KefuAgents DocsAgent Gateway Overview
Bring Your Own Bot: one MCP endpoint plus one key connects any agent to KefuAgents's supervised support loop.
You bring the agent, we keep the gate
The KefuAgents Agent Gateway is the BYOB (Bring Your Own Bot) surface: the platform provides context (threads, orders, knowledge, the forbidden-commitments list), your agent reasons with its own model, and the platform safety gate re-checks every draft. Reads and draft submissions consume zero platform AI credits.
The supervision loop:
get_pending_cards— see the pending decision cards;get_reply_context— fetch the full reply context for one thread (includingcontextVersionand theconstraintsforbidden-commitments list);- Draft with your own model →
submit_draft— the safety gate re-checks the draft; accepted drafts become pending human review (never auto-sent), rejections return namedviolations; - High-risk decisions (refund / replacement / complaint / chargeback /
safety) always stay human — the correct agent action is
create_review_requestor telling the merchant.
Endpoint & authentication
| Purpose | URL |
|---|---|
| MCP endpoint (Streamable HTTP) | https://kefuagents.com/api/gateway/mcp |
| REST base | https://kefuagents.com/api/gateway/v1 |
| OpenAPI description | https://kefuagents.com/gateway/openapi.json |
- One credential:
Authorization: Bearer <gateway key>. Keys are created in the KefuAgents settings panel, shown once; one key binds exactly one workspace (brand). - MCP tool names map 1:1 onto the OpenAPI
operationIds — 17 tools. - Protocol revision 2026-07-28. Clients on the earlier revisions are
still served unchanged: the server answers both
server/discoverand the olderinitializehandshake, so an agent configured before this change needs no edit. - Every request is authenticated independently (stateless); key revocation takes effect immediately.
Scopes
| Scope | Grants | Default |
|---|---|---|
gateway:read | the 5 read tools + webhook management | ✅ on |
gateway:submit_draft | submit drafts + the 4 non-approve write tools | ✅ on |
gateway:approve | approve/reject pending replies | ❌ off, explicit merchant opt-in |
Platform guarantees (structural — not dependent on agent obedience)
- The gate re-checks every draft; violating drafts are rejected with named violations.
- High-risk decisions are always made by humans — no scope combination
bypasses this (
HIGH_RISK_REQUIRES_HUMAN). - One key can only ever reach one workspace.
- Every call — including denials — is audited; the settings panel shows each key's recent calls.