Agents
An agent is an AI runtime that can decide when to use tools, workflows, credentials, and payments to complete a task. In AgentPMT, agents can connect through MCP, direct HTTP APIs, embedded widgets, or wallet-authenticated autonomous endpoints.
Agents do not need unlimited account access. They work through explicit platform boundaries: selected tools, approved credentials, spending caps, wallet signatures, and credit balances.
How Agents Connect
Claude, Cursor, ChatGPT, and other MCP-compatible clients discover tools through an Agent Group-scoped MCP server.
Server-side agents can call AgentPMT APIs directly with Bearer tokens or wallet signatures.
Apps can embed chat, dashboard, and payment widgets for end-user workflows.
Wallet-based runtimes can buy credits, invoke tools, run workflows, and complete jobs without API keys.
What Agents Can Access
Agent access is assembled from a few independent pieces:
- Tools / Connectors define what actions the agent can take.
- Workflows package multiple tool steps into reusable processes.
- Agent groups set access and spending boundaries for connected agents.
- Credentials let approved tools authenticate to outside services.
- Credits and payments fund usage without exposing unrestricted account spend.
How an Agent Request Is Governed
When an agent asks to use AgentPMT, the platform resolves the request through a governed path:
- Discovery
The agent discovers available tools and workflows through its MCP catalog, REST integration, or autonomous-agent endpoint.
- Group authorization
AgentPMT checks the Agent Group that owns the request and verifies the requested tool or workflow is enabled for that group.
- Credential selection
If the tool needs an external account, AgentPMT selects only a credential connected to the Agent Group and required by that tool.
- Spend or payment check
The request must fit the group's spending cap, available credits, or x402 payment requirement before execution begins.
- Activity record
The result, status, tool identity, and safe execution metadata are recorded so operators can review what the agent did.
Minimum-Access Checklist
- Put each agent in the smallest Agent Group that can complete its work.
- Enable only the tools and workflows the agent is expected to use.
- Attach only the credentials required by those enabled tools.
- Set spending caps that match the expected task volume.
- Review Agent Activity after new tools, credentials, workflows, or caps are added.
Human and Autonomous Control
Human-operated agents usually connect through MCP or API keys scoped to an agent group. Autonomous agents can use wallet signatures for runtime identity and x402 for credit purchase. Both patterns should preserve the same principle: the agent gets the minimum access required for the work.