Obol: The Control Plane for Trusted Agent Execution

Research context and background

Obol is a software platform designed to manage and secure the actions of AI agents in production environments. Think of it as a central command center that ensures every time an AI agent tries to use a tool or access data, it is safe, authorized, and tracked. The product solves a major problem where AI agents often run without oversight, leading to risks like exposed secrets, unauthorized actions, and unclear billing. Obol acts as a single gateway that handles all agent interactions, ensuring that vendor credentials are never exposed to the agent directly and that every action is logged and accountable.

Benefits

Obol offers several key advantages for teams building AI applications. First, it provides tool-aware authorization using a system called Cedar. This means the platform can check specific conditions before an action happens, such as ensuring an agent can only refund up to a certain dollar amount. Second, it creates a clear audit trail. Every tool call generates a receipt that proves what happened and who authorized it. Third, it simplifies billing. Instead of tracking token usage and tool calls in separate places, Obol combines them into one bill. Finally, it supports human approval for dangerous actions. If an agent needs to perform a destructive task, the system can pause and ask a human to approve it before the action completes.

Use Cases

Obol is ideal for companies that are deploying AI agents to perform real-world tasks. For example, a customer support team might use an AI agent to process refunds. Without Obol, the agent might accidentally refund too much money or use a secret key that gets leaked. With Obol, the system enforces a policy that limits refunds to $50 without approval. If a larger refund is needed, the system sends a link to a human manager for approval. Another use case is managing costs. Companies can set budgets for their AI usage and get a single bill that covers both the AI model costs and the tool calls made by the agent. The platform also works well for teams that need to self-host their software, as it can be deployed with a simple command and uses the customer's own security keys.

Pricing

Obol offers different pricing options depending on how the software is deployed. For teams that choose to self-host the platform, the infrastructure is free. However, there are costs associated with the actual usage of the tools. Read-only tools cost $0.30 per 1,000 calls. Idempotent writes, which are actions that can be safely repeated, cost $1.00 per 1,000 calls. Destructive tools, which can cause permanent changes, cost $3.00 per 1,000 calls. LLM requests are priced at $0.10 per 1,000 calls. It is important to note that policy denials and discovery probes do not cost anything. For hosted versions, there are Pro and Team plans available with features like SSO and assisted onboarding, while Enterprise plans offer custom contracts for larger organizations.

Vibes

While specific customer testimonials are not detailed in the available information, the product design suggests a strong focus on trust and reliability. The system uses a "one key, one policy, one bill" model that gives customers full control over their data and spending. The ability to generate receipts that separate the model's claim from the observed reality indicates a high level of transparency. The architecture ensures that vendor keys never appear in logs, which is a significant win for security-conscious teams. The inclusion of features like idempotency keys ensures that network retries do not cause duplicate charges or actions, adding another layer of reliability for production systems.

Additional Information

Obol is built with a modular architecture that supports both OpenAI-compatible APIs and the Model Context Protocol. The platform uses a virtual key system where the agent holds an Obol-minted key, while the actual vendor credentials are kept in a vault and injected only when needed. This separation ensures that sensitive keys are never exposed in the agent's configuration or logs. The system supports various trust levels for tool call results, ranging from strong gateway-observed evidence to untrusted sources. This allows teams to understand exactly how much they can trust the output of their AI agents. The platform also supports different deployment models, including self-hosted options that integrate with customer-managed key management systems like AWS or GCP.