Coding Agent — 22 Tools

The Coding Agent is Mode 2 of HelixMind and the core experience. It is your AI pair programmer with 22 built-in tools for reading, writing, searching, testing, and managing code.

Unlike other AI tools that forget everything between sessions, HelixMind has persistent spiral memory. It remembers your codebase, your patterns, and your decisions across every session. Ask it to refactor a module today, and next week it still knows why you chose that architecture.

Use it in your terminal with hx or from the web console. Same power, your choice.

Where does the Agent run?

Environment	How to start	What you get
CLI Terminal	Install with npm install -g helixmind, then run hx	Full interactive chat with all 22 tools. This is the primary experience.
Web Console	Open the Console tab in your HelixMind dashboard	Same agent, same tools, streamed to your browser. Start sessions with /auto or /security, or use the interactive console.

Both connect to the same spiral memory and project context. Work in the terminal at home, check progress from the web console on your phone. Everything stays in sync.

The 22 Tools

File Operations (4 tools)

The foundation of every coding session. The agent reads your code, writes new files, and makes surgical edits without rewriting entire files.

Tool	What it does
read_file	Reads any file in your project. The agent uses this constantly to understand your code before making changes.
write_file	Creates new files or overwrites existing ones. Used for new components, configs, tests.
edit_file	Surgically edits specific lines in a file. No full rewrites needed. just the lines that matter.
list_directory	Explores your project structure recursively. The agent maps out your codebase before diving into details.

Search and Discovery (2 tools)

Finding the right code in a large project is half the battle. These tools make it instant.

Tool	What it does
search_files	Searches file contents with regex. Finds patterns, function calls, imports, and references across your entire codebase.
find_files	Finds files by name pattern (glob). Need all test files? All TypeScript configs? Located instantly.

Command Execution (1 tool)

Tool	What it does
run_command	Runs any shell command: npm test, tsc, eslint, python, custom scripts, build pipelines.

Commands are automatically classified for safety. Running ls or npm test happens instantly. Running npm install asks for your OK. Running rm -rf always requires explicit approval. See the Permission System below for details.

Git Operations (4 tools)

The agent manages your version control so you can focus on the code. It checks status before committing, reviews diffs to understand what changed, and writes meaningful commit messages.

Tool	What it does
git_status	Shows staged, unstaged, and untracked files. The agent checks this before every commit.
git_diff	Shows exactly what changed, line by line. Used to verify edits before committing.
git_commit	Creates commits with meaningful messages. Can auto-stage relevant files.
git_log	Reviews commit history. The agent uses this to understand recent changes and conventions.

Spiral Memory (4 tools)

This is what makes HelixMind fundamentally different. The agent can read and write to your project's persistent memory. a 6-level knowledge graph that evolves over time.

Tool	What it does
spiral_search	Semantic search across all memory levels. Finds concepts, not just keywords. Ask about "authentication flow" and it finds relevant nodes even if they use different words.
spiral_store	Stores new knowledge with auto-generated vector embeddings. Decisions, patterns, architecture choices. all persisted.
spiral_status	Shows memory statistics per level: how many nodes, total size, last activity.
spiral_compact	Triggers memory evolution. Frequently accessed knowledge gets promoted. Unused nodes decay. Old context gets compressed into summaries.

Bug Tracking (2 tools)

The agent maintains a persistent bug journal across sessions. Bugs are never forgotten. even if you switch tasks and come back weeks later.

Tool	What it does
bug_report	Reports a bug to the persistent journal with description, context, and reproduction steps.
bug_list	Lists all tracked bugs across sessions. Filter by status, priority, or search terms.

Web Research (1 tool)

Tool	What it does
web_research	Searches the web via DuckDuckGo, fetches pages, extracts code examples and key points, then stores everything in spiral memory as L6 Web Knowledge nodes. The agent learns from the internet in real-time.

When the agent encounters an unfamiliar API, framework, or error message, it can research it on the spot. The extracted knowledge persists in your spiral memory for future sessions.

