diff --git a/.env.example b/.env.example
index a6768c5..664e97b 100644
--- a/.env.example
+++ b/.env.example
@@ -1,13 +1,13 @@
-# --- AI provider (required) ---
-# deepsec dispatches AI agents to investigate candidates. Two backends are
-# supported, both routed through Vercel AI Gateway by default.
+# --- AI provider (required for full scans) ---
+# Recommended: use one Vercel AI Gateway key for both supported backends.
+AI_GATEWAY_API_KEY=
 
-# Claude Agent SDK (default backend, --agent claude-agent-sdk)
-ANTHROPIC_AUTH_TOKEN=
-ANTHROPIC_BASE_URL=https://ai-gateway.vercel.sh
+# Direct Anthropic fallback for the Claude Agent SDK (--agent claude-agent-sdk).
+# Leave these unset when using AI_GATEWAY_API_KEY or local `claude` login.
+ANTHROPIC_API_KEY=
+ANTHROPIC_BASE_URL=
 
-# OpenAI Codex SDK (--agent codex). Optional; the gateway accepts the
-# Anthropic token at /v1/chat/completions too, so you can leave OPENAI_*
-# unset and Codex will fall back to ANTHROPIC_AUTH_TOKEN.
+# OpenAI Codex SDK (--agent codex). Leave these unset when using
+# AI_GATEWAY_API_KEY or local `codex` login.
 OPENAI_API_KEY=
-OPENAI_BASE_URL=https://ai-gateway.vercel.sh/v1
+OPENAI_BASE_URL=
diff --git a/README.md b/README.md
index 659e489..8403b3d 100644
--- a/README.md
+++ b/README.md
@@ -82,8 +82,9 @@ AI_GATEWAY_API_KEY=vck_...
 
 See [docs/vercel-setup.md](docs/vercel-setup.md) for getting a key and
 for the Vercel Sandbox setup. To bypass the gateway, set
-`ANTHROPIC_AUTH_TOKEN` + `ANTHROPIC_BASE_URL` (or the OpenAI pair)
-explicitly. Explicit values always win over the `AI_GATEWAY_API_KEY`
+`ANTHROPIC_API_KEY` (plus `ANTHROPIC_BASE_URL=https://api.anthropic.com`
+only when you need an explicit direct Anthropic base URL) or the OpenAI
+pair explicitly. Explicit values always win over the `AI_GATEWAY_API_KEY`
 expansion.
 
 If a `process` or `revalidate` run halts because the upstream credential
diff --git a/docs/configuration.md b/docs/configuration.md
index e0cb6ba..b8ef651 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -105,8 +105,9 @@ backend you're using.
 | Var | Used by | Purpose |
 |---|---|---|
 | `AI_GATEWAY_API_KEY` | all AI commands | Shortcut. Expands at CLI startup into `ANTHROPIC_AUTH_TOKEN` / `OPENAI_API_KEY` / `ANTHROPIC_BASE_URL` / `OPENAI_BASE_URL` — one key covers both Claude and Codex through Vercel AI Gateway. Any of those four vars set explicitly takes precedence. Falls back to `VERCEL_OIDC_TOKEN` (from `vercel env pull`) when unset, so users already authenticated for Sandbox don't need a separate gateway key. |
-| `ANTHROPIC_AUTH_TOKEN` | `process`, `revalidate`, `triage` (Claude backend) | API token for the Claude Agent SDK. AI Gateway-issued or Anthropic-issued. Set this if you don't use `AI_GATEWAY_API_KEY`. |
-| `ANTHROPIC_BASE_URL` | same | Default (when `AI_GATEWAY_API_KEY` is set): `https://ai-gateway.vercel.sh`. Set to `https://api.anthropic.com` for direct Anthropic. |
+| `ANTHROPIC_AUTH_TOKEN` | `process`, `revalidate`, `triage` (Claude backend) | Gateway-routed token for the Claude Agent SDK. Usually filled from `AI_GATEWAY_API_KEY`; set explicitly only for gateway-compatible proxies that expect this variable. |
+| `ANTHROPIC_API_KEY` | same | Direct Anthropic API key. Set this, not `ANTHROPIC_AUTH_TOKEN`, when bypassing the gateway and talking to Anthropic directly. |
+| `ANTHROPIC_BASE_URL` | same | Default (when `AI_GATEWAY_API_KEY` is set): `https://ai-gateway.vercel.sh`. Optional for direct Anthropic; set to `https://api.anthropic.com` when you need the base URL to be explicit. |
 
