feat(vertex): turnkey Claude Code + Vertex compression (+ fixes from the Vertex review) (#1113)

## Description Makes **Claude Code on Google Vertex AI** actually receive Headroom's prompt compression, and fixes the issues found in a deep review of the Vertex path. The headline is a turnkey path: `headroom wrap claude` (with the user's existing Vertex env) compresses each request and forwards to Vertex using the client's own GCP ADC token — Headroom holds no credentials. _No linked issue — this addresses the internal Vertex code review (`docs/proposals/vertex-claude-compression-review.md`)._ ## Type of Change - [x] Bug fix (non-breaking change that fixes an issue) - [x] New feature (non-breaking change that adds functionality) - [ ] Breaking change (fix or feature that would cause existing functionality to change) - [x] Documentation update - [ ] Performance improvement - [ ] Code refactoring (no functional changes) ## Changes Made - `cli/wrap.py`: `wrap claude` detects `CLAUDE_CODE_USE_VERTEX=1` and points Claude Code's Vertex endpoint at the proxy via `ANTHROPIC_VERTEX_BASE_URL` (Claude Code ignores `ANTHROPIC_BASE_URL` in Vertex mode). Client keeps its own GCP ADC auth. Adds `--backend`/`--region` flags (parity with `wrap aider`). - `providers/registry.py`: alias `litellm-vertex` → provider `vertex_ai`. Previously it resolved to `"vertex"` (not in the registry) → generic pass-through with the wrong model prefix, dropped region, and mishandled auth, even though all help text advertises `litellm-vertex`. - `providers/proxy_routes.py`: derive the Vertex upstream host per-request from the path's `locations/{location}` (handles `global`) instead of pinning the configured fixed-region host; explicit `--vertex-api-url` overrides still win. - `docs/content/docs/claude-code-vertex.mdx` (+ nav): simple user guide for running Claude Code on Vertex through Headroom. - `docs/proposals/vertex-claude-compression-review.md`: the deep-review findings these fixes address. - `tests/test_vertex_claude_compression.py`: new tests. ## Testing - [x] Unit tests pass (`pytest`) - [x] Linting passes (`ruff check .`) - [x] Type checking passes (`mypy headroom`) - [x] New tests added for new functionality - [ ] Manual testing performed ### Test Output ```text $ python -m pytest tests/test_vertex_claude_compression.py -q 8 passed $ python -m pytest tests/test_provider_proxy_routes.py tests/test_provider_registry.py tests/test_cli_proxy_env.py tests/test_backend_bugs.py -q 108 passed $ ruff check headroom/providers/registry.py headroom/providers/proxy_routes.py headroom/cli/wrap.py tests/test_vertex_claude_compression.py All checks passed! $ mypy headroom/providers/registry.py headroom/providers/proxy_routes.py headroom/cli/wrap.py Success: no issues found in 3 source files ``` ## Real Behavior Proof - Environment: local macOS, Python 3.12 `.venv`, branch `feat/vertex-claude-compression`. - Exact command / steps: ran the test suite above; verified in code that the native `:rawPredict` route (publisher=anthropic) delegates to `handle_anthropic_messages` with the region-derived host, that `create_proxy_backend("litellm-vertex")` resolves to provider `vertex_ai`, and that `wrap claude` sets `ANTHROPIC_VERTEX_BASE_URL` when `CLAUDE_CODE_USE_VERTEX` is set. - Observed result: 8 new tests + 108 existing tests pass; ruff + mypy clean; the alias, region derivation (incl. `global` and explicit override), and rawPredict→compression-handler delegation all behave as asserted. - Not tested: a live end-to-end run of Claude Code against a real Google Vertex project (no GCP credentials available in this environment). Recommend one smoke test against a live Vertex project before announcing GA. The Rust `headroom-proxy` Vertex path is intentionally out of scope (separate, unwired binary). ## Review Readiness - [x] I have performed a self-review - [x] This PR is ready for human review ## Checklist - [x] My code follows the project's style guidelines - [x] I have performed a self-review of my code - [x] I have commented my code, particularly in hard-to-understand areas - [x] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] New and existing unit tests pass locally with my changes - [ ] I have updated the CHANGELOG.md if applicable ## Additional Notes - CHANGELOG.md left unchanged — releases are managed by release-please from conventional commits. - Follow-ups (not in this PR): wire or formally retire the Rust `headroom-proxy` Vertex implementation; add a live-Vertex smoke test once CI has GCP credentials.
