orxhestra-code 0.0.53

AI coding agent powered by orxhestra — reads, writes, edits code and runs commands

AI coding agent for your terminal — reads, writes, edits code and runs commands.

Powered by orxhestra. Works with any LangChain-supported LLM provider.

Install

# Install with your preferred provider
pip install orxhestra-code[openai]       # GPT-5.4, o3, o4, etc.
pip install orxhestra-code[anthropic]    # Claude
pip install orxhestra-code[google]       # Gemini
pip install orxhestra-code[aws]          # Bedrock
pip install orxhestra-code[azure-ai]     # Azure OpenAI
pip install orxhestra-code[mistral]      # Mistral
pip install orxhestra-code[groq]         # Groq
pip install orxhestra-code[ollama]       # Ollama (local)
pip install orxhestra-code[deepseek]     # DeepSeek
pip install orxhestra-code[xai]          # xAI / Grok
pip install orxhestra-code[openrouter]   # OpenRouter (multi-provider)

# Or install all providers at once
pip install orxhestra-code[all]

Or with Homebrew:

brew tap NicolaiLassen/orxhestra
brew install orxhestra-code

Or from source:

git clone https://github.com/NicolaiLassen/orxhestra-code.git
cd orxhestra-code
uv sync

Usage

# Start with default model (Claude Sonnet)
orx-coder

# Use any LangChain provider
orx-coder --model openai/gpt-5.4
orx-coder --model anthropic/claude-sonnet-4-6
orx-coder --model google/gemini-2.5-pro
orx-coder --model mistral/mistral-large-latest
orx-coder --model groq/llama-3.3-70b-versatile
orx-coder --model ollama/qwen2.5-coder:32b
orx-coder --model deepseek/deepseek-chat
orx-coder --model xai/grok-3

# Control LLM reasoning effort
orx-coder --effort low      # fast responses, 5 iterations max
orx-coder --effort medium   # balanced reasoning, 15 iterations max
orx-coder --effort high     # deep reasoning, 30 iterations max (default)

# Permission modes
orx-coder --permission-mode plan           # read-only analysis
orx-coder --permission-mode accept-edits   # auto-approve edits, prompt for shell
orx-coder --auto-approve                   # skip all approval prompts
orx-coder --permission-mode trust          # full autonomous mode

# Resume previous sessions
orx-coder --continue             # continue most recent session
orx-coder --resume SESSION_ID   # resume a specific session

# Work in a specific directory
orx-coder --workspace /path/to/project

# Pipe a command
echo "fix the failing tests" | orx-coder

Tools

The agent has access to these built-in tools:

web_search and web_fetch do not require a separate search API key. They use open-web search/fetching locally inside the CLI.

File Operations

• ls — List files and directories
• read_file — Read files with line numbers, offset/limit pagination, image support
• write_file — Create or overwrite files, auto-creates parent directories
• edit_file — Find-and-replace edits with diff output
• mkdir — Create directories
• glob — Fast file pattern matching (**/*.py, src/*.ts)
• grep — Search file contents with regex, context lines, glob filtering

Shell

• shell_exec — Run shell commands (git, npm, pip, make, tests, etc.)

Web

• web_search — Search the open web without an extra search API key
• web_fetch — Fetch a URL and extract readable content from the page

Planning

• enter_plan_mode — Switch to read-only mode for codebase exploration. The agent reads files, searches, and analyzes before writing code.
• exit_plan_mode — Present an implementation plan for your approval. You can approve, reject, or request changes before the agent starts coding.

Task Management

• write_todos — Create and update structured task lists. The agent breaks complex work into steps and tracks progress.

Delegation

• task — Delegate a complex subtask to a fresh agent with isolated context. Useful for parallel research or exploratory work that would clutter the main conversation.

Artifacts

• save_artifact — Save files or data to the artifact store
• load_artifact — Load a previously saved artifact
• list_artifacts — List all available artifacts

User Interaction

• human_input — Ask you a clarifying question when the agent needs more information before proceeding

Permission Modes

Control what the agent can do without prompting:

Mode	Reads	Edits	Shell	Web	Use case
default	auto	prompt	prompt	prompt	Normal usage — you approve each destructive or external action
plan	auto	deny	deny	deny	Read-only exploration and analysis
accept-edits	auto	auto	prompt	prompt	Focused coding — edits are expected, shell/web need approval
auto-approve	auto	auto	auto	auto	Full speed — trust the agent completely
trust	auto	auto	auto	auto	Like auto-approve with no warnings

Switch mid-session:

/permissions plan           # switch to read-only
/permissions accept-edits   # auto-approve edits
/permissions cycle          # rotate to next mode
/permissions                # show current mode

When a tool needs approval, you'll see:

  ? Allow: shell_exec: npm test
  [y/n/a(ll)] >

Type y to allow once, n to deny, or a to auto-approve all for the rest of the session.

