code-muse 0.1.72

Muse — eternal guide of creators in the arts and sciences. An elegant AI coding assistant.

MUSE — where ancient inspiration meets modern craft

⭐ Star this repo if you seek the muse's favor ⭐

"Nine voices, one purpose: to shape raw thought into crafted form." — After Hesiod

---

Overview

In Greek myth, the nine Muses — Calliope of epic poetry, Clio of history, Erato of love verse, Euterpe of music, Melpomene of tragedy, Polyhymnia of sacred hymn, Terpsichore of dance, Thalia of comedy, and Urania of astronomy — breathed inspiration into mortals, transforming raw ambition into enduring craft.

MUSE carries that lineage forward. Each agent is a modern Muse: a specialist that illuminates its domain with precise, disciplined intelligence. Together, they orchestrate your work not with brute force, but with the clarity that comes from mastering a single art.

Where others would throw compute at complexity, MUSE channels inspiration — the oldest and most powerful force in creation. The path of the inspired craftsperson is always the more rewarding one.

MUSE is an AI-powered code generation agent, designed to understand programming tasks, generate high-quality code, and explain its reasoning — an open-source instrument for those who refuse to choose between velocity and elegance.

✨ Features

Muse is an open-source AI coding agent with an extensive feature set built around a plugin-based architecture. Below is a comprehensive catalog of everything Muse offers:

Core Intelligence

• Multi-Provider AI — Supports OpenAI (GPT-4o, o-series), Anthropic (Claude 3.5/4), Google Gemini (1.5/2), Cerebras, xAI Grok, Groq, Mistral, Together AI, Perplexity, DeepInfra, Cohere, AWS Bedrock, Ollama (local), and 65+ more providers via models.dev
• models.dev Integration — Browse and add models from 65+ providers with a single /add_model command, including interactive TUI, model preview, smart warnings for unsupported/no-tool-calling models, and offline fallback
• Round Robin Model Distribution — Rotate through multiple API keys/models with configurable rotate_every frequency to maximize throughput and stay within rate limits
• Custom Model Timeouts — Per-model timeout configuration for slow or unreliable endpoints
• Multi-Agent System — Specialist agents (Muse, Helios, QA Iris, Code Critic) with dynamic switching via /agent

Token Optimization

• Filter Engine — Smart command output classifier routing domain-specific output through compression strategies
• Shell Minimizer — Pipeline-based compression reducing shell output tokens by 60–90%
• Git Compression (~85%) — Compresses git status, git log, git diff into one-liners
• Test Compression (~90%) — Shows only failures + summary for pytest, cargo test, and more
• Lint Compression (~80%) — Groups lint errors by rule, not by file
• Code Compression (~50%) — Strips comments, trims boilerplate, removes excess blank lines
• Read Compression (~60%) — Smart-ranged file reading, grep/find output grouping
• Tree-sitter AST Compression — Language-aware abstract syntax tree compression via tree-sitter parsers (Python, JavaScript, Go)
• JSON Compression — Structural JSON output minimization
• Build Filter — Compresses build tool output (make, cargo, npm, pip, etc.)

Agent System

• Built-in Agents — Muse (default), Helios (Universal Constructor), QA Iris (browser testing), Code Critic (code review), Planning Agent, Agent Creator
• Custom JSON Agents — Create your own agents without Python knowledge via JSON configuration files
• Agent Creator — Interactive /agent agent-creator guided process for forging custom agents
• Agent Skills — Discoverable, installable skills with remote catalog, local catalogs, and interactive menus
• Sub-Agent Delegation — Invoke specialist sub-agents with full context passing

Plugin Architecture

• 40+ Built-in Plugins — All under code_muse/plugins/ with a standardized callback registration system
• Callback Hook System — 30+ lifecycle hooks (startup, shutdown, pre/post tool call, file permission, shell command, agent run events, streaming, custom commands, model registration, and more)
• User Plugin Support — ~/.muse/plugins/ with content-hash-based trust manifest
• Plugin Trust Model — Fail-closed by default, per-plugin trust approval, content hash verification

