Skip to content

Tools & Commands

The AI agent has a set of built-in tools it can invoke autonomously, plus any tools contributed by MCP servers.

Built-in Tools

Tool Description
read_file Read file contents with optional line offset and limit. Lines are numbered and long lines are truncated at 2000 characters.
write_file Create or overwrite a file. Parent directories are created automatically.
edit_file Exact string replacement in a file. Fails if the search string is ambiguous unless replace_all is set.
glob Find files matching a glob pattern (*, **, ?). Results sorted by modification time.
grep Search file contents by regular expression. Supports path and glob filters.
shell_execute Run a shell command. Output is truncated at 30 KB. Default timeout: 2 minutes.
endo_execute Evaluate Endo source code and capture its output. Same limits as shell_execute.
git Execute git operations. Read-only subcommands run automatically; destructive patterns are blocked.
explore Spawn a read-only sub-agent to research the codebase. The sub-agent's conversation is discarded after it returns.
submit_plan Submit a structured plan during plan mode exploration.
save_memory Persist a memory file to ~/.config/endo/agent-memory/. Memories are loaded into the system prompt on future sessions.
search Unified glob + grep search. Combines file pattern matching and content search in a single tool call.
web_search Search the web via DuckDuckGo, Brave, or Google. Configure in init.endo.
web_fetch Fetch a web page and return its content as readable text. HTML is converted to markdown, JSON is pretty-printed. Includes a 15-minute cache.
list_directory List directory contents with optional hidden file display and long format (sizes, dates, types).
ask_user Ask the user a clarifying question. Supports free-text input or 2–6 multiple-choice options.

MCP tools from configured servers appear alongside these built-in tools and are callable in the same way.

Slash Commands

Slash commands are typed directly in the agent prompt:

Command Description
/help List available commands
/tools List all active agent tools
/plan [query] Enter plan mode (also activated with Shift+Tab)
/rename <new-name> Rename the current session
/exit Leave agent mode (same as Ctrl+D)

Plan Mode

Plan mode lets the agent explore your codebase before making changes, producing a step-by-step plan for your review.

Entering Plan Mode

  • Type /plan followed by an optional query, or
  • Press Shift+Tab to cycle into plan mode.

How It Works

  1. Exploration -- the agent uses read-only tools (read_file, glob, grep, git) to understand the codebase. It has a budget of exploration turns (default: 15, configurable via plan_mode.max_exploration_turns).

  2. Plan submission -- the agent calls submit_plan with a structured plan containing:

    • A high-level summary
    • Ordered steps with descriptions, rationale, and file lists
    • Dependency graph between steps
    • Risk assessment and alternative approaches

  3. Review -- the plan is presented for your approval. You can accept, reject, or ask for revisions.

  4. Execution -- once approved, the agent carries out the plan step by step. If pause_between_steps is enabled in agent.yml, it pauses for confirmation after each step.

Agent Memory

The save_memory tool lets the agent persist notes across sessions. Memory files are stored in ~/.config/endo/agent-memory/ and automatically loaded into the system prompt when you next enter agent mode.

This is useful for recording project conventions, architectural decisions, or debugging insights that the agent should remember.

Further Reading