2026-08-27 14:17:10 -04:00 · 2026-06-18 00:56:23 -07:00 · 2026-06-18 00:56:23 -07:00 · 0e0591506c
commit 0e0591506c
parent e45cf4e061
7 changed files with 582 additions and 6 deletions
--- a/docs/content/docs/claude-code-vertex.mdx
+++ b/docs/content/docs/claude-code-vertex.mdx
@ -0,0 +1,117 @@
+---
+title: Claude Code on Vertex AI
+description: Run Claude Code against Claude models on Google Vertex AI, with Headroom compressing your prompts — fewer input tokens, same answers, your own GCP login.
+---
+
+If your Claude models live on **Google Vertex AI**, you can still get Headroom's
+prompt compression. Headroom sits between Claude Code and Vertex: it shrinks the
+big stuff in each request (file reads, logs, tool output) and forwards the rest to
+Vertex using **your own Google credentials**. You keep your GCP setup; Headroom
+just makes each call cheaper.
+
+## What you get
+
+- **Fewer input tokens** on every Claude Code request to Vertex (often 30–60% on
+  agent workloads), so you pay Vertex for less.
+- **Same answers** — compression is reversible and content-aware.
+- **No new secrets** — Headroom never holds your Google credentials. Claude Code
+  keeps authenticating to Vertex with its own ADC token; Headroom passes it through.
+
+## Before you start
+
+You should already have Claude Code working against Vertex **without** Headroom.
+That means these are set in your shell:
+
+```bash
+export CLAUDE_CODE_USE_VERTEX=1
+export ANTHROPIC_VERTEX_PROJECT_ID=<your-gcp-project>
+export CLOUD_ML_REGION=us-east5          # your Vertex region (or "global")
+gcloud auth application-default login    # or set GOOGLE_APPLICATION_CREDENTIALS
+```
+
+No `ANTHROPIC_API_KEY` is needed — Vertex mode uses your Google login.
+
+## Run it (one command)
+
+```bash
+pip install headroom
+headroom wrap claude
+```
+
+That's it. Because `CLAUDE_CODE_USE_VERTEX=1` is set, `headroom wrap claude`
+automatically:
+
+1. starts the Headroom proxy,
+2. points Claude Code's Vertex endpoint at it (`ANTHROPIC_VERTEX_BASE_URL`),
+3. leaves your project, region, and Google login untouched.
+
+You'll see a line like:
+
+```
+Vertex mode: ANTHROPIC_VERTEX_BASE_URL=http://127.0.0.1:8787
+  → compress, then forward to Vertex with your GCP ADC token
+```
+
+Use Claude Code exactly as you normally would.
+
+## How it works
+
+```
+Claude Code ──(Vertex request)──▶ Headroom ──(compressed)──▶ Vertex AI (Claude)
+   in Vertex mode                  compresses                 your project + region
+   (your ADC token) ───────────── passed through ───────────▶ authenticates you
+```
+
+Claude Code sends its normal Vertex `…:rawPredict` / `:streamRawPredict` request to
+Headroom. Headroom compresses the messages (keeping the Vertex request shape
+intact), then forwards to the correct regional Vertex host — derived from the
+request itself, so multi-region and `global` both work — using the Google token
+Claude Code already attached.
+
+## Check that compression is working
+
+1. Open the dashboard: [http://localhost:8787/dashboard](http://localhost:8787/dashboard).
+   "Tokens saved" should climb as you use Claude Code.
+2. Or look at the response headers on a request: `x-headroom-tokens-before`,
+   `x-headroom-tokens-after`, `x-headroom-tokens-saved`.
+
+If "tokens saved" stays at 0 on large prompts, see Troubleshooting below.
+
+## Troubleshooting
+
+- **It still goes straight to Google (no savings).** Make sure `CLAUDE_CODE_USE_VERTEX=1`
+  is exported *in the same shell* before `headroom wrap claude`. The wrapper only
+  switches to Vertex mode when it sees that variable.
+- **Wrong region / 404 from Vertex.** Confirm `CLOUD_ML_REGION` matches a region
+  where your Claude model is enabled. `global` is supported and maps to the
+  non-regional host.
+- **Auth errors.** Headroom forwards your token as-is — if `gcloud auth
+  application-default login` (or `GOOGLE_APPLICATION_CREDENTIALS`) works for Claude
+  Code without Headroom, it works with it.