Plugin Catalog

• Smart Compression — semantic_compression plugin for intelligent code/text compression
• Autonomous Memory — Long-term memory with BM25 scoring, session scanning, lease locking, memory consolidation, secret scanning
• Token Tracking — SQLite-based per-command, per-session, cumulative token usage with reports
• Token Caching — Cache hit tracking, cacheable prefix detection, stats display for Anthropic prompt caching
• Plan Mode — Three-mode cycling (DEFAULT → AUTO_EDIT → PLAN) with plan approval workflow, goal setting, plan editor
• Checkpointing — Conversation snapshots, restore, rewind, shadow git auto-commit on every assistant message
• Code Critic — Automated strict code review returning approved/rejected verdicts with detailed feedback
• Auto Review — Automatic code review on file changes with visibility control and caching
• Destructive Command Guard — Detects and blocks dangerous shell commands (rm -rf, dd, format, mkfs, etc.)
• Force Push Guard — Protection against accidental git push --force
• File Permission Handler — Granular file operation permission control with user approval flows
• Shell Safety — Command approval, command caching, agent shell safety boundaries
• Policy Engine — TOML-based policy files with approval flow integration
• Custom Commands — Slash commands from markdown files or TOML schemas
• Customizable Commands — User-customizable command definitions
• Build Filter — Build output compression strategies
• GAC (Git Auto Commit) — Automated git add/commit with intelligent prompts
• Hook Engine — Full hook system with trust, aliases, matchers, validators, executor, registry
• Hook Manager — Interactive hook management UI
• Hook Creator — Interactive hook creation wizard
• AWS Bedrock — Amazon Bedrock model integration
• ChatGPT OAuth — OAuth flow for ChatGPT Codex integration
• Claude Code Hooks & OAuth — Claude Code integration with token refresh heartbeat, fast mode
• Ollama Setup — Interactive local model setup with auto-completion
• Agent Skills — Full skill system with discovery, download, installation, metadata, config
• Tool Registry — Centralized tool definitions with standardized registration
• Universal Constructor — Safe user-generated tool execution in sandboxed subprocess workers with JSON-only serialization, eval/exec blocking, timeout enforcement
• TPS Meter — Real-time tokens-per-second metering
• Safety Status — /safety and /status commands to inspect security posture
• Frontend Emitter — WebSocket-based frontend event emission
• History Store — Session history persistence
• Prompt Newline — Configurable newline behavior in prompts
• Synthetic Status — API status checks and health monitoring
• Plugin Trust Manager — Plugin trust manifest UI

Developer Tools

• Chrome CDP — Full browser debugging via Chrome DevTools Protocol: navigate, screenshot, accessibility snapshots, console error reading, network capture, JavaScript eval, click, type, view source
• mitmproxy — Network traffic capture and inspection for debugging HTTP/S requests/responses
• Universal Constructor — Create any tool dynamically at runtime with sandboxed execution
• Image Analysis — Load and analyze images with vision-capable models

Development Environment

• Interactive TUI — Rich terminal UI with prompt toolkit, theming, spinners, shimmer effects
• Message Bus — Structured event-driven messaging system with queue, console, and renderer layers
• Diff Menu — Interactive diff viewing with syntax highlighting
• Colors Menu — Configurable color themes and appearance
• Model Settings Menu — Interactive model configuration and switching
• Autosave — Automatic session state persistence
• Onboarding Wizard — Interactive first-run setup with slides
• Agent Menu — Interactive agent switching
• File Path Completion — Tab completion with fuzzy matching
• Model Picker Completion — Fuzzy model selection

Configuration & Customization

• AGENTS.md Rules — Hierarchical agent rules (global → project) with .muse/ directory precedence
• TOML Config — ~/.muse/config.toml with per-project overrides
• Custom Slash Commands — .md files in .claude/commands/, .github/prompts/, or .agents/commands/
• Session Management — Save, load, resume, truncate, export sessions with JSON storage
• Secret Storage — Private file creation with 0o600 permissions