 ### Optional
 
diff --git a/docs/faq.md b/docs/faq.md
index bfd88ec..5db9e6e 100644
--- a/docs/faq.md
+++ b/docs/faq.md
@@ -70,12 +70,11 @@ the gateway.
 
 ```bash
 # Direct Anthropic
-ANTHROPIC_AUTH_TOKEN=sk-ant-...
+ANTHROPIC_API_KEY=sk-ant-...
 ANTHROPIC_BASE_URL=https://api.anthropic.com
 
 # AI Gateway (recommended)
-ANTHROPIC_AUTH_TOKEN=vck_...
-ANTHROPIC_BASE_URL=https://ai-gateway.vercel.sh
+AI_GATEWAY_API_KEY=vck_...
 ```
 
 If `claude` or `codex` is already logged in on this machine, non-sandbox
diff --git a/docs/getting-started.md b/docs/getting-started.md
index bf27969..0b4c7e0 100644
--- a/docs/getting-started.md
+++ b/docs/getting-started.md
@@ -37,8 +37,9 @@ Open `.env.local` and pick one of:
   expires after 12 hours; re-pull when you hit auth errors. Convenient
   if you're already using Vercel Sandbox (same token unlocks both).
 
-Prefer Anthropic directly? Set `ANTHROPIC_AUTH_TOKEN=sk-ant-…` and
-`ANTHROPIC_BASE_URL=https://api.anthropic.com` instead. If `claude` or
+Prefer Anthropic directly? Set `ANTHROPIC_API_KEY=sk-ant-…` instead.
+Add `ANTHROPIC_BASE_URL=https://api.anthropic.com` only if you need the
+base URL to be explicit. If `claude` or
 `codex` is already logged in on this machine, non-sandbox runs
 (`process` / `revalidate` / `triage`) skip the token and reuse the
 subscription. See [vercel-setup.md](vercel-setup.md).
diff --git a/docs/vercel-setup.md b/docs/vercel-setup.md
index d3b8fd1..05d0c8e 100644
--- a/docs/vercel-setup.md
+++ b/docs/vercel-setup.md
@@ -62,7 +62,7 @@ If the second command fails with `Missing AI credentials` or a `401`, see [Troub
 
 ### How it works
 
-deepsec expands whichever credential it finds (the API key first, the OIDC token as fallback) at startup into the four vars the agent SDKs read (`ANTHROPIC_AUTH_TOKEN`, `OPENAI_API_KEY`, `ANTHROPIC_BASE_URL`, `OPENAI_BASE_URL`), so a single credential covers both Codex (`--agent codex`, the default) and Claude (`--agent claude`).
+deepsec expands whichever gateway credential it finds (the API key first, the OIDC token as fallback) at startup into the four vars the agent SDKs read (`ANTHROPIC_AUTH_TOKEN`, `OPENAI_API_KEY`, `ANTHROPIC_BASE_URL`, `OPENAI_BASE_URL`), so a single gateway credential covers both Codex (`--agent codex`, the default) and Claude (`--agent claude`). For direct Anthropic, set `ANTHROPIC_API_KEY` instead.
 
 Any of those four vars you set explicitly takes precedence over the expansion — useful for mixing direct Anthropic with gateway-routed OpenAI, etc.
 
@@ -113,11 +113,11 @@ If you have your own Anthropic / OpenAI agreement, two options:
 
 **Through the gateway (recommended)** — configure [Bring Your Own Key (BYOK)](https://vercel.com/docs/ai-gateway/authentication-and-byok#bring-your-own-key-byok) at the team level. No gateway markup, with failover and observability on top.
 