+
+## Alternative: let Headroom talk to Vertex for you
+
+If you'd rather **not** run Claude Code in Vertex mode, you can have Headroom be the
+translator instead: Claude Code speaks plain Anthropic to Headroom, and Headroom
+calls Vertex on your behalf.
+
+```bash
+export HEADROOM_BACKEND=litellm-vertex_ai     # note the _ai suffix
+export HEADROOM_REGION=us-east5
+export VERTEXAI_PROJECT=<your-gcp-project>
+export GOOGLE_APPLICATION_CREDENTIALS=/path/sa.json   # or gcloud ADC
+export ANTHROPIC_API_KEY=placeholder          # Claude Code needs *a* key to start
+headroom wrap claude --backend litellm-vertex_ai --region us-east5
+```
+
+The native Vertex-mode flow above is recommended — it keeps your existing GCP auth
+and has the smallest moving parts. Use this alternative only if you can't set
+`CLAUDE_CODE_USE_VERTEX`.
+
+## Notes
+
+- Pick a Claude model that is enabled in your Vertex project/region
+  (e.g. `claude-sonnet-4-6`, `claude-haiku-4-5`).
+- Streaming, tool use, and prompt caching all work through Headroom.
+- Want to point at a private Vertex gateway instead of Google's host? Start the
+  proxy with `--vertex-api-url https://your-gateway` and Headroom will forward there.
--- a/docs/content/docs/meta.json
+++ b/docs/content/docs/meta.json
@ -32,6 +32,7 @@
    "agno",
    "strands",
    "litellm",
+    "claude-code-vertex",
    "mcp",
    "---Configuration---",
    "configuration",
--- a/docs/proposals/vertex-claude-compression-review.md
+++ b/docs/proposals/vertex-claude-compression-review.md
@ -0,0 +1,255 @@
+# Claude Code + Vertex AI + Headroom compression — does it work?
+
+*A plain-English deep code review. Last updated 2026-06-18.*
+
+## TL;DR (read this first)
+
+**Yes, Headroom can compress Claude traffic that goes to Vertex AI — but only if you
+set it up by hand, and only if you work around one real bug.** There is **no
+turnkey path today**: `headroom wrap claude` does not know anything about Vertex,
+and the one backend flag everybody is told to use (`--backend litellm-vertex`) is
+broken.
+
+Two setups actually work (both need manual environment variables):
+
+- **Setup A — keep Claude Code in Vertex mode**, and point its Vertex URL at
+  Headroom. Headroom compresses, then forwards to real Vertex.
+- **Setup B — run Claude Code in normal Anthropic mode**, and let Headroom be the
+  translator that talks to Vertex (`--backend litellm-vertex_ai`).
+
+Everything in the *middle* (compression, request/response translation, streaming,
+tool calls) is implemented correctly. The gaps are all at the **edges**: how the
+client is pointed at Headroom, one mis-named backend, and a few env vars Headroom
+never sets for you.
+
+> Correction to an earlier claim: it is **not** true that "the Python proxy just
+> passes Vertex through without compressing." For the Anthropic publisher it runs
+> the full compression pipeline. That earlier statement was based on an incomplete
+> read of the routing code; the verified behavior is in this doc.
+
+---
+
+## The thing we're trying to do
+
+An enterprise runs **Claude Code**, but their Claude models live on **Google
+Vertex AI** (not the direct Anthropic API). They want **Headroom** in the middle so
+their prompts get compressed (fewer input tokens = lower cost), without changing
+the answers.
+
+For that to happen, three things must all be true:
+
+1. **The client's traffic must actually reach Headroom** (the proxy must be in the path).
+2. **Headroom must compress it.**
+3. **Headroom must forward it to Vertex correctly** (right URL, right auth, right body shape) and translate the answer back so Claude Code understands it.
+
+This review checks all three.
+
+---
+
+## The map: where Claude-on-Vertex can run, and what compresses
+
+There are **two proxies** in this repo and **three** possible routes. Only some compress.
+
+| Route | What it is | Compresses? | Notes |
+|---|---|---|---|
+| **Python proxy, native Vertex `:rawPredict`** (publisher = `anthropic`) | Client sends a real Vertex request to Headroom | ✅ **Yes** | Runs the full Anthropic compression pipeline, keeps the Vertex body shape, forwards the client's own Google token. `proxy_routes.py:648` |
+| **Python proxy, `--backend litellm-vertex_ai`** | Client speaks plain Anthropic; Headroom translates to Vertex | ✅ **Yes** (correct string only) | Full Anthropic↔Vertex translation incl. streaming + tools. **`litellm-vertex` is broken — must use `litellm-vertex_ai`.** |
+| **Rust proxy, native Vertex `:rawPredict`** | A separate `headroom-proxy` binary | ✅ **Yes** | Correct and well-built — **but never run by `headroom proxy`/`wrap`.** Dead code for normal users. |
+| **Python proxy, passthrough** (any *other* publisher) | Generic verbatim forward | ❌ No | Only used for non-Anthropic, non-Google publishers. `openai.py:6014` |
+
+**Key takeaway:** the *compression engine* for Vertex+Claude exists and works in the
+Python proxy. The problems are getting traffic into it and one naming bug.
