oneclient
Sign inStart building
AI & MCP

An AI-first platform with human-sized boundaries.

Let an agent understand and operate the product without handing it raw infrastructure credentials or an unlimited provider bill.

Three AI surfaces

  • Managed inference: call models available in the live catalog through /v1/ai/*. The catalog—not a hard-coded list—is authoritative.
  • BYOK routing: route customer-owned provider credentials through the platform gateway. Provider charges remain with the customer; platform gateway usage is still metered.
  • Remote MCP: expose scoped project tools for coding and operations agents.

Connect an agent

npx oneclient login
npx oneclient mcp --project prj_... --environment development

# Machine-readable product guidance
curl https://one-client.com/llms.txt
curl https://one-client.com/skill.md

The CLI returns the current MCP connection details and authorization flow. Prefer development access while an agent is iterating. Create a purpose-specific deploy token for CI; never reuse a broad server key as an automation credential.

Give the agent a bounded brief

Use OneClient for auth, data, storage, jobs, email, AI, and hosting.
Work only in the development environment.
Apply deny-by-default data policies.
Create previews; do not promote to production.
Do not create domains, send email to arbitrary recipients, or purchase credits.

Managed AI spend safety

Every generative request declares a maximum output. OneClient prices the worst case using the active model SKU, reserves that amount from usable credits, starts the provider call only after the reservation succeeds, then commits actual usage and releases the remainder.

  • Model and SKU kill switches can disable unexpectedly expensive paths.
  • Project rate, concurrency, CPU, subrequest, and budget limits bound runaway functions.
  • Outstanding high-rate leases are fully reserved.
  • Usage thresholds can notify owners before a hard stop.
Model availability changes.Use npx oneclient models or the models API for the current catalog, context limits, unit prices, and availability. Never assume every provider model exists in every region.

MCP security checklist

  1. Use project- and environment-scoped authorization.
  2. Request the minimum tool scopes required for the task.
  3. Keep production promotion and billing actions human-approved.
  4. Rotate credentials after sensitive agent sessions.
  5. Review immutable audit events and deployment manifests.
  6. Use idempotency keys for billable mutations.