Agentoven
AgentOven is a **framework-agnostic agent control plane** that standardizes how AI agents are built, deployed, observed, and orchestrated across an enterprise. Think of it as a **clay oven** 🏺 — you put in raw ingredients (models, tools, data, prompts) and **production-ready agents come out the chimney**.
Install / Use
/learn @agentoven/AgentovenREADME
What is AgentOven?
AgentOven is a framework-agnostic agent control plane that standardizes how AI agents are built, deployed, observed, and orchestrated across an enterprise.
Think of it as a clay oven 🏺 — you put in raw ingredients (models, tools, data, prompts) and production-ready agents come out the chimney.
The Problem
- Agents are built ad-hoc with no consistency
- No governance, audit trail, or cost visibility
- Locked into single vendors (Databricks, Azure, LangChain)
- Multi-agent workflows are stitched together manually
- No standard protocol for agent-to-agent collaboration
The Solution
AgentOven provides a unified control plane with:
| Capability | Description | |---|---| | 🏺 Agent Registry | Version, discover, and manage agents as first-class resources | | 🔀 Model Router | Intelligent routing across providers with fallback, cost optimization | | 🤝 A2A Native | Agent-to-Agent protocol built-in from day 1 | | 🔧 MCP Gateway | Model Context Protocol for tool/data integration | | 📊 Observability | OpenTelemetry tracing on every invocation, cost & latency dashboards | | 🔄 Workflow Engine | DAG-based multi-agent orchestration via A2A task lifecycle | | 📝 Prompt Studio | Versioned prompt management with diff view and A/B variants | | 💬 Sessions | Multi-turn chat sessions with history, thinking mode, and streaming | | 🛡️ Guardrails | Pre/post processing content filters and safety checks | | 🧪 Evaluation | Automated evals with LLM judges and regression detection | | 💰 Cost Tracking | Per-request token counting, tenant-level chargeback | | 🔐 Governance | Pluggable auth (API keys, service accounts, SSO), RBAC, audit logs | | 🔎 RAG Pipelines | 5 retrieval strategies with vector stores and embedding management |
Architecture
┌─────────────────────────────────────────────────────────────┐
│ AgentOven Control Plane │
│ (Registry · Router · RBAC · Cost · Tenancy) │ ← Go
├──────────────────────┬──────────────────────────────────────┤
│ A2A Gateway │ MCP Gateway │
│ (Agent ↔ Agent) │ (Agent ↔ Tools/Data) │ ← Go
├──────────────────────┴──────────────────────────────────────┤
│ AgentOven Runtime │
│ (Execute · Instrument · Route · Enforce Policies) │ ← Rust
├─────────┬─────────┬──────────┬──────────┬───────────────────┤
│LangGraph│ CrewAI │OpenAI SDK│ AutoGen │ Custom Agents │
└─────────┴─────────┴──────────┴──────────┴───────────────────┘
Quick Start
Install the CLI
# macOS
brew install agentoven/tap/agentoven
# Cargo
cargo install agentoven-cli
# Or download the binary
curl -fsSL https://agentoven.dev/install.sh | sh
Install the Python SDK
pip install agentoven
Install the TypeScript SDK
npm install @agentoven/sdk
Bake your first agent
# Initialize a project
agentoven init --name my-agent --framework openai-sdk
# Set up a model provider
agentoven provider add my-openai --kind openai --api-key $OPENAI_API_KEY
# Register an agent
agentoven agent register summarizer \
--description "Summarizes documents with citations" \
--model-provider my-openai \
--model-name gpt-4o \
--system-prompt "You are a document summarizer."