-**Bypass the gateway entirely** — set the explicit base URL + token pairs in `.env.local`:
+**Bypass the gateway entirely** — set direct provider credentials in `.env.local`:
 
 ```bash
 # Anthropic direct
-ANTHROPIC_AUTH_TOKEN=sk-ant-…
+ANTHROPIC_API_KEY=sk-ant-…
 ANTHROPIC_BASE_URL=https://api.anthropic.com
 
 # OpenAI direct
@@ -125,7 +125,7 @@ OPENAI_API_KEY=sk-…
 # (OPENAI_BASE_URL defaults to api.openai.com — only set it for proxies)
 ```
 
-Mix freely — gateway for Claude, direct for OpenAI, etc. The explicit values always win over the `AI_GATEWAY_API_KEY` expansion.
+Mix freely — gateway for Claude, direct Anthropic, direct OpenAI, etc. The explicit values always win over the `AI_GATEWAY_API_KEY` expansion.
 
 ---
 
diff --git a/packages/deepsec/src/__tests__/credential-brokering.test.ts b/packages/deepsec/src/__tests__/credential-brokering.test.ts
index 5f08dcc..62b32d5 100644
--- a/packages/deepsec/src/__tests__/credential-brokering.test.ts
+++ b/packages/deepsec/src/__tests__/credential-brokering.test.ts
@@ -3,6 +3,7 @@ import { buildSandboxEnv, resolveBrokeredCredentials } from "../sandbox/setup.js
 
 const TOUCHED_KEYS = [
   "ANTHROPIC_AUTH_TOKEN",
+  "ANTHROPIC_API_KEY",
   "ANTHROPIC_BASE_URL",
   "OPENAI_API_KEY",
   "OPENAI_BASE_URL",
@@ -34,6 +35,20 @@ describe("credential brokering", () => {
       expect(c.openaiToken).toBeUndefined();
     });
 
+    it("captures direct ANTHROPIC_API_KEY from the orchestrator env", () => {
+      process.env.ANTHROPIC_API_KEY = "sk-ant-real";
+      const c = resolveBrokeredCredentials("claude-agent-sdk");
+      expect(c.anthropicToken).toBe("sk-ant-real");
+      expect(c.openaiToken).toBeUndefined();
+    });
+
+    it("prefers ANTHROPIC_AUTH_TOKEN over direct ANTHROPIC_API_KEY", () => {
+      process.env.ANTHROPIC_AUTH_TOKEN = "vck_real";
+      process.env.ANTHROPIC_API_KEY = "sk-ant-real";
+      const c = resolveBrokeredCredentials("claude-agent-sdk");
+      expect(c.anthropicToken).toBe("vck_real");
+    });
+
     it("falls back to ANTHROPIC token for OpenAI on the codex path", () => {
       process.env.ANTHROPIC_AUTH_TOKEN = "vck_real";
       const c = resolveBrokeredCredentials("codex");
diff --git a/packages/deepsec/src/__tests__/preflight.test.ts b/packages/deepsec/src/__tests__/preflight.test.ts
index 02915f1..48c537b 100644
--- a/packages/deepsec/src/__tests__/preflight.test.ts
+++ b/packages/deepsec/src/__tests__/preflight.test.ts
@@ -32,12 +32,14 @@ describe("assertAgentCredential", () => {
   beforeEach(() => {
     saved = {
       ANTHROPIC_AUTH_TOKEN: process.env.ANTHROPIC_AUTH_TOKEN,
+      ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY,
       OPENAI_API_KEY: process.env.OPENAI_API_KEY,
       CLAUDE_HOME: process.env.CLAUDE_HOME,
       CODEX_HOME: process.env.CODEX_HOME,
       PATH: process.env.PATH,
     };
     delete process.env.ANTHROPIC_AUTH_TOKEN;
+    delete process.env.ANTHROPIC_API_KEY;
     delete process.env.OPENAI_API_KEY;
     // Point CLAUDE_HOME / CODEX_HOME and PATH at empty tmp dirs so the
     // suite is hermetic — the dev running tests may have a real
@@ -65,9 +67,14 @@ describe("assertAgentCredential", () => {
     expect(() => assertAgentCredential("claude-agent-sdk")).not.toThrow();
   });
 