Security

• JSON Session Storage — No pickle/RCE risk (legacy pickle rejected unless explicitly imported)
• Secret Redaction — Logs scrub Authorization: Bearer ..., sensitive query params; token length never logged
• Shell Approval — Commands require approval by default; yolo_mode is off
• Workspace Boundaries — File tools enforce cwd containment, block .env, .ssh, etc.
• Hook Trust — Content-hash based trust verification
• Grep Safety — Patterns passed after -- to prevent CLI flag injection
• Secret Redaction — High-performance secret redaction with pure-Python pattern matching

Quality of Life

• 100% Privacy — Zero telemetry, zero prompt logging, zero behavioral profiling, zero third-party data sharing
• Beads Issue Tracking — Integrated bd issue tracker with local Dolt DB and git sync
• Version Checker — Automatic version checking with update notifications
• Event Streaming — Real-time streaming of agent events
• Sub-agent Streaming — Streaming output from sub-agent invocations
• Agent Run Events — Lifecycle hooks with full metadata for agent execution tracking

Quick start

uvx code-muse -i

If uvx still starts an older cached version after a release, refresh the tool environment:

uvx --refresh-package code-muse code-muse -i

🪙 Token Savings

Muse compresses shell command output the way a sculptor removes excess marble — reducing token usage by 60–90%.

Strategy	What it does	Savings
Git	Compresses status/log/diff into one-liners	~85%
Test	Shows only failures + summary	~90%
Lint	Groups errors by rule, not by file	~80%
Code	Strips comments, trims boilerplate	~50%
Read	Smart-ranged file reading	~60%

Quick start: Run /init in your project to lay the foundation.

See FEATURES.md for detailed examples of each strategy.

Installation

UV (Recommended)

macOS / Linux

# Install UV if you don't have it
curl -LsSf https://astral.sh/uv/install.sh | sh

uvx code-muse

Windows

On Windows, we recommend installing code-muse as a global tool for the best experience with keyboard shortcuts (Ctrl+C/Ctrl+X cancellation):

# Install UV if you don't have it (run in PowerShell as Admin)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

uvx code-muse

Changelog (By Kittylog!)

📋 View the full changelog on Kittylog

Usage

Adding Models from models.dev 🆕

While there are several models configured right out of the box from providers like Synthetic, Cerebras, OpenAI, Google, and Anthropic, Muse integrates with models.dev to let you browse and add models from 65+ providers with a single command:

/add_model

This opens an interactive TUI where you can:

• Browse providers - See all available AI providers (OpenAI, Anthropic, Groq, Mistral, xAI, Cohere, Perplexity, DeepInfra, and many more)
• Preview model details - View capabilities, pricing, context length, and features
• One-click add - Automatically configures the model with correct endpoints and API keys

Live API with Offline Fallback

The /add_model command fetches the latest model data from models.dev in real-time. If the API is unavailable, it falls back to a bundled database:

📡 Fetched latest models from models.dev     # Live API
📦 Using bundled models database              # Offline fallback

Supported Providers

Muse integrates with https://models.dev giving you access to 65 providers and >1000 different model offerings.

There are 39+ additional providers that already have OpenAI-compatible APIs configured in models.dev!

These providers are automatically configured with correct OpenAI-compatible endpoints, but have not been tested thoroughly:

Provider	Endpoint	API Key Env Var
xAI (Grok)	https://api.x.ai/v1	XAI_API_KEY
Groq	https://api.groq.com/openai/v1	GROQ_API_KEY
Mistral	https://api.mistral.ai/v1	MISTRAL_API_KEY
Together AI	https://api.together.xyz/v1	TOGETHER_API_KEY
Perplexity	https://api.perplexity.ai	PERPLEXITY_API_KEY
DeepInfra	https://api.deepinfra.com/v1/openai	DEEPINFRA_API_KEY
Cohere	https://api.cohere.com/compatibility/v1	COHERE_API_KEY
AIHubMix	https://aihubmix.com/v1	AIHUBMIX_API_KEY