Browser Automation (5 tools)

The agent can open, see, and interact with web pages. Test your frontend, fill out forms, take screenshots for visual analysis, or automate repetitive browser tasks.

Tool	What it does
browser_open	Opens URLs in headless or headful Chrome. Headful mode lets you watch it work.
browser_screenshot	Takes screenshots for visual analysis. The agent can see layout issues, broken UIs, and design problems.
browser_click	Clicks elements on web pages by selector or coordinates.
browser_type	Types into input fields, search boxes, forms.
browser_close	Closes the browser session and releases resources.

Permission System

Every tool call goes through a permission check. Three levels protect you from unintended changes.

Level	Behavior	Example tools
Safe	Runs without asking	read_file, search_files, git_status, spiral_search
Ask	Prompts for your approval	write_file, edit_file, git_commit, run_command
Dangerous	Always requires explicit approval	Destructive commands like rm -rf, DROP TABLE, chmod 777

Permission Modes

Choose your comfort level:

• Normal mode (default): Safe operations auto-run. Ask operations prompt you. Dangerous operations always prompt.
• Skip Permissions (hx -s): Safe and Ask operations auto-run. Dangerous operations still prompt. Good for when you trust the agent but want a safety net.
• YOLO Mode (hx --yolo): Everything auto-runs. No prompts. Full speed, full trust. Use this when you know what you are asking for and want zero interruptions.

How the Agent Works

Here is what happens when you type a message:

1. You describe what you want in natural language. "Add input validation to the signup form" or "Why is this test failing?"
2. The agent queries spiral memory to understand your project's conventions, architecture, and history.
3. It reads relevant files to get the current state of your code.
4. It decides which tools to use based on what it learned.
5. Permission check runs for each tool call (safe / ask / dangerous).
6. The tool executes and a checkpoint is auto-created so you can undo.
7. The result feeds back to the agent for the next step.
8. Repeat up to 25 tool calls in a single turn.

The agent does not just execute instructions blindly. It uses spiral memory to understand your project before making changes. It checks git_status before committing. It reads existing code before editing. It searches for patterns before adding new ones.

Quick Start

# Start interactive chat
hx

# Send a single instruction
hx -m "Add input validation to the signup form"

# Full autonomous mode. no permission prompts
hx --yolo

# Skip confirmation prompts (but not dangerous ops)
hx -s

Once inside the chat, you can use slash commands like /help, /undo, /diff, /brain, /sessions, and many more. See the Slash Commands reference for the full list.

Using the Agent in the Web App

The web console gives you the same agent experience in your browser:

1. Open the Console tab in your HelixMind dashboard.
2. Type commands or use the /auto and /security quick-start buttons.
3. Watch the agent work in real-time with streaming output.
4. All tool calls, decisions, and file changes are visible as they happen.
5. The same spiral memory, same checkpoints, and same undo system are available.

The web console is especially useful for monitoring long-running autonomous sessions. Start a /security scan or /auto fix session from the CLI, then check progress from anywhere in the web console.

What makes this different

Persistent memory. The agent remembers everything across sessions. Your architecture decisions, code patterns, past bugs, and project conventions are all stored in spiral memory and available in every future conversation.

22 specialized tools. This is not just code generation pasted into your editor. The agent reads files, runs tests, manages git, searches the web, controls a browser, and tracks bugs. It is a complete development workflow.

Checkpoint safety. Every single tool call creates an undo point. Made a mistake? Press ESC ESC to open the checkpoint browser and revert to any previous state. Code changes, chat history, or both.

Browser vision. The agent can open your web app, take screenshots, and analyze visual issues. It sees broken layouts, missing elements, and CSS problems that text-only tools miss entirely.

Web research. When the agent encounters an unfamiliar API or error, it searches the web, extracts knowledge, and stores it in spiral memory. Next time it sees the same problem, it already knows the answer.

Bug journal. The agent automatically detects bugs from your messages (in English and German) and tracks them in a persistent journal. Bugs survive across sessions and can be reviewed anytime with /bugs.