Claude Code does not use the OpenAI chat endpoint. It sends Anthropic Messages requests to /v1/messages, so its base URL and credentials must use Claude Code’s Anthropic settings. GetVibe supports the Messages format for specific catalog models; GPT-6 Luna, GPT-6 Sol, and GPT-6.1 Sol are not compatible with this CLI path. This guide follows the GetVibe Claude Code setup and the existing custom endpoint reference.
Install and configure Claude Code
Install the CLI with Node.js 22 or later:
npm install -g @anthropic-ai/claude-code
mkdir -p ~/.claudeEdit ~/.claude/settings.json and use the exact fields below. The base URL intentionally has no /v1; Claude Code appends /v1/messages. The empty ANTHROPIC_API_KEY is significant: without it, Claude Code can prefer its normal direct Anthropic authentication instead of the custom endpoint.
{
"apiKeyHelper": "echo $GETVIBE_API_KEY",
"env": {
"ANTHROPIC_BASE_URL": "https://getvibeapi.com",
"ANTHROPIC_API_KEY": "",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5.5",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5.5",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4.5"
}
}Set the key in the shell before launching Claude Code: export GETVIBE_API_KEY='your-dashboard-key'. That is the environment variable used by GetVibe’s integration files. The settings file references the variable; it should not contain the actual key. Avoid putting a direct Anthropic credential in this profile because credential precedence can route around the custom endpoint.
Verify the endpoint before granting write access
Launch claude in a disposable repository and run /status. The Anthropic base URL should be https://getvibeapi.com. Ask a read-only question first, then request a small edit and inspect the diff. If the request still goes directly to Anthropic, check that ANTHROPIC_API_KEY is an empty string and that the helper can read GETVIBE_API_KEY. If you see a doubled /v1/v1/messages, remove /v1 from the base URL.
Which models work through Messages?
GetVibe’s current integration guide lists Claude Opus 5.5, Claude Sonnet 5, DeepSeek V4.1 Flash, Gemini 3.8 Flash, GLM-5.3, and Kimi K3 as examples supported on the Anthropic Messages route. Check the current model catalogue before changing a slot; protocol support is model-specific. GPT-6 Luna, GPT-6 Sol, and GPT-6.1 Sol do not work through Claude Code’s Messages endpoint. Use Codex CLI for those Responses-compatible models instead. The public catalog snapshot checked on October 3, 2026 shows Sonnet 5.5 at $0.168 input · $0.84 output per 1M tokens per million tokens; the displayed price remains live.
Assign the model slots intentionally
Claude Code supports Opus, Sonnet, and Haiku slots. The GetVibe settings example chooses current IDs for those slots, and uses CLAUDE_CODE_SUBAGENT_MODEL for subagent work. Change one slot at a time. After restarting the CLI, ask /status which model is active and run a harmless test prompt. Do not assume an OpenAI model belongs in a Claude slot: the protocol and model-family handling are different.
For cost, compare Sonnet and Opus on the same task, repository snapshot, permissions, and output request. Record all agent turns and repair steps. The live Sonnet 5.5 price is shown as $0.168 input · $0.84 output per 1M tokens per million input and output tokens. The catalog snapshot date is October 3, 2026; use the linked model page when budgeting a new session.
Keep long sessions under control
Agent sessions generate repeated context. Keep instructions concise, exclude build artifacts and vendored directories, and avoid asking the agent to reread large files when a focused file list will do. Claude requests can use cache breakpoints; GetVibe documents the caching behavior in caching docs. Keep the prefix stable and check usage fields to see whether cached input is actually reported. Do not treat a lower theoretical cache rate as savings until the session shows cache reads.
Set a clear task boundary, use normal permission prompts, and review commands and diffs. A low token price does not make an unsafe shell command safe. For repeatable benchmark comparisons, use the same permission mode and tool access for each model.
Common setup failures
- 401: verify
GETVIBE_API_KEY, the helper output, and the emptyANTHROPIC_API_KEY. - Traffic bypasses GetVibe: check credentials in your shell profile and verify the base URL using
/status. - Model unavailable: copy an exact current ID and verify Messages compatibility on its model page.
- Bad URL: set the root domain without
/v1; Claude Code adds the route suffix.
Build a cost baseline
Before rollout, run one short task and one realistic multi-file change. Record model ID, input and output usage, cache reads if reported, elapsed time, and whether tests pass. Keep the key in a secret manager and set an internal alert. These observations give the team an actual cost-per-change baseline instead of assuming every coding session consumes the same context. The current price can be checked on the Sonnet 5.5 model page.
Keep the model slots explicit
Claude Code has Opus, Sonnet, and Haiku slots. Set only the slots your team plans to use, and confirm the active choice with /status after changing settings. A Claude slot name is not a protocol adapter: an OpenAI-only model will not become Messages-compatible because it is assigned to the Sonnet slot.
Measure caching in a real session
Run one short and one multi-file task with the same stable instructions. Compare cache-read usage on later turns and record the total session cost. Keep permissions, repository snapshot, and model slot unchanged between repeated trials; otherwise the comparison cannot tell whether a cost difference came from caching, extra tool calls, or a different task.
Protect edits and credentials
Use a disposable worktree for the first write-enabled task, review commands and diffs, and keep credentials in a shell-managed secret. If a key lands in terminal history or a shared settings file, rotate it. These simple boundaries let developers explore the agent without giving every prompt access to production data or an unreviewed working tree.