Connect Claude Code

Claude Code speaks the Anthropic Messages API, so pointing it at ForceAI is a configuration change, not a code change: you set a base URL and a ForceAI key, and every turn flows through the gateway instead of straight to Anthropic. Guardrails, spend and audit apply automatically because it is ordinary gateway traffic

What can and cannot be routed

The gateway only governs clients that speak the Anthropic API to a configurable endpoint. That is Claude Code and anything using the SDK, not the subscription surfaces

Surface	Through the gateway?
Claude Code CLI (`claude` in a terminal)	Yes
Claude Code in an IDE extension (VS Code, JetBrains)	Yes
The Code tab inside the Claude Desktop app	Yes
The Claude Desktop chat app	No
`claude.ai` in a browser	No

The desktop chat app and claude.ai web are the subscription product. They talk to Anthropic directly and have no configurable endpoint, so nothing can route them through ForceAI. Govern those through the Claude Enterprise admin console instead. Everything a developer runs as Claude Code can be routed.

Point Claude Code at the gateway

Two settings matter. The base URL is the gateway origin with no /v1 (Claude Code appends /v1/messages itself), and the credential is an ordinary ForceAI virtual key sent as a bearer token

Choose where the config lives

Use a shell export for a one-off session in a single terminal, or ~/.claude/settings.json to apply it to every Claude Code session on the machine, including the desktop Code tab. See Session vs persistent

Set the base URL, key and models

Claude Code otherwise defaults to Sonnet/Opus model IDs. Pin it to a model the gateway serves, or register the IDs Claude Code sends

export ANTHROPIC_BASE_URL=https://your-gateway        # origin, NO /v1
export ANTHROPIC_AUTH_TOKEN=sk-your-forceai-key       # sent as Bearer
export ANTHROPIC_MODEL=claude-haiku-4-5               # a registered model
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5 # background tasks too
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
claude

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://your-gateway",
    "ANTHROPIC_AUTH_TOKEN": "sk-your-forceai-key",
    "ANTHROPIC_MODEL": "claude-haiku-4-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Run it

Start claude. Every turn now posts to your-gateway/v1/messages and appears in the dashboard under Logs with its guardrail trace

Setting only ANTHROPIC_BASE_URL with no credential variable leaves Claude Code on your saved claude.ai subscription login. The key (or apiKeyHelper) is what actually moves it onto the gateway, and it bills through the gateway’s upstream, not your subscription seat.

Session vs persistent config

A shell export is scoped to that one terminal: it affects only the claude you launch there and its children. It never reaches the desktop app, which the OS launches without your shell environment ~/.claude/settings.json is Claude Code’s own config. It applies wherever Claude Code runs, including the desktop Code tab. Settings are read at startup, so after editing the file, restart the Claude Desktop app or start a new session to pick it up. This file configures Claude Code only; it has no effect on the desktop chat app

Register the models

The gateway fails closed on an unrecognized model and the error names the exact id to register, so the loop is: run Claude Code once, read the rejected id, register it. Add it in the dashboard under Models + Endpoints -> Add Model, or:

curl -X POST "$FORCEAI_URL/model/new" \
  -H "Authorization: Bearer $FORCEAI_ADMIN_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model_name":"claude-sonnet-4-5-20250929",
       "forceai_params":{"model":"anthropic/claude-sonnet-4-5",
                          "api_key":"os.environ/ANTHROPIC_API_KEY"}}'

Pinning ANTHROPIC_MODEL and the ANTHROPIC_DEFAULT_*_MODEL variables to already registered names is the quick alternative. Note that background tasks use the Haiku model, so register (or pin) it even if nobody selects it

Verify it is governed

curl -s -o /dev/null -w '%{http_code}\n' \
  "$FORCEAI_URL/v1/messages" \
  -H "Authorization: Bearer $FORCEAI_KEY" \
  -H 'anthropic-version: 2023-06-01' \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-haiku-4-5","max_tokens":10,
       "messages":[{"role":"user","content":"hi"}]}'

A 200 confirms the gateway serves the Anthropic endpoint for the model. Then run Claude Code and open the request in the dashboard under Logs. The guardrail trace on that row is the proof the session is governed rather than merely proxied

A pre_call guardrail sits in the latency path of every turn. An interactive coding tool feels per-turn latency far more than a batch caller does, so measure it on a real session before a wide rollout.

Roll it out to a team

For a fleet, enforce the gateway centrally with Claude Code managed settings rather than relying on each developer’s shell or user settings. An org owner pushes the same env block from the claude.ai admin console (server-managed settings, no device management) or through MDM, and allowedProviders: ["customEndpoint"] makes the pinned base URL mandatory so a developer cannot point Claude Code elsewhere. See the enterprise rollout runbook in the repository for the full pattern, including the provisioning of per-developer keys and the direct-to-Anthropic side channels to close

Get started

LLM providers

A2A agents

Skills

Knowledge Catalog

Single sign-on

Admin settings

Tools

Guardrails

Command line

Coding tools

What can and cannot be routed

Point Claude Code at the gateway

Session vs persistent config

Register the models

Verify it is governed

Roll it out to a team