+
+---
+
+## Does it work end-to-end? The honest answer
+
+**Through `headroom wrap claude` with zero extra setup: no.** `wrap claude` only
+sets `ANTHROPIC_BASE_URL`. If Claude Code is in Vertex mode it ignores that and
+talks straight to Google — Headroom is never in the path. And `wrap claude` has no
+`--backend`/`--region` flags and sets no Vertex environment variables.
+
+**With manual setup: yes, one of two ways.** Both are below. Both work *around*
+issues, not because the product wires them for you.
+
+---
+
+## Setup A — Claude Code stays in Vertex mode (recommended for Vertex shops)
+
+Idea: Claude Code keeps using its native Vertex mode and its own Google login.
+You just tell it "send Vertex requests to Headroom instead of straight to Google,"
+and you tell Headroom where the real Vertex endpoint is.
+
+```bash
+# 1) Run Headroom, telling it the real Vertex endpoint (match your region!)
+headroom proxy --port 8787 \
+  --vertex-api-url https://us-east5-aiplatform.googleapis.com   # use YOUR region
+
+# 2) Run Claude Code in Vertex mode, but point its Vertex base URL at Headroom
+export CLAUDE_CODE_USE_VERTEX=1
+export ANTHROPIC_VERTEX_PROJECT_ID=<your-gcp-project>
+export CLOUD_ML_REGION=us-east5
+export ANTHROPIC_VERTEX_BASE_URL=http://127.0.0.1:8787   # <-- the redirect that makes it work
+claude
+```
+
+What happens: Claude Code → `ANTHROPIC_VERTEX_BASE_URL` (Headroom) → Headroom
+matches the `:rawPredict` route, sees `publisher=anthropic`, **compresses**, then
+forwards to the real Vertex endpoint using Claude Code's own Google token.
+
+Caveats: you must set `--vertex-api-url` to your region (see Issue #6), and
+`wrap claude` won't set `ANTHROPIC_VERTEX_BASE_URL` for you (Issue #3).
+
+---
+
+## Setup B — Claude Code in normal Anthropic mode; Headroom talks to Vertex
+
+Idea: Claude Code thinks it's talking to plain Anthropic. Headroom holds the Google
+credentials and is the one that actually talks to Vertex.
+
+```bash
+# Headroom does the Vertex talking — note the backend name carefully
+export HEADROOM_BACKEND=litellm-vertex_ai        # NOT "litellm-vertex" (that's broken — Issue #1)
+export HEADROOM_REGION=us-east5                  # becomes the Vertex location
+export VERTEXAI_PROJECT=<your-gcp-project>        # Headroom does NOT set this for you (Issue #4)
+export GOOGLE_APPLICATION_CREDENTIALS=/path/sa.json   # or use `gcloud auth application-default login`
+export ANTHROPIC_API_KEY=placeholder-not-used    # Claude Code needs *a* key to start (Issue #5)
+
+# Do NOT set CLAUDE_CODE_USE_VERTEX here — Claude Code must stay in normal mode
+headroom wrap claude
+```
+
+What happens: Claude Code → Headroom (plain Anthropic `/v1/messages`) →
+**compresses** → LiteLLM converts to Vertex and calls Claude on Vertex → converts
+the answer back to Anthropic shape → Claude Code reads it.
+
+Caveats: the backend-name bug (Issue #1), the missing project env (Issue #4), and
+this path has **no automated tests** (Issue #8) — smoke-test it before relying on it.
+
+---
+
+## How to verify compression is really happening
+
+1. Open the dashboard: `http://localhost:8787/dashboard` — "tokens saved" should
+   climb as you use Claude Code.
+2. Or check response headers on a request: `x-headroom-tokens-before`,
+   `x-headroom-tokens-after`, `x-headroom-tokens-saved`.
+3. Confirm it actually hit Vertex (proxy logs show a `vertex_ai/claude-…` model or
+   a Vertex host, not `api.anthropic.com`).
