orxhestra

Compose multi-agent AI systems with async event streaming, agent hierarchies, and built-in support for MCP and A2A protocols.

Orx CLI

Turn any orx.yaml agent setup into an interactive terminal agent. Ships with a coding agent out of the box — or compose your own.

Looking for a full-featured coding agent? Check out orxhestra-code — an enhanced coding agent built on orxhestra with permissions, multi-file editing, and project-aware context.

pip install orxhestra[cli,openai]
orx

+-- orx - terminal coding agent ------------------------------------+
|  model: gpt-5.4   workspace: ~/my-project   /help for commands    |
+-------------------------------------------------------------------+

orx> add error handling to the API routes

  > read_file(src/api/routes.py)
  > grep(pattern="raise", path=src/api/)
  > write_todos(3 tasks)

  Tasks
  * Add try/except to all route handlers  [in progress]
  - Add custom error response model
  - Write tests for error cases

  > edit_file(src/api/routes.py)
  > shell_exec(pytest tests/test_api.py)
  4 passed

  Done - added structured error handling to all 4 route handlers
  with a custom ErrorResponse model. All tests pass.

Features

• 29 LLM providers — OpenAI, Azure OpenAI, Anthropic, Google, Mistral, Cohere, Groq, DeepSeek, Ollama, and 20 more via --model
• Streaming — real-time token rendering with Markdown formatting
• Tool approval — prompts before destructive operations (write, edit, shell)
• Task planning — structured todo lists visible in the terminal
• Sub-agent delegation — spawn isolated agents for complex subtasks
• Auto-memory — persistent per-project memories across sessions (4 types: user, feedback, project, reference)
• Dark/light theme — auto-detects terminal, toggle with /theme
• Background tasks — spawn and monitor async sub-agent tasks
• Smart file reading — offset/limit pagination with line numbers, 256KB size guard
• Local context injection — auto-detects language, git state, package manager, project tree
• Context summarization — auto-compacts long conversations, /compact command
• Orx YAML — run any orx.yaml agent team: orx my-agents.yaml

Usage

orx                               # interactive REPL (default model)
orx --model claude-sonnet-4-6     # use a specific model
orx -c "fix the failing tests"    # single-shot command
orx my-agents.yaml                # run a custom orx file
orx --auto-approve                # skip approval prompts
orx orx.yaml --serve -p 9000      # start as A2A server

Commands

Command	Description
/model <name>	Switch model mid-session
/clear	Reset conversation
/compact	Summarize old messages to free context
/todos	Show current task list
/memory	List saved memories
/theme	Switch dark/light theme
/session	Session info (includes active signer DID when identity is on)
/undo	Remove last turn
/retry	Re-run last message
/copy	Copy last response
/help	Show all commands
/exit	Exit

orx identity — Ed25519 signing

Opt-in identity for every agent the CLI spawns. Events get signed with Ed25519, chained per branch, and (optionally) audited by an AttestationProvider.

orx identity init                          # generate a keypair at ~/.orx/identity.key
orx identity show                           # print the DID + public-key multibase
orx identity did-web example.com agents     # render a did.json for hosting

orx --identity ~/.orx/identity.key          # attach identity to every agent
export ORX_IDENTITY=~/.orx/identity.key     # or via env

See Composer → Identity, trust, and attestation for the YAML equivalents.

---

Quickstart (SDK)

pip install orxhestra
# or
uv add orxhestra

from orxhestra import LlmAgent, Runner, InMemorySessionService

agent = LlmAgent(
    name="assistant",
    model="gpt-5.4",
    instructions="You are a helpful assistant.",
)

runner = Runner(agent=agent, session_service=InMemorySessionService())
response = await runner.run(user_id="user1", session_id="s1", new_message="Hello!")

for event in response:
    print(event.content)

[!TIP] For persistent database sessions, install the database extra: pip install orxhestra[database]

[!TIP] For full documentation, guides, and API reference, visit docs.orxhestra.com.

