MPAS - A new primitive for software approvals
Don't let a single actor, agent or human, unilaterally:
- Delete a production database.
- Violate compliance with a social post.
- Exfiltrate data.
- Push to production.
- Take a large trading position.
MPAS puts independent, verifiable approval between an intention and its consequences, and keeps the credentials out of reach of the actor.
The most consequential actions should not be unilateral decisions.
MPAS is for multi-party approvals what OAuth is for delegated authorization. Today, one credential is often enough to act alone. Systems that do ask for approval invent their own way.
Give AI agents autonomy, not unilateral authority.
Agents handle routine work. Consequential steps wait for independent approval.
One phished account shouldn't be enough to compromise your organization.
Critical actions can require multiple approvals, so one compromised account is not enough to act.
Propose. Approve. Verify. Execute.
Routine actions go straight through.
01. Propose
An agent or person proposes the exact operation, target, and parameters. They never receive the credential needed to run it.
Keep the stack you have. Just add approvals.
Identity, policy engines, and agent runtimes coordinate your actions. MPAS decides which ones can run.
Identity and access
OAuth, OpenID Connect, enterprise IAM, DIDs, OMATrust
Policy engines
OPA, Cedar, custom rules
Agent runtimes
MCP, agent control hooks
Signing
Keys, passkeys, HSMs, MPC, smart-contract wallets
Applications
APIs, cloud infrastructure, local apps, blockchains
Help make multi-party approvals a common building block for software.
Join OMA3
Shape MPAS and the open standards around it as an OMA3 member.
Establish a liaison
Align MPAS with your work through a formal liaison agreement.
Contribute
Review the specification, surface issues, and contribute code.
- An approved OMA3 open-source project
- Code under Apache 2.0
- Specifications under the OMA3 IPR Policy
- Draft specification