Coral Bricks Client Agent Endpoint Handoff (Onboarding)

Coral Bricks Client Agent Endpoint Handoff (Onboarding)

1. 01

   Issue a scoped cb_ API key per client from the Coral Bricks dashboard

   Create one cb_... key per client engagement rather than sharing a single agency-wide key. This keeps token spend attributable when a client asks which agent loop burned the budget, and it lets you revoke one client's access without touching other delivery work.

2. 02

   Send OpenAI wire-format requests to https://inference.coralbricks.ai/v1 against a named model slug

   Pin the exact slug in the handoff doc, for example coralbricks/glm-5.3-fp4 or coralbricks/deepseek-v4.1-flash-fast-fp4. Never hand a client a generic 'Coral Bricks endpoint' note; the slug is the contract, and swapping it later changes both latency and per-token cost.

3. 03

   Confirm streaming tool calls and 1M-token context handling on a representative agent loop before go-live

   Run the client's actual multi-step plan, not a hello-world prompt. Coral Bricks publishes setup complexity as medium, so budget onboarding effort without promising a fixed day count to the client.

4. 04

   Point the client's existing agent tool at the endpoint

   OpenCode, Codex CLI, GitHub Copilot, Cursor, Cline, Continue, and aider all route through the OpenAI-compatible surface with minimal configuration. Reuse whatever tool the client already runs instead of forcing a migration during onboarding.

5. 05

   Document the caching behavior in the client runbook

   Coral Bricks offers free cached reads, so repeated context in a long agent loop costs less than a cold read. Tell the client which parts of their prompt are stable context (cache-friendly) and which change every call.

6. 06

   Record the pay-per-token rates the client is being billed against

   GLM 5.3 input is $1.12 per 1M tokens and cache write is $1.68 per 1M tokens. Put these figures in the handoff sheet so the retainer conversation about usage is grounded in published rates, not estimates.

7. 07

   Flag the dedicated-capacity path for clients with committed throughput needs

   The 'Own your AI' plan runs on the client's VPC with private KV storage, custom models, and resource-based pricing rather than per-token. It is contact-sales only, so raise it early if the client's compliance team will not accept shared inference.

8. 08

   Hand the client a one-page escalation path back to the agency

   Name who owns model slug changes, key rotation, and usage reviews. Coral Bricks is an inference API, not a white-label product, so the client must understand the agency owns the integration layer and the vendor owns the endpoint.