+
+If `tokens-saved` is 0 on large prompts, compression isn't running — re-check the
+setup against the issues below.
+
+---
+
+## Every issue we found (the full list)
+
+Severity: **BROKEN** = doesn't work; **GAP** = works only with manual workaround;
+**BUG** = wrong behavior in an edge case; **HOUSEKEEPING** = confusing but harmless.
+
+### 1. BROKEN — `--backend litellm-vertex` never reaches Vertex
+The backend name is turned into a provider by chopping off `litellm-`, so
+`litellm-vertex` becomes the provider `vertex`. But the Vertex integration is keyed
+on `vertex_ai`, not `vertex`. So Headroom falls back to a generic "unknown
+provider" mode: it builds the wrong model name (`vertex/claude-…` instead of
+`vertex_ai/claude-…`), **ignores the region**, and mishandles auth.
+**You must use `--backend litellm-vertex_ai`.** Worse: every help message and the
+`wrap` example tell users the broken `litellm-vertex`.
+*Where:* `providers/registry.py:174-178`, `backends/litellm.py:291,326-336,681-682`;
+help text at `cli/proxy.py:524`, `cli/wrap.py:3645`, `proxy/server.py:3913`.
+*Fix (small):* alias `vertex` → `vertex_ai` in `create_proxy_backend`, or add a
+`"vertex"` entry to the provider registry. Then fix the help text.
+
+### 2. GAP — `headroom wrap claude` has no Vertex support
+The `claude` wrap command has no `--backend` and no `--region` (the `aider` wrap
+command has both). It only ever sets `ANTHROPIC_BASE_URL`. So there's no flag to
+turn on a Vertex backend for Claude Code — you must pre-export env vars.
+*Where:* `cli/wrap.py:2780-2819` (vs `cli/wrap.py:3612,3615` for aider).
+*Fix:* add `--backend`/`--region` to `wrap claude`, mirroring `aider`.
+
+### 3. GAP — Vertex-mode Claude Code bypasses the proxy, and Headroom never sets the fix
+With `CLAUDE_CODE_USE_VERTEX=1`, Claude Code ignores `ANTHROPIC_BASE_URL` and goes
+straight to Google. There **is** a documented override — `ANTHROPIC_VERTEX_BASE_URL`
+— that points Claude Code's Vertex traffic at a gateway. But Headroom never sets it
+(0 references in the codebase). So the proxy has the right routes, but nothing
+connects the client to them automatically.
+*Where:* repo-wide grep for `ANTHROPIC_VERTEX_BASE_URL` = 0 hits.
+*Fix:* in a Vertex-aware `wrap claude`, set `ANTHROPIC_VERTEX_BASE_URL` to the proxy.
+
+### 4. GAP — the GCP project is never passed to LiteLLM
+For Setup B, Headroom passes the region to LiteLLM but never the project. You must
+export `VERTEXAI_PROJECT` (or `GOOGLE_CLOUD_PROJECT`) yourself or it fails.
+*Where:* `backends/litellm.py:682` sets only `vertex_location`.
+*Fix:* thread a project config/env through to the LiteLLM call.
+
+### 5. GAP — no placeholder API key for Claude Code
+With a custom `ANTHROPIC_BASE_URL`, Claude Code needs *an* `ANTHROPIC_API_KEY` to
+start, even though the proxy uses Google creds upstream. `wrap claude` never sets a
+placeholder, so the user must.
+*Where:* `cli/wrap.py:2984-2988`.
+*Fix:* set a placeholder key (or `ANTHROPIC_AUTH_TOKEN`) when launching.
+
+### 6. BUG — region/host mismatch
+Headroom pins the Vertex host to one region (default `us-central1`) but throws away
+the region in the client's request path. If your client targets, say,
+`europe-west1` while Headroom is on the default host, the request goes to the wrong
+region unless you set `--vertex-api-url` to match.
+*Where:* `copilot_auth.py:936` (host = base + path, no region reconciliation),
+`providers/proxy_routes.py:647` (path `location` is discarded), `registry.py:16`.
+*Fix:* derive the upstream host from the request path's `location`, or validate they match.
+
+### 7. HOUSEKEEPING — the Rust Vertex proxy is correct but unwired
+There's a second, Rust proxy (`crates/headroom-proxy/src/vertex/`) that compresses
+Vertex traffic correctly. But `headroom proxy` and `headroom wrap` run the **Python**
+server and never call it — it's a separate binary you'd run by hand. This is a
+frequent source of "but I thought Vertex compression was added" confusion: it was,
+in Rust, on a path nobody runs by default.
