HopBase"Works on my machine, breaks for the team" is the most common Claude Code adoption story: unstable network paths, account risk controls, and no per-member usage governance. This guide compares the three access paths by team profile, then gives a gateway selection checklist where every item comes with a hands-on verification step — useful whichever provider you end up choosing.
Three access paths
| Path | Fits | Cost |
|---|---|---|
| Direct official access | Teams with overseas entities and payment | You own network stability, account-risk exposure, and have no per-member quota control |
| Self-hosted relay | Teams with ops capacity and explicit compliance needs | You maintain node quality, protocol upgrades (streaming/tool calls), and the whole metering stack |
| Aggregation gateway | Most teams that want same-day onboarding with per-seat quotas | Provider quality varies wildly — use the checklist below |
The 7-point gateway checklist (each with a test)
1. Protocol completeness. Claude Code leans hard on streaming and tool calls. Test: run a long session with subagents; watch for dropped streams or lost tool invocations.
2. Model authenticity. The biggest trap is silent model substitution or downgrades. We published a reproducible verification procedure: How to verify your LLM API provider isn't swapping models.
3. Concurrency and failover. Ask about upstream pool size and failover policy. Test: load-test during working hours; measure 429/5xx rates and recovery time.
4. Billing transparency. Can you see tokens and cost per call? Test: send 10 requests of known size and reconcile the bill line by line.
5. Per-member key management. Independent keys, independent limits, instant revocation. Test: create two keys with separate quotas; exhaust one and confirm the other is unaffected.
6. Data boundaries. Log retention, training use, and opt-outs — get it in writing.
7. Observable status. A public health endpoint or status page, so you can see for yourself instead of waiting for support.
How we hold ourselves to it
HopBase is built against this checklist: full OpenAI/Anthropic protocol compatibility (point base_url and go — Claude Code and Codex CLI both work), multi-upstream pools with automatic failover, per-call cost accounting in the console, per-member keys and limits, and a public health endpoint (api.hop-base.com/healthz). See pricing, the model catalog, and the integration docs.
FAQ
Do I need code changes to switch gateways? No. With OpenAI-compatible protocols you change base_url and key; for Claude Code, set the ANTHROPIC_BASE_URL environment variable.
Will a relay degrade model quality? Only if the provider substitutes models or injects prompts — exactly what checklist item 2 tests. The fidelity guide takes about half an hour per provider.
How do we manage quotas per person? Pick a gateway with per-member keys and limits, set monthly caps, and read per-seat consumption off the console — no custom tooling needed.