Agent Tools¶

Agent tools let AI coding agents interact with your development environment: reading files, running commands, searching code, and connecting to external services.

Tool categories:

Built-in tools: Read, Write, Edit, Bash, Grep, Glob
Shell tools through bash (git, npm, docker, curl)
MCP tools for external services requiring structure or authentication

Built-in Tools Comparison¶

Both Claude Code and Opencode provide similar core tools for file operations and shell access:

Capability	Claude Code	Opencode	Description
Read files	`Read`	`read`	View file contents
Write files	`Write`	`write`	Create or overwrite files
Edit files	`Edit`	`edit`	Modify existing files
Shell commands	`Bash`	`bash`	Execute terminal commands
Pattern search	`Grep`	`grep`	Search with regex
File discovery	`Glob`	`glob`	Find files by pattern

Additional Opencode Tools¶

Opencode provides extra built-in tools:

list - directory contents
lsp - language server integration (hover, go-to-definition)
patch - edit multiple locations at once
skill - run registered skills
todoread / todowrite - task tracking
webfetch - fetch URLs
question - prompt the user

Bash-First Philosophy¶

From Anthropic's best practices: the agent has access to your shell environment. You can build convenience scripts and functions for it like you would for yourself.

Bash is the primary tool layer. The agent inherits your shell environment with all your functions and aliases.

For reusable workflows, create skills that agents can invoke via slash commands. Prefer standard CLI tools (gh, git, npm) over MCP equivalents.

When to Use Which Tool¶

Task	Recommended Tool	Rationale
Run commands	`Bash`	Direct shell access
Git operations	`Bash`	Full git CLI available
Install packages	`Bash`	npm/pip/etc. work natively
Create files	`Write`	Transactional safety, prevents partial failures
Modify files	`Edit`	Atomic changes with precise targeting
Complex search	`Grep` + `Glob`	Purpose-built for pattern matching
External APIs	MCP	Only when auth/state needed

Custom Bash Tools¶

You can create shell functions that agents will discover and use:

# In ~/.bashrc or project .envrc
function project-test() {
    npm run test -- --coverage
}

function project-lint() {
    npm run lint -- --fix
}

For complex workflows, wrap these in skills so agents can invoke them via /project-test or /project-lint.

Tool Selection Guide¶

Task Type	First Choice	Fallback
Run tests	`Bash`	—
Format code	`Bash`	—
Search codebase	`Grep` / `Glob`	`Bash` with rg/fd
Read file	`Read`	`Bash` with cat
Edit file	`Edit`	`Write` (full replacement)
Create file	`Write`	`Bash` with cat/echo
Git commit	`Bash`	—
API calls (no auth)	`Bash` with curl	MCP
Database queries	MCP	`Bash` with psql/mysql
JIRA/Linear tickets	MCP	—

Best Practices¶

Start with Bash. Move to built-in tools when you need better error handling, and MCP only when you need the constraints it provides.

Keep MCP servers to a minimum since each one consumes context tokens. Wrap reusable workflows in skills. Give subagents only the tools they need.

Runtime Support¶

Runtime	Built-in Tools	MCP Support	Documentation
Claude Code	✓	✓	tools
Opencode	✓	✓	tools
Cursor	✓	✓	tools
Cline	✓	✓	tools
Copilot	Limited	✗	—