+*Where:* `crates/headroom-proxy/Cargo.toml` (`[[bin]]`), no Python→Rust bridge to it.
+*Fix:* either document that the Rust proxy is separate, or wire/retire it.
+
+### 8. GAP — no tests for the Vertex compression paths
+No test instantiates the LiteLLM Vertex backend with a mocked Vertex call, and the
+native-Vertex compression route isn't covered against a real Vertex shape. The
+translation code is correct by inspection, but unproven by CI.
+*Fix:* add a mocked round-trip test (Anthropic in → compressed → Vertex call asserted → Anthropic out).
+
+### 9. HOUSEKEEPING — stale Rust doc comment
+`crates/headroom-proxy/src/vertex/mod.rs:42-47` describes a "synthetic model
+injection" strategy the code no longer implements. Doc only; behavior is correct.
+
+---
+
+## What's actually solid (so we don't over-correct)
+
+These were verified and are **correct**:
+
+- **Native Vertex `:rawPredict` for `publisher=anthropic` compresses** and preserves
+  the Vertex body shape (keeps `anthropic_version`, never injects `model`).
+  `proxy_routes.py:648`, `anthropic.py:604-606,1941-1949`.
+- **LiteLLM translation is real and complete** (with the `vertex_ai` provider):
+  response is rebuilt into Anthropic shape (`backends/litellm.py:575-628`), streaming
+  emits proper Anthropic SSE events (`streaming.py:1344-1472`, `litellm.py:736-947`),
+  and tool calls round-trip both directions (`litellm.py:525-565,593-602`).
+- **Compression runs before the backend dispatch** (`anthropic.py:1671` then
+  `:1781`), so the backend always gets the compressed body.
+- **Auth is forwarded correctly** on the native path — the client's Google token is
+  passed through untouched (`copilot_auth.py:1156-1157`).
+
+---
+
+## Recommended fixes, smallest-first
+
+1. **Fix the backend name (Issue #1)** — one-line alias `vertex`→`vertex_ai`, then
+   correct the help text. This is the highest-impact, lowest-effort fix; it turns the
+   *documented* command from broken to working.
+2. **Add `--backend`/`--region` to `wrap claude` (Issue #2)** — copy from `aider`.
+3. **Add a Vertex mode to `wrap claude` (Issues #3, #5)** — detect/set
+   `ANTHROPIC_VERTEX_BASE_URL` → proxy, set a placeholder API key, and configure the
+   proxy's Vertex upstream — so Setup A becomes one command.
+4. **Pass the GCP project (Issue #4)** and **reconcile the region/host (Issue #6).**
+5. **Add a mocked round-trip test (Issue #8).**
+6. **Decide the Rust proxy's fate (Issue #7)** — document-as-separate or wire it in.
+
+After #1–#3, the honest customer message becomes: *"`headroom wrap claude` works
+with Vertex out of the box."* Until then, it's *"works with a documented manual
+setup."*
--- a/headroom/cli/wrap.py
+++ b/headroom/cli/wrap.py
@ -2814,6 +2814,18 @@ def unwrap() -> None:
        "ENABLE_TOOL_SEARCH env var is respected."
    ),
 )
+@click.option(
+    "--backend",
+    default=None,
+    help="API backend for the proxy: 'anthropic' (default), 'litellm-vertex_ai', etc. "
+    "(env: HEADROOM_BACKEND). For Vertex, prefer CLAUDE_CODE_USE_VERTEX=1 (native, "
+    "keeps your GCP auth) over a litellm backend.",
+)
+@click.option(
+    "--region",
+    default=None,
+    help="Cloud region for Vertex/Bedrock backends (env: HEADROOM_REGION).",
+)
@click.option("--verbose", "-v", is_flag=True, help="Verbose output")
@click.option("--prepare-only", is_flag=True, hidden=True)
@click.argument("claude_args", nargs=-1, type=click.UNPROCESSED)
@ -2827,6 +2839,8 @@ def claude(
    learn: bool,
    memory: bool,
    tool_search: str | None,
+    backend: str | None,
+    region: str | None,
    verbose: bool,
    prepare_only: bool,
    claude_args: tuple,
@ -2928,6 +2942,16 @@ def claude(
        if os.environ.get("CLAUDE_CODE_USE_FOUNDRY"):
            foundry_upstream = os.environ.get("ANTHROPIC_FOUNDRY_BASE_URL")

+        # Detect Vertex mode: with CLAUDE_CODE_USE_VERTEX=1, Claude Code IGNORES
+        # ANTHROPIC_BASE_URL and authenticates to Google Vertex with GCP ADC. The
+        # documented way to route its Vertex :rawPredict / :streamRawPredict
+        # traffic through a gateway is ANTHROPIC_VERTEX_BASE_URL. Point it at
+        # Headroom and the proxy compresses the request, then forwards to the
+        # real regional Vertex host (derived per-request from the path's
+        # location) using Claude Code's own ADC token — no API key, no creds held
+        # by Headroom. This is the turnkey Vertex compression path.