Features

• Agent ensemble - LLM, ReAct, Sequential, Parallel, and Loop agents
• 29 LLM providers - OpenAI, Azure OpenAI, Anthropic, Google, Mistral, Cohere, Groq, DeepSeek, Ollama, and 20 more
• Event streaming - Async event-driven architecture with real-time streaming
• Composer - Declarative YAML with four pluggable registries: custom agent types, LLM providers, built-in tools, and tool-type resolvers
• Tools - Function tools, filesystem tools, agent-as-tool, shell, transfer routing, long-running tools, and register_tool_resolver for whole new tool kinds
• Planners - Choreograph task execution with PlanReAct and TaskPlanner strategies
• Skills - Reusable, composable agent repertoires (Agent Skills Protocol)
• MCP - Full-spec Model Context Protocol client (tools, resources, prompts, sampling, logging, progress, elicitation) plus adapters that turn MCP prompts into LangChain messages or tools
• A2A - Full v1.0 server + client with Ed25519 message signing and verification_method on agent cards
• Identity / Trust / Attestation (opt-in) - Sign every event, verify peers via DID, hash-chained audit log with a pluggable AttestationProvider — all wireable from a YAML block or a single orx --identity flag
• Auto-memory - Persistent memories with save_memory tool (user, feedback, project, reference)
• Background tasks - Async sub-agent task lifecycle with spawn and monitor
• Deprecation decorators - @deprecated and @deprecated_param for clean API evolution
• Tracing - Built-in support for Langfuse, LangSmith, and custom callbacks

Agents at a glance

Agent	Description
LlmAgent	Chat model agent with tools, instructions, and structured output
ReActAgent	Reasoning + acting loop with automatic tool use
SequentialAgent	Runs sub-agents in order
ParallelAgent	Runs sub-agents concurrently
LoopAgent	Repeats a sub-agent until exit condition
A2AAgent	Connects to remote agents via A2A protocol

Composer

Define entire agent orchestras in a single YAML file — no Python wiring needed. Compose LLM agents, loops, pipelines, tools, and review cycles declaratively. The example below builds a coding agent that plans, implements with filesystem + shell access, and self-reviews in a loop. Identity signing + local audit are opt-in — remove the last two blocks to turn them off.

defaults:
  model:
    provider: openai
    name: gpt-5.4

tools:
  exit:
    builtin: "exit_loop"
  filesystem:
    builtin: "filesystem"
  shell:
    builtin: "shell"

agents:
  planner:
    type: llm
    description: "Plans the implementation steps for the coder agent."
    instructions: |
      Output a numbered list of concrete steps the coder
      should execute. Each step must be an actionable file
      operation or shell command.

  coder:
    type: llm
    description: "Implements code changes with filesystem and shell access."
    instructions: |
      Follow the plan from the previous step exactly.
      Use filesystem tools to create files and shell to
      run commands. Never ask the user to do anything.
    tools:
      - filesystem
      - shell

  reviewer:
    type: llm
    description: "Reviews changes and approves or requests fixes."
    instructions: |
      Check files exist and look correct. If done, call
      exit_loop. Otherwise describe what needs fixing.
    tools:
      - exit

  dev_loop:
    type: loop
    agents: [coder, reviewer]
    max_iterations: 10

  coordinator:
    type: sequential
    agents: [planner, dev_loop]

main_agent: coordinator

runner:
  app_name: coding-agent
  session_service: memory

# Optional: sign every event + write a hash-chained audit log.
identity:
  signing_key: ./keys/agent.key         # orx identity init --path ./keys/agent.key
  did_method: key
attestation:
  provider: local
  path: ./audit

Run it as an interactive CLI or expose it as an A2A server:

orx orx.yaml                    # interactive terminal agent
orx orx.yaml --serve -p 9000    # A2A server on port 9000

# test the server
curl -X POST http://localhost:9000/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0", "id": "1",
    "method": "message/send",
    "params": {
      "message": {
        "role": "user",
        "parts": [{"text": "Hello!", "mediaType": "text/plain"}]
      }
    }
  }'

Docker

docker run -e OPENAI_API_KEY=$OPENAI_API_KEY \
  -v ./orx.yaml:/app/orx.yaml \
  nicolaimtlassen/orxhestra

Documentation

• Getting Started — Install and run your first agent (YAML or Python)
• Composer overview — YAML-based agent composition (recommended starting point)
• Composer schema reference — Field-by-field reference for every orx.yaml block
• Extending the composer — Register custom agent types, LLM providers, built-in tools, and tool resolvers
• Agents — Agent types and configuration
• Tools — Built-in and custom tools
• Integrations — MCP and A2A setup
• Skills — Code-level CLI skill references (agent-tools, callbacks, planners, streaming, and more)
• orxhestra-code — Enhanced coding agent with permissions, multi-file editing, and project context

---

Acknowledgments

This project is built on the shoulders of several outstanding open-source projects and research efforts:

• LangChain
• Google Agent Development Kit (ADK)
• LangGraph
• Model Context Protocol (MCP)
• Agent-to-Agent Protocol (A2A)
• attestix — inspired the shape of orxhestra's identity / trust / attestation layer (hash-chained audit logs, Ed25519-signed claims, DID-based agent identity). orxhestra ships the provider-agnostic AttestationProvider protocol; attestix plugs in as an external implementation.

Special thanks to the open-source AI community for pushing the boundaries of what's possible with agent frameworks.