# Bake (deploy) the agent
agentoven agent bake summarizer
# Test it interactively
agentoven agent test summarizer --interactive
Or use the Python SDK
Simple — single model, no extras:
from agentoven import Agent, AgentOvenClient
agent = Agent("summarizer",
description="Summarizes documents with citations",
model_provider="my-openai",
model_name="gpt-4o",
system_prompt="You are a document summarizer.",
)
client = AgentOvenClient()
client.register(agent)
client.bake(agent, environment="production")
Advanced — multi-model fallback, tools, and MCP:
from agentoven import Agent, Ingredient, AgentOvenClient
agent = Agent("summarizer",
description="Summarizes documents with citations",
ingredients=[
Ingredient.model("gpt-4o", provider="my-openai"),
Ingredient.model("claude-sonnet", provider="anthropic", role="fallback"),
Ingredient.tool("document-reader", protocol="mcp"),
Ingredient.prompt("system", text="You are a document summarizer."),
],
)
client = AgentOvenClient()
client.register(agent)
client.bake(agent, environment="production")
# The agent is now discoverable via A2A
# Other agents can find it at:
# /.well-known/agent-card.json
Multi-Agent Recipes
from agentoven import Recipe, Step, AgentOvenClient
# A Recipe is a multi-agent workflow
recipe = Recipe("document-review",
steps=[
Step("planner", agent="task-planner", timeout="30s"),
Step("researcher", agent="doc-researcher", parallel=True),
Step("summarizer", agent="summarizer"),
Step("reviewer", agent="quality-reviewer"),
Step("approval", human_gate=True, notify=["team-leads"]),
],
)
# Bake the recipe via the client
client = AgentOvenClient()
client.bake(recipe, input='{"document_url": "https://..."}')
CLI Reference
The agentoven CLI provides 55+ commands across 13 command groups for complete control of your agent infrastructure.
Global Flags
--url <url> Control plane URL (env: AGENTOVEN_URL)
--api-key <key> API key (env: AGENTOVEN_API_KEY)
-k, --kitchen Kitchen/workspace scope (env: AGENTOVEN_KITCHEN)
--output <fmt> Output format: text, json, table
--help Show help for any command
Commands Overview
| Command Group | Subcommands | Description |
|---|---|---|
| agentoven init | — | Initialize a new project with agentoven.toml |
| agentoven agent | register, list, get, update, delete, bake, recook, cool, rewarm, retire, test, invoke, config, card, versions | Full agent lifecycle management |
| agentoven provider | list, add, get, update, remove, test, discover | Model provider management (OpenAI, Anthropic, Ollama, LiteLLM) |
| agentoven tool | list, add, get, update, remove | MCP tool management |
| agentoven prompt | list, add, get, update, remove, validate, versions | Versioned prompt template management |
| agentoven recipe | create, list, get, delete, bake, runs, approve | Multi-agent workflow orchestration |
| agentoven session | list, create, get, delete, send, chat | Multi-turn chat session management |
| agentoven kitchen | list, get, settings, update-settings | Workspace/tenant management |
| agentoven trace | ls, get, cost, audit | Observability, cost tracking, audit logs |
| agentoven rag | query, ingest | RAG pipeline operations |
| agentoven dashboard | — | Start the control plane + open the dashboard UI |
| agentoven login | — | Authenticate with the control plane |
| agentoven status | — | Show control plane health and agent count |
Agent Lifecycle
register → bake → ready
↓ ↑
cool → rewarm
↓
retire
| Command | Description |
|---|---|
| agentoven agent register <name> | Register a new agent (accepts --config, --framework, --model-provider, --guardrail, etc.) |
| agentoven agent bake <name> | Deploy an agent — resolves ingredients, validates config, sets status to ready |
| agentoven agent recook <name> | Hot-swap agent configuration without full redeployment |
| agentoven agent cool <name> | Pause a running agent (preserves state) |
| agentoven agent rewarm <name> | Bring a cooled agent back to ready |
| agentoven agent retire <name> | Permanently decommission an agent |
| agentoven agent invoke <name> | Run a managed agent with full agentic loop and execution trace |
| agentoven agent test <name> | One-shot or interactive playground for testing agents |
| agentoven agent card <name> | Show the A2A Agent Card (discovery metadata) |
| agentoven agent versions <name> | Show version history |
Multi-turn Sessions
# Create a session
agentoven session create my-agent
# Interactive chat with thinking mode
agentoven session chat my-agent <session-id> --t
