands, file edits and file reads, with what went in and what came back (the limits below say how much is kept). Each AI tool reports its calls to Thrum through a hook, a script it runs at set moments, which gives the log its name. Antigravity agents aren’t in the hook log yet, but their Thrum commands are in the command log.

Both are always on. There is nothing to set up.

Ask what an agent did

Section titled “Ask what an agent did”

Ask your coordinator, in its terminal or in My Inbox, your inbox in Thrum’s web UI, the page you open in your browser to see your agents and their messages. For example:

What Thrum commands did impl_api run today?
What shell commands did impl_api run today?

The second question uses the hook log, so it works for agents on every supported AI tool except Antigravity.

Thrum starts a new log file about once a day. Older files are compressed and kept for a week, and your coordinator can search those too: ask, for example, What did impl_api run yesterday?

The logs keep times in UTC, so “today” and “yesterday” can mean UTC days. If it matters, say which time zone you mean.

Each machine keeps its own logs. To ask about an agent on another machine, ask the coordinator on that machine.

What is kept, and where

Section titled “What is kept, and where”
  • The logs are files in your repository’s .thrum/var/log/ folder, and only your user account can read them. Every worktree, a separate checkout of your repository on its own branch, writes to the same folder.
  • Thrum never sends them anywhere: they don’t go through sync, pairing or Git. A backup made with thrum backup includes them, so they go wherever your backups go.
  • Before it writes the hook log, Thrum hides anything that looks like a secret: tokens, passwords, keys, and anything named like one, such as API_TOKEN=…. For example, the shell command export API_TOKEN=abc123456789xyz is kept as export <REDACTED>. A secret that doesn’t look like one, such as a passphrase in plain words, is kept as written.
  • In the command log, the text of messages you send with a Thrum command is hidden. If the same text is inside a shell command, the hook log keeps it.
  • In the hook log, what went in and what came back are each cut at 4 KB.
  • In the hook log, a tool call that was refused, interrupted or crashed can show what went in with nothing coming back. That is normal.
  • On Copilot, the hook log can’t match the start of a call to its end, and in some setups, such as agents working in a linked worktree, each call is logged twice. On Muse, a shell command that fails logs what went in and nothing more.

What the logs are not

Section titled “What the logs are not”
  • Not tamper-proof. Any program running as your user account, including your agents, can read, change or delete them. Treat them as a record, not as proof.
  • Not complete after a crash. If your machine stops suddenly, the last few lines can be missing.
  • Not a transcript. They don’t keep what you type to an agent or what it says back.

Free and Pro

Section titled “Free and Pro”

Both logs, and searching them, are Free.

Next steps

Section titled “Next steps”