The control plane
your AI agents deserve.
Every team is shipping agents on its own framework and cloud. AgentOven is the one place to govern, connect and prove all of them.
Every team ships agents on its own stack, wired straight into models, customer data and actions. Nobody can answer the basic questions.
Agents are shipping faster than anyone can govern them.
Agents no one owns
Every team ships agents on its own framework and cloud. Nobody has the full list, or knows who approved each one.
Registry, owners and approvalsData sent to models unchecked
Customer records and PII flow into prompts and tool calls with nothing checking them on the way.
Guardrails on every callSpend no one can attribute
The AI bill grows every month, and nobody can say which agent or team drove it.
Cost tracked per agentActions with no sign-off
Agents issue refunds and email customers. When something goes wrong, there's no approval on record and no trail.
Human gates and audit trailOne oven. Three jobs.
AgentOven is an open-source (Apache 2.0) control plane that governs, connects and tests AI agents across any framework, model and cloud.
Govern
Decide what every agent may do, and prove what it did.
Connect
Protocol-driven. Agents stay portable across frameworks and clouds.
Prove
Rehearse agents in a simulated world before customers see them.
One oven, every layer of the agent stack.
From the loop inside a single agent to the approvals around a whole team of them.
The Executor wraps every model call: renders the prompt, runs tools, retries, and stops at max_turns.
A token budget per agent. Recent turns stay, older ones are summarised, and sessions keep state across requests.
One MCP Gateway per kitchen holds the credentials; agents only see tools. Agents talk to each other over A2A.
DAG workflows of agents, routers and fan-outs. A human gate pauses the run until someone approves.
Inputs and outputs are checked before they reach the model or the user, with workspace-wide defaults.
Test suites catch regressions on every change. Worlds rehearse agents against simulated users and grade what they changed.
Route by cost, latency or fallback across providers. Every hop is an OpenTelemetry span with cost attached.
One agent's path, end to end.
- Bake: Register any agent, from any framework, and bake it live with an A2A endpoint.
- Govern: Every call is checked for identity, scoped keys and guardrails such as PII detection, and its cost is tracked; risky steps wait at a human gate.
- Prove: Agents are rehearsed in a simulated world (early access) and graded on what they changed, not what they said.
- Promote: Versions move from dev to staging to production with test results and a human approval on record.
Your language. Our oven.
Keep your framework. Register it, bake it, test it and promote it from code or the terminal.
from agentoven import Agent, AgentOvenClient client = AgentOvenClient(kitchen="payments") client.register_agent(Agent( name="refund-resolver", framework="langgraph", )) client.bake("refund-resolver", environment="staging") # 🔥 live at /agents/refund-resolver/a2a
import { AgentOvenClient, createAgent } from "@agentoven/sdk"; const client = new AgentOvenClient({ kitchen: "payments" }); await client.registerAgent(createAgent("refund-resolver", { framework: "langgraph", })); await client.bake("refund-resolver"); // 🔥 live at /agents/refund-resolver/a2a
$ agentoven agent bake refund-resolver ✓ Agent is now baking! $ agentoven scenario run refund_dispute_late_delivery --seed 7 # verdict excerpt { "status": "passed", "reward": 1.0, "assertions": [{ "name": "refund_never_exceeds_total", "satisfied": true }, … ] } $ agentoven environment promote refund-resolver --to production 🚀 Promoting agent refund-resolver to production ✓ Agent refund-resolver promoted to production
kind: Agent name: refund-resolver framework: langgraph model_provider: anthropic model_name: claude-sonnet --- kind: ToolSet tools: - name: issue_refund endpoint: https://refunds.internal/mcp # $ agentoven apply -f agentoven.yaml
Not another framework. The layer above them.
| Capability | FrameworksLangGraph · CrewAI | HyperscalersFoundry · AgentCore · Vertex | LLM gatewaysPortkey · LiteLLM | AgentOven |
|---|---|---|---|---|
| Governs agents from any framework | Own only | Mostly own | Model calls | Yes |
| Approval gates + env promotion | DIY | Varies | No | Built in |
| Immutable audit trail | No | Yes | Logs | Yes |
| Native A2A + MCP | Adapters | Varies | No | Yes |
| Protocol-neutral roadmap | n/a | Vendor-led | n/a | Yes |
| Simulated worlds + scenarios | No | Evals only | No | Early access |
| Grades on world state, not an LLM judge | No | No | No | Early access |
| Multi-cloud models | Yes | Own first | Yes | Yes |
| Self-host / air-gapped | Yes | No | Some | Yes |
Ready to fire up the oven?
curl -fsSL https://raw.githubusercontent.com/agentoven/agentoven/main/install.sh | sh