Smart Warnings

• ⚠️ Unsupported Providers - Providers like Amazon Bedrock and Google Vertex that require special authentication are clearly marked
• ⚠️ No Tool Calling - Models without tool calling support show a big warning since they can't use Muse's file/shell tools

Custom Commands

Create markdown files in .claude/commands/, .github/prompts/, or .agents/commands/ to define custom slash commands. The filename becomes the command name and the content runs as a prompt.

# Create a custom command
echo "# Code Review

Please review this code for security issues." > .claude/commands/review.md

# Use it in Muse
/review with focus on authentication

Requirements

• Python 3.14+
• OpenAI API key (for GPT models)
• Gemini API key (for Google's Gemini models)
• Cerebras API key (for Cerebras models)
• Anthropic key (for Claude models)
• Ollama endpoint available

Agent Rules

Muse supports AGENTS.md files for defining coding standards, project conventions, and behavioral guidelines — the laws of your workshop. These rules cover formatting, naming conventions, architectural patterns, and project-specific instructions.

For examples and more information about agent rules, visit https://agent.md

AGENTS.md Search Order

Muse loads rules from multiple locations, combining them in order:

Priority	Location	Purpose
1	~/.muse/AGENTS.md	Global rules (applied to all projects)
2	.muse/AGENTS.md	Project rules (preferred location)
3	./AGENTS.md	Project rules (alternate location)

Key behaviors:

• Global and project rules are combined (global first, then project)
• .muse/ directory takes precedence over project root
• All filename variants are supported: AGENTS.md, AGENT.md, agents.md, agent.md

Round Robin Model Distribution

Muse supports Round Robin model distribution — cycling through configured models the way the Muses cycled through their domains, each taking its turn. This feature automatically rotates through models with each request, maximizing API usage while staying within rate limits.

Configuration

Add a round-robin model configuration to your ~/.muse/extra_models.json file:

export CEREBRAS_API_KEY1=csk-...
export CEREBRAS_API_KEY2=csk-...
export CEREBRAS_API_KEY3=csk-...

{
  "qwen1": {
    "type": "cerebras",
    "name": "qwen-3-coder-480b",
    "custom_endpoint": {
      "url": "https://api.cerebras.ai/v1",
      "api_key": "$CEREBRAS_API_KEY1"
    },
    "context_length": 131072
  },
  "qwen2": {
    "type": "cerebras",
    "name": "qwen-3-coder-480b",
    "custom_endpoint": {
      "url": "https://api.cerebras.ai/v1",
      "api_key": "$CEREBRAS_API_KEY2"
    },
    "context_length": 131072
  },
  "qwen3": {
    "type": "cerebras",
    "name": "qwen-3-coder-480b",
    "custom_endpoint": {
      "url": "https://api.cerebras.ai/v1",
      "api_key": "$CEREBRAS_API_KEY3"
    },
    "context_length": 131072
  },
  "cerebras_round_robin": {
    "type": "round_robin",
    "models": ["qwen1", "qwen2", "qwen3"],
    "rotate_every": 5
  }
}

Then just use /model and tab to select your round-robin model!

The rotate_every parameter controls how many requests are made to each model before rotating to the next one. In this example, the round-robin model will use each Qwen model for 5 consecutive requests before moving to the next model in the sequence.

Custom Model Timeouts

For custom model endpoints (custom_openai, custom_anthropic, custom_gemini, cerebras), you can configure custom timeout values to handle slow or unreliable endpoints. The default timeout for these custom endpoint models is 180 seconds.

