Tool Execution

Cluster Build can execute tools on your behalf — reading files, running commands, and editing code.

Tool Execution

The Cluster Build agent can call a set of built-in tools to interact with your system. Every destructive tool requires your approval by default — the CLI shows you exactly what it wants to do and you choose whether to allow it.

Available Tools

File operations

ToolDescription
readRead the contents of a file
writeWrite a file (create or overwrite)
editTargeted search-and-replace edits in a file
multieditBatch of edits applied atomically to one file
lsList files in a directory
ToolDescription
grepContent search across files (powered by ripgrep)
globFile pattern matching (e.g. **/*.ts)

Shell & jobs

ToolDescription
bashRun shell commands. Long-running commands can be launched as background jobs
job_outputFetch output from a background job
job_killTerminate a background job

Web

ToolDescription
web_searchSearch the web
web_fetchFetch and extract content from a URL
downloadDownload a file from a URL

Developer tools

ToolDescription
lsp_diagnosticsRead LSP diagnostics for the current project
lsp_referencesFind references to a symbol via LSP
lsp_restartRestart the language server
code_sandboxExecute code in an isolated sandbox

Browser (headless)

When the browser is enabled the agent can drive a headless browser with screenshot, click, type, evaluate, and content extraction actions.

Interaction

ToolDescription
ask_user_questionPause and ask you a clarifying question

Approval Flow

By default, the agent asks for confirmation before running any tool that modifies your system (write, edit, bash, etc.). Read-only tools like read, grep, and glob run automatically.

To skip approvals entirely — for example, during a long unattended task — pass the --dangerously-allow-all flag when starting the session:

cluster --dangerously-allow-all

Use this with care. The agent can run any bash command and edit any file you have access to.

On this page