+        use_vertex = bool(os.environ.get("CLAUDE_CODE_USE_VERTEX"))
+
        proxy_holder[0] = _ensure_proxy(
            port,
            no_proxy,
@ -2935,6 +2959,8 @@ def claude(
            memory=memory,
            agent_type="claude",
            code_graph=code_graph,
+            backend=backend,
+            region=region,
            anthropic_api_url=foundry_upstream,
        )

@ -2970,7 +2996,12 @@ def claude(
        proxy_url = _claude_proxy_base_url(port)
        click.echo()
        click.echo("  Launching Claude Code (API routed through Headroom)...")
-        if foundry_upstream:
+        if use_vertex:
+            click.echo(
+                f"  Vertex mode: ANTHROPIC_VERTEX_BASE_URL={proxy_url} "
+                "→ compress, then forward to Vertex with your GCP ADC token"
+            )
+        elif foundry_upstream:
            click.echo(
                f"  Foundry mode: ANTHROPIC_FOUNDRY_BASE_URL={proxy_url} → upstream {foundry_upstream}"
            )
@ -2982,7 +3013,12 @@ def claude(
        click.echo()

        env = os.environ.copy()
-        if foundry_upstream:
+        if use_vertex:
+            # Claude Code stays in Vertex mode (keeps CLAUDE_CODE_USE_VERTEX,
+            # ANTHROPIC_VERTEX_PROJECT_ID, CLOUD_ML_REGION, ADC — all inherited);
+            # we only redirect its Vertex endpoint to Headroom.
+            env["ANTHROPIC_VERTEX_BASE_URL"] = proxy_url
+        elif foundry_upstream:
            env["ANTHROPIC_FOUNDRY_BASE_URL"] = proxy_url
        else:
            env["ANTHROPIC_BASE_URL"] = proxy_url
--- a/headroom/providers/proxy_routes.py
+++ b/headroom/providers/proxy_routes.py
@ -28,6 +28,27 @@ def _api_target(proxy: Any, provider_name: str) -> str:
    return cast(str, getattr(proxy, legacy_attr, proxy.provider_runtime.api_target(provider_name)))


+def _vertex_target_for_location(proxy: Any, location: str) -> str:
+    """Resolve the Vertex upstream host for a request, region-aware.
+
+    The Vertex regional host must match the ``locations/{location}`` in the
+    request path (e.g. a ``europe-west1`` request cannot go to a
+    ``us-central1`` host). The configured target is a single fixed-region host
+    (default ``us-central1``), so unless the operator pinned an explicit
+    non-default upstream (e.g. a private gateway), derive the host from the
+    request's own location. ``global`` maps to the unprefixed host.
+    """
+    from headroom.providers.registry import DEFAULT_VERTEX_API_URL
+
+    configured = _api_target(proxy, "vertex")
+    if configured and configured != DEFAULT_VERTEX_API_URL:
+        # Operator pinned an explicit upstream (gateway / specific host) — honor it.