Note: Other model types have different default timeouts:

• ChatGPT/Codex models: 300 seconds (5 minutes)
• Regular Anthropic models: 180 seconds
• Gemini models: 180 seconds

Configuration

Add a timeout field to your model configuration in ~/.muse/extra_models.json:

{
  "slow_model": {
    "type": "custom_openai",
    "name": "gpt-4",
    "custom_endpoint": {
      "url": "https://slow-endpoint.example.com/v1",
      "api_key": "$API_KEY",
      "timeout": 600
    }
  },
  "fast_model": {
    "type": "cerebras", 
    "name": "llama3.1-8b",
    "custom_endpoint": {
      "url": "https://api.cerebras.ai/v1",
      "api_key": "$CEREBRAS_API_KEY"
    },
    "timeout": 300
  }
}

The timeout value can be specified either:

• Inside the custom_endpoint object (recommended for endpoint-specific timeouts)
• At the top level of the model config (affects all custom endpoint types)

Timeout values must be positive numbers (integers or floats) representing seconds. If no timeout is specified, the default 180-second timeout is used for custom endpoint models.

---

Create Your Own Agent

Muse features a flexible agent system — each agent a specialist Muse in its own right — that allows you to work with tailored AI assistants for different domains of craft. The system supports both built-in Python agents and custom JSON agents that you can forge yourself.

Quick Start

Check Current Agent

/agent

Shows current active agent and all available agents

Switch Agent

/agent <agent-name>

Switches to the specified agent

Create New Agent

/agent agent-creator

Switches to the Agent Creator for building custom agents

Truncate Message History

/truncate <N>

Truncates the message history to keep only the N most recent messages while protecting the first (system) message. For example:

/truncate 20

Would keep the system message plus the 19 most recent messages, removing older ones from the history.

This is useful for managing context length when you have a long conversation history but only need the most recent interactions.

Available Agents

Muse (Default)

• Name: muse
• Specialty: General-purpose coding assistant — the Calliope of code
• Personality: Playful, sharp, pedantic about craftsmanship
• Tools: Full access to all tools
• Best for: All coding tasks, file management, execution
• Principles: Clean, concise code following YAGNI, SRP, DRY principles
• File limit: Max 600 lines per file (enforced!)

Agent Creator 🏛️

• Name: agent-creator
• Specialty: Forging custom JSON agent configurations — the Hephaestus of agents
• Tools: File operations, reasoning
• Best for: Building new specialized agents
• Features: Schema validation, guided creation process

Agent Types

Python Agents

Built-in agents forged in Python with full system integration:

• Discovered automatically from code_muse/agents/ directory
• Inherit from BaseAgent class
• Full access to system internals
• Examples: muse, agent-creator

JSON Agents

Agents you craft yourself, defined in JSON files:

• Stored in user's agents directory
• Easy to create, share, and modify
• Schema-validated configuration
• Custom system prompts and tool access

Creating Custom JSON Agents

Using Agent Creator (Recommended)

1. Switch to Agent Creator:

   /agent agent-creator
   

2. Request agent creation:

   I want to create a Python tutor agent
   

3. Follow guided process to define:

   • Name and description
   • Available tools
   • System prompt and behavior
   • Custom settings

4. Test your new agent:

   /agent your-new-agent-name
   

Manual JSON Creation

Create JSON files in your agents directory following this schema:

{
  "name": "agent-name",              // REQUIRED: Unique identifier (kebab-case)
  "display_name": "Agent Name 🤖",   // OPTIONAL: Pretty name with emoji
  "description": "What this agent does", // REQUIRED: Clear description
  "system_prompt": "Instructions...",    // REQUIRED: Agent instructions
  "tools": ["tool1", "tool2"],        // REQUIRED: Array of tool names
  "user_prompt": "How can I help?",     // OPTIONAL: Custom greeting
  "tools_config": {                    // OPTIONAL: Tool configuration
    "timeout": 60
  }
}

Required Fields

• name: Unique identifier (kebab-case, no spaces)
• description: What the agent does
• system_prompt: Agent instructions (string or array)
• tools: Array of available tool names