Plan Mode

For non-trivial tasks, the agent enters plan mode to explore and design before coding:

orx-coder> build a REST API for user management

  ┌ enter_plan_mode
  └ done
  ┌ glob(**/*.py), read_file(app/models.py), grep(def.*user, path=app/)
  └ done
  ┌ exit_plan_mode
  │ IMPLEMENTATION PLAN
  │ ============================================================
  │ 1. Create app/routers/users.py with CRUD endpoints
  │ 2. Add UserCreate/UserUpdate schemas to app/schemas.py
  │ 3. Add pagination to list endpoint
  │ ============================================================
  │
  │ ? Approve this plan? [y/n/e(dit)] >

• y — Approve and start implementing
• n — Reject, agent asks what you'd prefer
• e — Request changes, agent revises the plan

Reasoning Effort

The --effort flag maps to each provider's native reasoning API:

Provider	Parameter	What it does
Anthropic / AWS	thinking.budget_tokens	Extended thinking with token budget
OpenAI / Azure	reasoning_effort + Responses API	Internal reasoning depth
Google / Vertex	thinking_level	Gemini thinking depth
xAI	reasoning_effort	Grok reasoning depth
DeepSeek	reasoning_effort	DeepSeek reasoning depth
Mistral	reasoning_effort	Magistral reasoning depth
Groq	reasoning_effort	Groq reasoning depth
Cohere	thinking.budget_tokens	Command A reasoning budget

REPL Commands

Command	Description
/model <name>	Switch model mid-session
/permissions <mode>	Switch permission mode
/perm cycle	Cycle to next permission mode
/cost	Show session token usage
/diff [scope] [full]	Show git changes with previews (/diff staged, /diff full)
/compact	Summarize history to free context (also auto-triggers at 80K chars)
/todos	Show task list
/session	Session info
/undo	Remove last turn
/retry	Re-run last message
/copy	Copy last response
/memory	Browse saved memories
/theme	Switch theme
/clear	Clear conversation
/help	Show all commands
/exit	Quit

Configuration

Create ~/.orx-coder/config.yaml for persistent defaults:

model: anthropic/claude-sonnet-4-6
effort: high
permission_mode: default

Environment variables

Variable	Description
ORX_MODEL	Override model (e.g. openai/gpt-5.4)
ORX_EFFORT	Override effort (low, medium, high)
ORX_PERMISSION_MODE	Override permission mode
ANTHROPIC_API_KEY	Anthropic API key
OPENAI_API_KEY	OpenAI API key
GOOGLE_API_KEY	Google AI API key
GROQ_API_KEY	Groq API key
MISTRAL_API_KEY	Mistral API key
TOGETHER_API_KEY	Together API key
FIREWORKS_API_KEY	Fireworks API key

Project Instructions

The agent loads instructions from multiple sources (closest files have highest priority):

File	Scope	Description
CLAUDE.md	Project (shared)	Team-shared project rules, version controlled
CLAUDE.local.md	Project (personal)	Personal rules, gitignored
.claude/CLAUDE.md	Project	Alternative location
.orx/instructions.md	Project	Alternative location
~/.claude/CLAUDE.md	User	Global rules for all projects
~/.orx-coder/CLAUDE.md	User	Global rules for all projects

Instructions support @path/to/file import directives for composing rules from multiple files. HTML comments are stripped to save tokens. Files are truncated at 50KB with a warning.

# Project rules

- Use pytest for testing
- Follow PEP 8
- Always run tests before committing

@.claude/style-guide.md

Session Persistence

Sessions are automatically saved to ~/.orx-coder/sessions.db (SQLite). Context is auto-compacted when it exceeds 80K characters.

orx-coder -c                    # continue most recent session
orx-coder -r SESSION_ID         # resume specific session

License

Apache-2.0