+  it("passes for claude-agent-sdk when direct Anthropic API key is set", () => {
+    process.env.ANTHROPIC_API_KEY = "x";
+    expect(() => assertAgentCredential("claude-agent-sdk")).not.toThrow();
+  });
+
   it("throws actionable message for claude-agent-sdk when no token and no claude CLI", () => {
     expect(() => assertAgentCredential("claude-agent-sdk")).toThrow(/--agent claude/);
-    expect(() => assertAgentCredential("claude-agent-sdk")).toThrow(/ANTHROPIC_AUTH_TOKEN/);
+    expect(() => assertAgentCredential("claude-agent-sdk")).toThrow(/ANTHROPIC_API_KEY/);
     expect(() => assertAgentCredential("claude-agent-sdk")).toThrow(/AI_GATEWAY_API_KEY/);
     expect(() => assertAgentCredential("claude-agent-sdk")).toThrow(
       /https:\/\/github\.com\/vercel-labs\/deepsec\/blob\/main\/docs\/vercel-setup\.md/,
diff --git a/packages/deepsec/src/preflight.ts b/packages/deepsec/src/preflight.ts
index 30ef742..c27e7e7 100644
--- a/packages/deepsec/src/preflight.ts
+++ b/packages/deepsec/src/preflight.ts
@@ -162,6 +162,7 @@ export function assertAgentCredential(
   if (agentType !== undefined && !KNOWN_BACKENDS.has(agentType)) return;
 
   const anthropic = process.env.ANTHROPIC_AUTH_TOKEN;
+  const anthropicApiKey = process.env.ANTHROPIC_API_KEY;
   const openai = process.env.OPENAI_API_KEY;
 
   if (isCodex(agentType)) {
@@ -177,14 +178,14 @@ export function assertAgentCredential(
     );
   }
 
-  if (anthropic) return;
+  if (anthropic || anthropicApiKey) return;
   if (!options.inSandbox && hasLocalClaudeAgent()) return;
   const displayAgent =
     agentType === "claude-agent-sdk" || agentType === undefined ? "claude" : agentType;
   throw new Error(
     `Missing AI credentials for --agent ${displayAgent}.\n` +
       `\n` +
-      `  Add to .env.local:    AI_GATEWAY_API_KEY=vck_…   (or ANTHROPIC_AUTH_TOKEN=…)\n` +
+      `  Add to .env.local:    AI_GATEWAY_API_KEY=vck_…   (or ANTHROPIC_API_KEY=… for direct Anthropic)\n` +
       `  Setup: ${SETUP_DOC_URL}`,
   );
 }
diff --git a/packages/deepsec/src/sandbox/setup.ts b/packages/deepsec/src/sandbox/setup.ts
index ac2f3f4..1108f43 100644
--- a/packages/deepsec/src/sandbox/setup.ts
+++ b/packages/deepsec/src/sandbox/setup.ts
@@ -37,7 +37,7 @@ export interface UploadBundle {
 // --- Sandbox env vars ---
 
 // Routing / debug knobs the agent reads inside the sandbox. AI credential
-// env vars (ANTHROPIC_AUTH_TOKEN, OPENAI_API_KEY, AI_GATEWAY_API_KEY) are
+// env vars (ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY, OPENAI_API_KEY, AI_GATEWAY_API_KEY) are
 // deliberately absent — they're brokered via firewall header injection
 // (see resolveBrokeredCredentials + buildWorkerNetworkPolicy below) so a
 // compromised in-VM agent can't read them out of /proc/<pid>/environ.
@@ -66,11 +66,11 @@ interface BrokeredCredentials {
 /**
  * Resolve the orchestrator-side AI credentials that will be brokered into
  * the sandbox. AI Gateway issues one token per team that authenticates both
- * Claude and OpenAI traffic, so when only ANTHROPIC_AUTH_TOKEN is set and
- * the worker is going to run codex, fall it back as the OpenAI token.
+ * Claude and OpenAI traffic, so when only an Anthropic token is set and the
+ * worker is going to run codex, fall it back as the OpenAI token.
  */
 export function resolveBrokeredCredentials(agentType: string | undefined): BrokeredCredentials {
-  const anthropicToken = process.env.ANTHROPIC_AUTH_TOKEN;
+  const anthropicToken = process.env.ANTHROPIC_AUTH_TOKEN ?? process.env.ANTHROPIC_API_KEY;
   const explicitOpenai = process.env.OPENAI_API_KEY;
   // Only borrow ANTHROPIC for OPENAI on the codex path — and only when the
   // user hasn't pinned an explicit OpenAI key. Outside codex this fallback