Tools: What Bolt Agents Can Actually Do - Bolt

Agents in Bolt do work by calling tools. Every action you see in the TUI (reading a file, running tests, searching the web) is the model choosing a tool and Bolt executing it. Each agent has a declared access level that determines which tools it may call without asking you first (see Permissions).

Built-in tools

Tool	What it does
`read`	Read a file from disk.
`write`	Create a new file or overwrite an existing one.
`edit`	Apply a targeted edit (find/replace) to a file.
`apply_patch`	Apply a unified-diff style patch across one or more files.
`glob`	List files matching a glob pattern.
`grep`	Search file contents by regex/text.
`bash`	Run a shell command.
`task`	Delegate a subtask to a subagent (`@general`, `@explore`, or a custom subagent).
`todowrite`	Maintain the agent’s structured task list for multi-step work.
`webfetch`	Fetch a URL and return its contents.
`websearch`	Search the web.
`skill`	Invoke a named skill (see Skills).
`memory_save`	Persist a fact to project memory.
`memory_recall`	Look up a previously-saved memory.
`question`	Ask you a clarifying question and wait for a reply.
`bash_background`	Start a shell command as a managed background process (dev servers, watchers, long builds).
`process_output`	Read accumulated output and status for a background process, optionally waiting for it to finish.
`list_processes`	List background processes with their id, status, and command.
`kill_process`	Terminate a background process (and its process tree) by id.
`test_run`	Detect and run the project’s tests, returning structured pass/fail results and failure details.

Experimental

Tool	What it does
`lsp`	Query a language server for definitions, references, and diagnostics.
`execute`	Run code in an isolated code-mode sandbox.
`plan`	Produce a structured plan before making changes.

Experimental tools may change or be renamed between releases.

Extending the tool set

Beyond built-ins, Bolt can call tools from three sources:

Custom tools in .opencode/tool/ (or .opencode/tools/) at the project root. See Custom Tools.
Plugin tools contributed by any installed plugin.
MCP tools from any MCP server configured for the project.

All three feel identical from the agent’s point of view: they show up in the tool list, honor the agent’s access level, and appear in the TUI’s tool-call output.

Agents — which agents can call which tools.
Permissions — how tool access is gated per agent.
MCP Servers — the most common way to add real-world tools.