Tools

Tools available to TurboDev agents.

Agents have access to a set of tools for interacting with your codebase. Tool availability can be controlled per-agent via the tools configuration.

Reference

Tool	Description
bash	Execute shell commands
edit_file	Create or edit files
git	Execute Git operations
github	Execute GitHub operations
grep	Search file contents with regex
list_files	List directory contents
mkdir	Create directories
question	Ask the user a question
read_file	Read file contents
load_skill	Load skill instructions or resource files
save_memory	Save a durable fact to persistent memory
task	Invoke a subagent

read_file

Read the full content of a file.

Args: { path: string }

read_file({ "path": "src/index.ts" })

edit_file

Create or edit a file. If old_str is empty, creates the file with new_str. Otherwise, finds and replaces the first occurrence of old_str with new_str.

Args: { path: string, old_str: string, new_str: string }

edit_file({ "path": "src/app.ts", "old_str": "hello", "new_str": "world" })

list_files

List files and directories in a given path.

Args: { path?: string } (defaults to current directory)

list_files({ "path": "src" })

mkdir

Create a new directory, including parent directories.

Args: { path: string }

mkdir({ "path": "src/components/ui" })

grep

Search file contents using regular expressions.

Args: { pattern: string, include?: string, path?: string }

grep({ "pattern": "TODO", "include": "*.ts" })

bash

Execute a shell command and return output.

Args: { command: string, timeout?: number, workdir?: string }

bash({ "command": "npm test", "timeout": 60000 })

question

Ask the user a question and wait for their response.

Args: { question: string, options?: string[] }

question({ "question": "Which framework?", "options": ["react", "vue"] })

load_skill

Load full instructions or a specific resource file from an agent skill. The LLM calls this automatically when it determines a skill is relevant to the current task.

Args: { name: string, resource?: string }

load_skill({ "name": "react-component" })
load_skill({ "name": "react-component", "resource": "references/patterns.md" })

When resource is omitted, returns the full SKILL.md instructions. When provided, returns the content of the specified file within the skill directory.

TIP

This tool is always enabled and cannot be disabled per-agent. The LLM only has access to skill metadata in the system prompt until it calls load_skill.

save_memory

Save a durable fact to persistent memory (.turbodev/memory.md) for future sessions. The LLM calls this when it learns something worth remembering.

Args: { content: string, category?: string }

save_memory({ "content": "User prefers pnpm over npm", "category": "preferences" })
save_memory({ "content": "Project uses PostgreSQL 16" })

Categories: preferences, decisions, architecture, facts (default). Only save information relevant to future sessions — not transient task state.

TIP

This tool is always enabled and writes only to .turbodev/memory.md (gitignored). It cannot modify other files. Memory survives compaction because it lives in the system prompt.

See Persistent Memory for the full reference.

task

Invoke a subagent for a specialized task.

Args: { agent: string, prompt: string, description: string }

task({ "agent": "explore", "prompt": "find all API routes", "description": "Explore API routes" })

WARNING

The task tool is disabled for the plan agent by default.

git

Execute Git operations via simple-git. Supports 26 operations including status, log, diff, add, commit, push, pull, branch, stash, remote, and more.

Args: { operation: string, args?: any }

git({ "operation": "status" })
git({ "operation": "log", "args": { "maxCount": 10 } })
git({ "operation": "commit", "args": { "message": "fix: resolve bug" } })

Available operations: status, log, diff, add, commit, push, pull, branch, checkout, stash, remote, fetch, reset, revert, tag, merge, rebase, init, clone, addRemote, removeRemote, listRemotes, raw, diffSummary, show, clean.

DANGER

Git operations that modify the repository (commit, push, reset, revert, clean, etc.) require explicit permission from the user.

github

Execute GitHub operations via the gh CLI. Supports 15 operations including PR management, issues, releases, and authentication.

Args: { operation: string, args?: any }

github({ "operation": "createPr", "args": { "title": "Fix bug", "body": "Description" } })
github({ "operation": "listPrs" })
github({ "operation": "createIssue", "args": { "title": "Bug report", "body": "Steps to reproduce" } })

Available operations: authStatus, createPr, listPrs, viewPr, mergePr, closePr, createIssue, listIssues, closeIssue, createRelease, listReleases, repoView, repoClone, runList, runView.

Requires the GitHub CLI (gh) to be installed and authenticated. Run /gh auth to set up authentication.

Controlling Tool Access

Tools can be enabled or disabled per-agent in the Markdown configuration:

tools:
  edit_file: false
  bash: false
  task: false

When a tool is disabled (false), the agent receives an error if it tries to use it: Tool "edit_file" is denied for agent "plan".

MCP Tools (dynamic)

In addition to the native tools above, agents can invoke MCP tools registered dynamically from external MCP servers declared in .turbodev/mcp.json. These tools appear automatically in the system prompt when at least one server is connected.

MCP tool names follow the convention mcp__<server>__<tool> — for example, a filesystem server exposing a read_file tool is callable as:

tool: mcp__filesystem__read_file({"path": "/tmp/example.txt"})

MCP tools are not listed in the tools: config (they're dynamic). To disable a specific MCP tool, use its full prefixed name:

tools:
  "mcp__filesystem__delete": false

Permissions for MCP tools default to ask and are configured via permission.mcp. See Permissions and MCP configuration for details.

Learn more about permissions