Skip to content

Binex

Debuggable runtime for AI agent pipelines

Binex orchestrates multi-agent workflows defined in YAML. It executes DAG-based pipelines with any combination of LLM providers, records every step, and lets you trace, replay, debug, and diff runs — giving you full observability into your AI agent systems.

Key Features

  • YAML-defined workflows — Describe multi-agent pipelines as directed acyclic graphs with a simple, declarative format. No code required.
  • Multi-provider LLM support — Mix OpenAI, Anthropic, Gemini, Ollama, Groq, Mistral, DeepSeek, Together, and OpenRouter in a single workflow via LiteLLM routing.
  • Full run observability — Every node execution is recorded. Trace timelines, inspect artifacts, debug failures, and replay past runs.
  • Root-cause analysis — Automatic failure diagnosis with cascade detection, latency anomaly flagging, and actionable recommendations.
  • Run bisection — Find the exact divergence point between two runs with content similarity analysis.
  • Run diffing — Compare two workflow runs side-by-side to understand what changed between executions.
  • Output schema validation — Define JSON Schema for node outputs with automatic retry on validation failure.
  • Streaming LLM output — Watch LLM tokens arrive in real-time with auto-detection for TTY terminals.
  • Agent-to-Agent (A2A) protocol — Connect to remote A2A-compatible agent servers with built-in Gateway proxy for capability-based routing, automatic failover, and health monitoring.
  • Pattern Nodes — 9 built-in agentic patterns (critic, debate, reflexion, best_of_n, scatter, FSM, constitutional, chain_of_verification, plan_execute) that expand into full sub-DAG pipelines at runtime.
  • Framework adapters — Integrate LangChain chains, CrewAI crews, and AutoGen teams as workflow nodes via thin-wrapper adapters. Install only what you need with optional extras.
  • Plugin system — Extend Binex with custom adapters using Python entry points or inline adapter_class configuration.
  • OpenTelemetry tracing — Optional run-level and node-level spans for external collectors (Jaeger, Tempo), with zero overhead when disabled.
  • Workflow versioning — Schema versioning with migration framework, plus workflow snapshots stored in SQLite for run reproducibility.
  • Export & webhooks — Export run data to CSV/JSON, webhook notifications on run lifecycle events.
  • Eval & regression testing — YAML eval suites with blessed baselines (binex eval run|bless|baselines) plus per-node assertions and golden-run gating (binex eval golden) for CI.
  • MCP server — Expose runs, debugging, and evals to Claude Code, Cursor, or any MCP-compatible agent via binex mcp serve.
  • OTel trace import — Turn OTLP/JSON traces from any agent framework into Binex runs (binex import otel, live binex collect collector).
  • Cron scheduler — Register workflows with a schedule: field and run them on cron via binex scheduler.
  • Web UI — Visual drag-and-drop editor, live run view (SSE), debugging, diffing, cost dashboards — 20 pages of full CLI parity (binex ui).
  • Interactive CLI — Project scaffolding, workflow validation, a built-in doctor command, and a start wizard to get you productive quickly.

Install

pip install binex

Quick Demo

binex hello                    # run a built-in demo workflow
binex run examples/simple.yaml # run a sample pipeline
binex debug <run-id>           # inspect the completed run

See the Quickstart for a full walkthrough.

Documentation

Section Description
Quickstart Install Binex and run your first workflow in under 5 minutes
CLI Reference All commands: hello, init, run, resume, debug, trace, replay, diff, artifacts, dev, doctor, validate, scaffold, cancel, start, explore, diagnose, bisect, gateway, plugins, export, workflow, eval, import, collect, mcp, scheduler, ui, cost
Concepts Core concepts: agents, workflows, artifacts, execution model, lineage tracking
Architecture Runtime internals: orchestrator, stores, adapters, scheduler, DAG engine
Workflow Format YAML schema reference with node specs, variables, conditionals, and defaults
Pattern Nodes 9 built-in agentic patterns that expand into sub-DAG pipelines
Multi-Provider LLM Using multiple LLM providers in a single workflow
Contributing Development setup, testing guide, and code style

License

Binex is released under the MIT License.