Neon AI Gateway

A credential is valid on the branch it was created on and on every branch descended from it, so one credential created on main covers preview and feature branches forked from main. It is not valid on a branch outside that lineage, where the gateway returns credential not authorized for this branch with a 403.

To rotate, create the new credential first, update the environment, then revoke the old one.

Neon exposes the OpenAI Responses API under /openai/v1 rather than /v1. Most OpenAI models answer on both endpoints, but a few are served only through Responses and are rejected by chat completions. Check Neon's endpoint guide and Responses model list before choosing a model. Models that require Responses need the configuration below.

Reach them with a second provider entry using the openai-responses API and that base URL, which is the same branch host you substituted above with /openai/v1 on the end:

{  models: {    providers: {      "neon-responses": {        baseUrl: "https://<your-neon-branch-host>/openai/v1",        apiKey: "${NEON_AI_GATEWAY_TOKEN}",        api: "openai-responses",        models: [          {            id: "gpt-5-3-codex",            name: "GPT-5.3 Codex",            reasoning: true,            input: ["text", "image"],            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },            contextWindow: 400000,          },        ],      },    },  },}

This endpoint takes OpenAI model IDs only. Sending a Gemini, Llama or Qwen ID returns 400 model "<model-id>" is not available on the openai_responses endpoint.

Neon also exposes the Anthropic Messages API under /anthropic, where the Anthropic SDK appends /v1/messages. This dialect is Claude only, and it reaches native Anthropic features that the OpenAI-compatible /v1 path cannot express, such as extended thinking and prompt caching. Set cacheRetention explicitly to choose retention, as the configuration below shows. Reach it with a second provider entry using the anthropic-messages API and the /anthropic base URL, the same shape as the bundled Anthropic-compatible provider examples:

{  models: {    providers: {      "neon-anthropic": {        baseUrl: "https://<your-neon-branch-host>/anthropic",        apiKey: "${NEON_AI_GATEWAY_TOKEN}",        authHeader: true,        api: "anthropic-messages",        models: [          {            id: "claude-sonnet-4-6",            name: "Claude Sonnet 4.6",            reasoning: true,            input: ["text", "image"],            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },            contextWindow: 200000,          },        ],      },    },  },  agents: {    defaults: {      model: { primary: "neon-anthropic/claude-sonnet-4-6" },      models: {        "neon-anthropic/claude-sonnet-4-6": {          params: { cacheRetention: "short" },        },      },    },  },}

The same cost, maxTokens and contextWindow caveats from the OpenAI entry apply here: set cost from the Neon model catalog, and take contextWindow from that catalog as well.

Set authHeader: true because Neon requires a bearer credential, including on this Anthropic-compatible endpoint.

Set params.cacheRetention under agents.defaults.models, as the example above does, to choose caching behavior for this custom anthropic-messages route. Use "short" for the 5-minute cache or "long" for the 1-hour TTL; "none" disables it. See Prompt caching.

Two behaviors matter here:

• This endpoint takes Claude model IDs only. Sending a non-Anthropic ID returns 400 model "<model-id>" is not available on the anthropic_messages endpoint, so keep this entry separate from the /v1 entry, which reaches every model.
• On a non-api.anthropic.com host, OpenClaw suppresses implicit anthropic-beta headers, such as interleaved thinking, so a proxy does not reject them. Set headers with an anthropic-beta value on the provider if you need a specific beta.

Neon enforces account-level token-per-minute rate limits. Going over returns 429 Too Many Requests with a message naming the model. Upstream output-token limits also apply, so a 429 can happen on output tokens alone.

Neon also enforces an account-level spend cap that returns 429 with error code REQUEST_LIMIT_EXCEEDED.

For most models message.content is a plain string. Some models return an array of typed content blocks instead; see Neon's content-shape guidance. OpenClaw's openai-completions transport flattens those blocks on the way in, treating text blocks as assistant output and reasoning blocks as thinking, so nothing in your config has to handle the array form.

A low output-token limit can cut a response off before the text block appears. If the response contains reasoning but no answer, check the finish reason and output-token budget before raising maxTokens.

• Custom reasoning models marked reasoning: true support OpenAI-style reasoning controls. With openai-completions, selecting a thinking level can send reasoning_effort to Neon. If the upstream model rejects that field, set compat.supportsReasoningEffort: false on that model entry.
• Inference uses prepaid credits on a paid plan. Neon charges the model provider's per-token rate with no markup. See the AI Gateway overview.

Neon's troubleshooting guide covers the errors you are most likely to hit:

• For account-access errors, check the model's enabled field in GET /v1/models. If your project has a paid plan and prepaid credits but access is disabled, contact Neon support. For credential errors, check the ai_gateway:invoke scope and branch lineage.
• 400 model "<model-id>" is not available on the <endpoint> endpoint means the model ID does not belong to the dialect you called, such as an OpenAI ID sent to /anthropic or a Claude ID sent to /openai/v1. Use the provider entry that matches the model.
• AI Gateway is generally available on paid plans and is not available in every region, so a project created in an unsupported region or on a free plan cannot reach it. See Regions for current coverage.