Peeve over MCP, both directions.

Claude, ChatGPT and other OpenAI agents, Hermes, OpenClaw, and any client that speaks MCP. The user points their agent at a scoped Peeve endpoint and it receives your capabilities as typed tools, each stating its effect before it runs. Nothing about the agent is special-cased; if it speaks the protocol, it connects.

Each user gets a personal, scoped endpoint of the form mcp.peeve.ai/<workspace>/u/<contact>. They add that URL in their agent (the same way they would add any custom MCP connector) and the agent discovers your capabilities as typed tools. The endpoint is minted per user, so access is theirs to hold and yours to revoke.

Your capabilities, rendered as typed tools with a dry-run effect string, for example change_plan, download_invoice or list_orders. It is the same capability registry that renders as a cursor and narration for a human. You define a capability once and it serves both the person and their agent, never maintained twice.

Exactly the calling user's permissions, and no more. An agent acting for a user can do what that user can do, nothing they cannot. Each grant is scoped to that user, revocable by them and visible to you, and rate-limited per grant.

Yes. Every grant is revocable and rate-limited per user. A user can revoke their agent's access at any time, and you can see and manage grants. Growth includes 5 grants per user; Scale is unlimited.

Yes. Anything that changes state is confirmed, actions run within a workspace allow-list, and every action writes a signed audit entry. The model requests a capability by name and never holds a credential, so an injected instruction can request a capability but never compose a raw request or reach a key.

Yes, on the Growth plan and above. Paste your MCP server URL, and an access token if it needs one. Peeve does a live handshake, discovers the tools your server exposes, and you choose which the assistant may use. Anything that changes state runs through the same confirmation guardrail.

Peeve initializes a client connection, lists the tools your server offers, and shows them to you with read or write marked on each. You enable the ones the assistant may call. Reads run live during a conversation; writes are confirmed before they run, and credentials are encrypted per workspace.

Yes. A read tool runs automatically to answer with live data. A write tool, like issue_refund, states its effect and runs only after approval, inside an allow-list, and is written to a signed audit log. This holds for both your custom MCP server and the capabilities you expose to agents.

No. Agent tool calls over MCP are logged but not billed today. Grants are limited by plan (5 per user on Growth, unlimited on Scale), and if the billing model changes you get a full billing cycle's notice, never a retroactive charge.

Agents need structured actions, explicit permissions and machine-readable context, not a chat window. Peeve renders one capability registry two ways: cursor motion and narration for people, and typed MCP tools with a stated effect for agents. Both get the same capabilities in the form each can use.

Peeve publishes a server card at /.well-known/mcp/server-card.json and an agent guide at /AGENTS.md, and every page has a Markdown twin (append .md to any URL, or send Accept: text/markdown). These let agents and crawlers read the product directly.

Give agents trustworthy, permissioned, action-ready access to your product. Peeve provides that today: a scoped MCP endpoint where a user's agent gets your capabilities as typed tools, with the same confirmation and audit guardrails as the human path, so an agent can transact on a user's behalf safely.