Optional Fields

• display_name: Pretty display name (defaults to title-cased name + 🤖)
• user_prompt: Custom user greeting
• tools_config: Tool configuration object

Available Tools

Agents can access these tools based on their configuration:

• list_files: Directory and file listing
• read_file: File content reading
• grep: Text search across files
• create_file: Create new files or overwrite existing ones
• replace_in_file: Targeted text replacements in existing files
• delete_snippet: Remove a text snippet from a file
• delete_file: File deletion
• agent_run_shell_command: Shell command execution
• agent_share_your_reasoning: Share reasoning with user

Tool Access Examples

• Read-only agent: ["list_files", "read_file", "grep"]
• File editor agent: ["list_files", "read_file", "create_file", "replace_in_file"]
• Full access agent: All tools (like Muse)

System Prompt Formats

String Format

{
  "system_prompt": "You are a helpful coding assistant that specializes in Python development."
}

Array Format (Recommended)

{
  "system_prompt": [
    "You are a helpful coding assistant.",
    "You specialize in Python development.",
    "Always provide clear explanations.",
    "Include practical examples in your responses."
  ]
}

Example JSON Agents

Python Tutor

{
  "name": "python-tutor",
  "display_name": "Python Tutor 🐍",
  "description": "Teaches Python programming concepts with examples",
  "system_prompt": [
    "You are a patient Python programming tutor.",
    "You explain concepts clearly with practical examples.",
    "You help beginners learn Python step by step.",
    "Always encourage learning and provide constructive feedback."
  ],
  "tools": ["read_file", "create_file", "replace_in_file", "agent_share_your_reasoning"],
  "user_prompt": "What Python concept would you like to learn today?"
}

Code Reviewer

{
  "name": "code-reviewer",
  "display_name": "Code Reviewer 🔍",
  "description": "Reviews code for best practices, bugs, and improvements",
  "system_prompt": [
    "You are a senior software engineer doing code reviews.",
    "You focus on code quality, security, and maintainability.",
    "You provide constructive feedback with specific suggestions.",
    "You follow language-specific best practices and conventions."
  ],
  "tools": ["list_files", "read_file", "grep", "agent_share_your_reasoning"],
  "user_prompt": "Which code would you like me to review?"
}

DevOps Helper

{
  "name": "devops-helper",
  "display_name": "DevOps Helper ⚙️",
  "description": "Helps with Docker, CI/CD, and deployment tasks",
  "system_prompt": [
    "You are a DevOps engineer specialized in containerization and CI/CD.",
    "You help with Docker, Kubernetes, GitHub Actions, and deployment.",
    "You provide practical, production-ready solutions.",
    "You always consider security and best practices."
  ],
  "tools": [
    "list_files",
    "read_file",
    "create_file",
    "replace_in_file",
    "agent_run_shell_command",
    "agent_share_your_reasoning"
  ],
  "user_prompt": "What DevOps task can I help you with today?"
}

File Locations

JSON Agents Directory

• All platforms: ~/.muse/agents/

Python Agents Directory

• Built-in: code_muse/agents/ (in package)

Best Practices

Naming

• Use kebab-case (hyphens, not spaces)
• Be descriptive: "python-tutor" not "tutor"
• Avoid special characters

System Prompts

• Be specific about the agent's role
• Include personality traits
• Specify output format preferences
• Use array format for multi-line prompts

Tool Selection

• Only include tools the agent actually needs
• Most agents need agent_share_your_reasoning
• File manipulation agents need read_file, create_file, replace_in_file
• Note: "edit_file" still works in tool lists (auto-expands to the three individual tools)
• Research agents need grep, list_files

Display Names

• Include relevant emoji for personality
• Make it friendly and recognizable
• Keep it concise

System Architecture

Agent Discovery

The system automatically discovers agents by:

1. Python Agents: Scanning code_muse/agents/ for classes inheriting from BaseAgent
2. JSON Agents: Scanning user's agents directory for *-agent.json files
3. Instantiating and registering discovered agents

