Skip to main content

Tools

Tools are what agents can actually do during a run — read files, run commands, browse the web, delegate to children. You choose each agent's tool set in Settings → Tools, per tool or with the quick-select buttons:

  • All — everything on. The default, and the right choice for most agents — an agent that can't write files can't fix its own mistakes.
  • Read-only — only tools that don't modify anything (Read, Glob, Grep, TOC, WebSearch, WebFetch, Notify). Good for review and research agents.
  • None — everything off.

Two tools ignore the checkboxes: Pause is always available so the agent can always ask for input, and Message is always available in delegated/subrun contexts so a child can always report progress. One tool lives elsewhere entirely: Admin is controlled by the Full system access toggle in Settings → Advanced, not by the Tools tab.

File and search tools​

  • Read — read a file (text or images), with line numbers.
  • Write — create a file, or overwrite one completely.
  • Edit — targeted search-and-replace in an existing file. The main tool for code changes.
  • ApplyPatch — apply a diff-style patch to one file; better than Edit for sprawling multi-line changes.
  • TOC — outline of a file's headings and symbols, for navigating big files before reading.
  • Glob — find files by pattern (src/**/*.ts).
  • Grep — regex search across file contents.
  • Sync — copy files between the container and remote targets. Only relevant with target grants.

Bash​

Runs shell commands in a persistent session — the working directory and environment carry over between calls. This is the workhorse: builds, tests, git, package installs, everything command-line.

Web tools​

  • WebSearch — search the web, get titles, URLs, and snippets.
  • WebFetch — fetch a URL as readable text.
  • Browser — a headless browser the agent can drive: navigate, click, type, screenshot, run JavaScript. For pages where fetching isn't enough — JavaScript rendering, forms, multi-step flows.

Control tools​

These are how agents manage their own workflow:

  • Pause — stop and wait for input: from you if the agent is top-level, from the parent if it was delegated. Always available.
  • Message — send a progress update to the parent run without finishing ("phase 1 done, starting phase 2"). Always available to children and subruns.
  • Notify — push a notification to the Taurus app: work finished, blocker hit, something failed.
  • Subrun — spin up a subtask within the same agent, in a separate context window. See Subruns.
  • Inspect — read run history: the agent's own past runs, or a child's.
  • Wait — wait for background subruns/delegations to finish, or just sleep.

Multi-agent tools​

  • Delegate — send a task to a child agent, blocking or in the background. Can also resume one of the child's earlier runs to iterate on existing work.
  • Supervisor — direct management of children: list them, create/update/delete them, send messages into their runs, stop their runs.
  • Admin — the account-wide version of Supervisor: manages any agent in your account, not just children. Not in the Tools list — it comes from the Full system access toggle and requires confirmation. See System access.

See Multi-Agent Collaboration for how these fit together.

Tools on remote targets​

When an agent has target grants, the file, search, and shell tools (Bash, Read, Write, Edit, ApplyPatch, Glob, Grep, TOC, Sync) accept a target parameter — omit it and the tool runs in the agent's container, set it and the tool runs on that remote machine.

What the agent may do remotely is controlled by the grant:

  1. Operation permissions — Shell, Read, Write, Sync toggles.
  2. Path policy — path rules that constrain file operations (exact paths and descendant prefixes, not globs).

The safety distinction to remember: path policy does not restrict Shell. If a grant allows remote Bash, the agent acts with the full privileges of the underlying SSH or relay account, path rules or not. Grant Shell deliberately.