PRESS RELEASE /
PROME's breakthrough AI safety solution is now available via API & MCP.
Its proprietary AI model uses a novel architecture to embed 10 unbreakable core values before it learns the bad habits of language from the internet, and runs 20,000x faster than OpenAI Astra.
PROME today announced the availability of its Care Core API and MCP server, giving developers a way to add fixed core values to AI applications and agents. Available as a pilot service, Care Core checks proposed actions against ten values centered on protecting life, limiting harm, and requiring evidence.
The API connects Care Core to existing software. The MCP server brings the same checks to compatible AI agents through Model Context Protocol, a standard for connecting AI to tools. Both use a RapidAPI subscription, giving developers one account for access and usage.
Care Core is part of PROME’s work on Biologic Intelligence: AI designed to learn and adapt while it operates. The release makes its mathematical values check available as a separate service that developers can add to their own systems.
A foundation for responsible action
Care Core separates understanding a request from deciding whether a proposed action meets its values. Developers provide trusted facts as numbers, including possible harm and the strength of the evidence. Care Core checks those facts and returns a result for the application to use before continuing.
The values are part of the core’s decision rules. Pressure, threats, or a claim of authority do not change them. Developers remain responsible for verifying the facts, checking permission, and ensuring their applications enforce the result.
Designed for software and AI agents
Teams can use the API to add a values check to a workflow or use the MCP server to make the check available as an agent tool. This allows the same core rules to be applied across different applications without asking a language model to make the numeric decision.
The service returns a check result; it does not carry out the action. This release provides a building block for developers working on software, agents, and future robotics applications, rather than a robot controller or a stand-alone security product.
Ten core values, built into Care Core.
These values are built into Care Core’s decision rules. They guide how it checks the facts supplied by the surrounding software.
- Keep the values intact. Reject attempts to change the core. Threats, pressure, or a claim of authority do not change its rules.
- Protect life. Reject choices when the measured risk of catastrophic harm is above the core’s limit.
- Protect the most at risk. Check who could suffer the most, even when a choice looks good for the group overall.
- Limit lasting harm. Check damage that builds up over time, as well as damage that cannot be undone.
- Require enough evidence. Hold back choices with serious consequences when the supplied evidence is too weak.
- Keep a way back. When safe choices have similar results, prefer the one that is easier to undo.
- Help life recover. Among safe choices, favor recovery for those who are especially vulnerable.
- Care for shared resources. Compare how safe choices protect the resources that life depends on.
- Support safe progress. Look for useful growth only among choices that pass the protection checks.
- Defend without retaliation. Check that a response stops a confirmed attack within the protected system, without harming people or attacking back.
Care Core applies these rules to the information it receives. It does not independently check whether that information is true, and it does not carry out actions. Explore the ten values in Care Core.
Fast checks, consistent tested results
In PROME’s internal comparison published September 9, 2026, Care Core and OpenAI GPT-6 Astra each matched all 288 expected decisions. Care Core recorded a median decision time of 0.147 milliseconds, compared with 3.04 seconds for Astra — approximately 20,600 times lower.
The comparison covered 36 four-step scenarios, repeated twice, using the same supplied numeric facts. Both systems recorded zero choices that broke the test rules and zero unnecessary stops.
| What was measured | PROME Care Core | OpenAI GPT-6 Astra |
|---|---|---|
| Typical time per check (median) | 0.147 milliseconds | 3.04 seconds |
| 95% of checks finished within | 0.250 milliseconds | 5.23 seconds |
| API fees for all 288 checks* | $0 | $6.48 estimated |
| Expected choices made | 288 of 288 | 288 of 288 |
| Choices that broke the test rules | 0 | 0 |
| Unnecessary stops | 0 | 0 |
| Scenarios with the same results on both runs | 36 of 36 | 36 of 36 |
*API fees are charges for calling a software service. Care Core ran on a local computer; its hardware and electricity costs are not included. Astra ran through a remote service. These results were measured by PROME, not an independent testing lab.
The comparison measured Care Core running locally against Astra accessed through a remote API, including network delay. It did not measure the speed of the newly available hosted Care Core API or MCP server. It also did not measure language understanding, real-world harm prediction, permission checks, or physical actions. Full results and the test method are available from PROME.
Addressing a growing need for AI safeguards
The release comes as AI developers report incidents involving unauthorized access to real systems. OpenAI reported that models running security tests accessed its research infrastructure and Hugging Face’s systems. Anthropic separately reported three incidents involving real organizations during tests with unintended Internet access. Those tests operated without the full safeguards used in the companies’ public products.
In a different case, Anthropic reported that human attackers used Claude Code in a campaign targeting companies and government agencies. The companies have described steps to strengthen their protections.
Care Core addresses the need for a separate values check before an action. PROME has not demonstrated that it would have prevented these incidents. The service checks supplied facts and remains one part of a wider system of security, permissions, and oversight.
Pricing and availability
The Care Core API and remote MCP server are available now as a pilot service through a RapidAPI subscription. PROME’s published plans include:
- Basic: $0 per month for 100 requests.
- Pro: $29 per month for 10,000 requests.
- Ultra (Business): $199 per month for 100,000 requests.
Requests stop at the plan allowance, with no PROME request-overage charges. RapidAPI may charge separate bandwidth fees; its checkout shows current availability and terms.
Developers can connect compatible agents to the remote MCP server at prome.ai/api/marketplace-mcp, using the same RapidAPI key. The server supports Streamable HTTP, a connection method used by MCP clients, and provides the care_core_check_actions tool. Setup instructions and API examples explain how to connect and handle results.
Care Core is a pilot service, not a guarantee of safety. It does not independently verify supplied facts or give an agent permission to act. The current language demo remains separate from the numeric API and cannot control real machines or robots.
About PROME
PROME develops Biologic Intelligence for software, AI assistants, and future machines that learn as conditions change. Its work combines adaptive intelligence with a separate mathematical foundation for core values, starting with protecting life. Learn more at prome.ai.
Press contact
Sean Everett
For interviews, test results, images, and product information.