JSONAgent Implementation

JSON agents are powered by the JSONAgent class (code_muse/agents/json_agent.py):

• Inherits from BaseAgent for full system integration
• Loads configuration from JSON files with robust validation
• Supports all BaseAgent features (tools, prompts, settings)
• Cross-platform user directory support
• Built-in error handling and schema validation

BaseAgent Interface

Both Python and JSON agents implement this interface:

• name: Unique identifier
• display_name: Human-readable name with emoji
• description: Brief description of purpose
• get_system_prompt(): Returns agent-specific system prompt
• get_available_tools(): Returns list of tool names

Agent Manager Integration

The agent_manager.py provides:

• Unified registry for both Python and JSON agents
• Seamless switching between agent types
• Configuration persistence across sessions
• Automatic caching for performance

System Integration

• Command Interface: /agent command works with all agent types
• Tool Filtering: Dynamic tool access control per agent
• Main Agent System: Loads and manages both agent types
• Cross-Platform: Consistent behavior across all platforms

Adding Python Agents

To create a new Python agent:

1. Create file in code_muse/agents/ (e.g., my_agent.py)
2. Implement class inheriting from BaseAgent
3. Define required properties and methods
4. Agent will be automatically discovered

Example implementation:

from .base_agent import BaseAgent

class MyCustomAgent(BaseAgent):
    @property
    def name(self) -> str:
        return "my-agent"

    @property
    def display_name(self) -> str:
        return "My Custom Agent ✨"

    @property
    def description(self) -> str:
        return "A custom agent for specialized tasks"

    def get_system_prompt(self) -> str:
        return "Your custom system prompt here..."

    def get_available_tools(self) -> list[str]:
        return [
            "list_files",
            "read_file",
            "grep",
            "create_file",
            "replace_in_file",
            "delete_snippet",
            "delete_file",
            "agent_run_shell_command",
            "agent_share_your_reasoning"
        ]

Troubleshooting

Agent Not Found

• Ensure JSON file is in correct directory
• Check JSON syntax is valid
• Restart Muse or clear agent cache
• Verify filename ends with -agent.json

Validation Errors

• Use Agent Creator for guided validation
• Check all required fields are present
• Verify tool names are correct
• Ensure name uses kebab-case

Permission Issues

• Make sure agents directory is writable
• Check file permissions on JSON files
• Verify directory path exists

Advanced Features

Tool Configuration

{
  "tools_config": {
    "timeout": 120,
    "max_retries": 3
  }
}

Multi-line System Prompts

{
  "system_prompt": [
    "Line 1 of instructions",
    "Line 2 of instructions",
    "Line 3 of instructions"
  ]
}

Future Extensibility

The agent system is built for expansion — new muses for new domains:

• Specialized Agents: Code reviewers, debuggers, architects
• Domain-Specific Agents: Web dev, data science, DevOps, mobile
• Personality Variations: Different communication styles
• Context-Aware Agents: Adapt based on project type
• Team Agents: Shared configurations for coding standards
• Plugin System: Community-contributed agents

The Craft of JSON Agents

1. Easy Customization: Create agents without Python knowledge
2. Team Sharing: JSON agents can be shared across teams
3. Rapid Prototyping: Quick agent creation for specific workflows
4. Version Control: JSON agents are git-friendly
5. Built-in Validation: Schema validation with helpful error messages
6. Cross-Platform: Works consistently across all platforms
7. Backward Compatible: Doesn't affect existing Python agents

Implementation Details

Files in System

• Core Implementation: code_muse/agents/json_agent.py
• Agent Discovery: Integrated in code_muse/agents/agent_manager.py
• Command Interface: Works through existing /agent command
• Testing: Comprehensive test suite in tests/test_json_agents.py

JSON Agent Loading Process

1. System scans ~/.muse/agents/ for *-agent.json files
2. JSONAgent class loads and validates each JSON configuration
3. Agents are registered in unified agent registry
4. Users can switch to JSON agents via /agent <name> command
5. Tool access and system prompts work identically to Python agents

