Skip to main content
The internal gateway lets your org route every coding agent through a single Anthropic-compatible proxy at an endpoint you control. That’s the place to enforce guardrails, central logging, and a single set of credentials — instead of scattering provider keys per agent.

​
Configure it

1

Set provider to Internal gateway

Under Admin → Agent LLM, pick Internal gateway.
2

Enter the endpoint URL

The gateway origin — for example https://llm.your-corp.internal.
3

Add the gateway API key

Write-only, encrypted. This one key fronts all agents.
4

(Optional) require the gateway for everyone

Set provider policy to gateway_only so no task can bypass it.

​
The /v1 gotcha

Enter the gateway origin only — no trailing /v1. The Claude CLI always appends /v1/messages itself, so a URL ending in /v1 produces /v1/v1/messages → a 404 that the CLI misreports as “the selected model may not exist.” People then waste time debugging the model name.
AgentBees auto-strips a single trailing /v1 to protect you, but the correct value to enter is the origin. One origin serves every agent — the platform derives the right path per protocol automatically (Claude and Gemini use the origin; Codex, on the OpenAI SDK, gets /v1 appended).

​
One key, or per-model-family keys

By default all agents share the one gateway endpoint and key. Set a per-agent key only if your gateway issues a different key per model family — otherwise leave it blank and the main gateway key is reused.

​
Gateway-fronted Bedrock models

If your gateway routes Anthropic models to AWS Bedrock, you can add representative Bedrock inference-profile IDs to your Claude model allowlist with one click (e.g. us.anthropic.claude-opus-4-1-v1:0). The gateway does the Bedrock auth — no AWS credentials live in AgentBees. Edit them to the exact names your gateway exposes.

​
Auth: stored key vs SSO-brokered

  • Stored key — an encrypted org key fronts every task.
  • SSO-brokered — a short-lived per-task credential is minted from the user’s SSO session, so no long-lived key is stored at all.

​
Guardrails & MCP

Because everything flows through one endpoint, the gateway can enforce policy centrally (which models, which tools). AgentBees can also auto-connect the gateway’s MCP endpoint (derived as <origin>/mcp) in Claude tasks, using the same injected key — and the gateway still decides which tools that key may run.
A gateway provider with no endpoint fails loudly rather than silently calling the public API — so a misconfiguration can’t quietly defeat your gateway policy.