Error Handling

• Invalid JSON syntax: Clear error messages with line numbers
• Missing required fields: Specific field validation errors
• Invalid tool names: Warning with list of available tools
• File permission issues: Helpful troubleshooting guidance

Future Possibilities

• Agent Templates: Pre-built JSON agents for common tasks
• Visual Editor: GUI for creating JSON agents
• Hot Reloading: Update agents without restart
• Agent Marketplace: Share and discover community agents
• Enhanced Validation: More sophisticated schema validation
• Team Agents: Shared configurations for coding standards

Contributing

Releases

Maintainer release steps live in docs/RELEASING.md.

Sharing JSON Agents

1. Create and test your agent thoroughly
2. Ensure it follows best practices
3. Submit a pull request with agent JSON
4. Include documentation and examples
5. Test across different platforms

Python Agent Contributions

1. Follow existing code style
2. Include comprehensive tests
3. Document the agent's purpose and usage
4. Submit pull request for review
5. Ensure backward compatibility

Agent Templates

Consider contributing agent templates for:

• Code reviewers and auditors
• Language-specific tutors
• DevOps and deployment helpers
• Documentation writers
• Testing specialists

---

Security & Trust Boundaries

Muse guards your workshop with the vigilance of a temple guardian — multiple safety layers protect your secrets, filesystem, and runtime:

• Sessions use JSON by default — legacy pickle sessions are rejected unless explicitly imported with --import-legacy-pickle-session (RCE risk warning).
• Secrets are redacted — token files are created with 0o600, logs scrub Authorization: Bearer ... and sensitive query params, and token length is never logged.
• Shell commands require approval by default — yolo_mode is off; background commands require approval before Popen.
• Workspace boundaries — file tools enforce cwd containment, block sensitive paths (.env, .ssh, etc.), and cap huge files/diffs before full read.
• Hook trust — project hooks from .claude/settings.json require explicit trust (keyed by content hash); untrusted hooks are blocked.
• Universal Constructor safety — user-generated tools run in a subprocess worker with JSON-only serialization, dangerous patterns (eval, exec, subprocess) are blocked or approval-gated, and timeouts kill the worker process.
• Grep safety — search patterns are passed after -- so they are treated as data, not CLI flags.

Run /safety or /status inside Muse to inspect the current risk posture — no secrets exposed.

For full details, see docs/SECURITY.md.

---

Muse Privacy Commitment

Zero-compromise privacy. Always.

Unlike other agentic coding tools, this project has no corporate or investor backing — meaning zero pressure to compromise our principles for profit. Privacy isn't a feature we bolted on; it is the bedrock on which Muse was built.

What Muse absolutely does not collect — now and forever:

• ❌ Zero telemetry – no usage analytics, crash reports, or behavioral tracking
• ❌ Zero prompt logging – your code, conversations, or project details are never stored
• ❌ Zero behavioral profiling – we don't track what you build, how you code, or when you use the tool
• ❌ Zero third-party data sharing – your information is never sold, traded, or given away

What data flows where:

• LLM Provider Communication: Your prompts are sent directly to whichever LLM provider you've configured (OpenAI, Anthropic, local models, etc.) – this is unavoidable for AI functionality
• Complete Local Option: Run your own VLLM/SGLang/Llama.cpp server locally → zero data leaves your network. Configure this with ~/.muse/extra_models.json
• Direct Developer Contact: All feature requests, bug reports, and discussions happen directly with me – no middleman analytics platforms or customer data harvesting tools

The privacy-first architecture:

Muse is designed with privacy-by-design principles. Every feature has been evaluated through a privacy lens, and every integration respects your data sovereignty. When you use Muse, you are the craftsperson — never the product.

This commitment is structurally impossible to violate. No external pressures, no investor demands, no quarterly earnings targets. Just principled code that respects your craft and your privacy.

License

This project is licensed under the MIT License - see the LICENSE file for details.