+        return configured
+    if not location or location == "global":
+        return "https://aiplatform.googleapis.com"
+    return f"https://{location}-aiplatform.googleapis.com"
+
+
 def _select_passthrough_base_url(proxy: Any, headers: dict[str, str]) -> str:
    # Codex CLI subscription mode hits a wide surface under
    # `/backend-api/*` (rate-limit polling, agent identity, JWT
@ -644,11 +665,11 @@ def register_provider_routes(app: FastAPI, proxy: Any) -> None:
        publisher: str,
        model: str,
    ):
-        del api_version, project, location
+        del api_version, project
        if publisher == "anthropic":
            return await proxy.handle_anthropic_messages(
                request,
-                _api_target(proxy, "vertex"),
+                _vertex_target_for_location(proxy, location),
                "vertex:anthropic",
                model,
            )
@ -665,11 +686,11 @@ def register_provider_routes(app: FastAPI, proxy: Any) -> None:
        publisher: str,
        model: str,
    ):
-        del api_version, project, location
+        del api_version, project
        if publisher == "anthropic":
            return await proxy.handle_anthropic_messages(
                request,
-                _api_target(proxy, "vertex"),
+                _vertex_target_for_location(proxy, location),
                "vertex:anthropic",
                model,
                True,
--- a/headroom/providers/registry.py
+++ b/headroom/providers/registry.py
@ -173,6 +173,12 @@ def create_proxy_backend(

    normalized_backend = backend if backend.startswith("litellm-") else f"litellm-{backend}"
    provider = normalized_backend.replace("litellm-", "")
+    # `litellm-vertex` is the name in our docs/help, but LiteLLM (and our
+    # provider registry) keys Google Vertex on `vertex_ai`. Without this alias
+    # the provider falls through to a generic pass-through: wrong model prefix
+    # (`vertex/…` instead of `vertex_ai/…`), region dropped, auth mishandled.
+    if provider in ("vertex", "google-vertex", "googlevertex"):
+        provider = "vertex_ai"
    try:
        backend_cls = litellm_backend_cls or _load_litellm_backend()
        instance = cast("Backend", backend_cls(provider=provider, region=bedrock_region))
--- a/tests/test_vertex_claude_compression.py
+++ b/tests/test_vertex_claude_compression.py
@ -0,0 +1,140 @@
+"""Turnkey Claude Code + Vertex compression wiring.
+
+Covers the fixes that let `headroom wrap claude` + Vertex actually deliver
+compression:
+
+- `litellm-vertex` -> `vertex_ai` provider alias (registry),
+- the Vertex upstream host derived per-request from the path's `location`
+  (so a europe-west1 request is not sent to a us-central1 host),
+- the native `:rawPredict` route running the compression handler (not the
+  verbatim passthrough) for the `anthropic` publisher.
+
+No real GCP/Vertex is contacted — handlers and the backend class are stubbed.
+"""
+
+from __future__ import annotations
+
+import logging
+import types
+from typing import Any
+
+from fastapi.responses import JSONResponse
+from fastapi.testclient import TestClient
+
+from headroom.providers import proxy_routes, registry
+from headroom.providers.registry import DEFAULT_VERTEX_API_URL
+from headroom.proxy.server import HeadroomProxy, ProxyConfig, create_app
+
+
+# --------------------------------------------------------------------------
+# Region-aware Vertex upstream host
+# --------------------------------------------------------------------------
+def _stub_proxy(vertex_url: str) -> Any:
+    # `_api_target` reads `proxy.VERTEX_API_URL`, but its getattr default eagerly
+    # evaluates `proxy.provider_runtime.api_target(...)`, so the stub needs both.
+    return types.SimpleNamespace(
+        VERTEX_API_URL=vertex_url,
+        provider_runtime=types.SimpleNamespace(api_target=lambda name: vertex_url),
+    )
+
+
+def test_vertex_target_derives_region_from_location_on_default() -> None:
+    proxy = _stub_proxy(DEFAULT_VERTEX_API_URL)
+    assert (
+        proxy_routes._vertex_target_for_location(proxy, "europe-west1")
+        == "https://europe-west1-aiplatform.googleapis.com"
+    )
+
+
+def test_vertex_target_global_uses_unprefixed_host() -> None:
+    proxy = _stub_proxy(DEFAULT_VERTEX_API_URL)
+    assert (
+        proxy_routes._vertex_target_for_location(proxy, "global")
+        == "https://aiplatform.googleapis.com"
+    )
+
+
+def test_vertex_target_empty_location_uses_unprefixed_host() -> None:
+    proxy = _stub_proxy(DEFAULT_VERTEX_API_URL)
+    assert (
+        proxy_routes._vertex_target_for_location(proxy, "") == "https://aiplatform.googleapis.com"
+    )
+
+
+def test_vertex_target_honors_explicit_override() -> None:
+    # An operator who pinned a non-default upstream (private gateway) wins,
+    # regardless of the path's location.
+    proxy = _stub_proxy("https://vertex-gateway.internal")
+    assert (
+        proxy_routes._vertex_target_for_location(proxy, "europe-west1")
+        == "https://vertex-gateway.internal"
+    )
+
+
+# --------------------------------------------------------------------------
+# litellm-vertex -> vertex_ai provider alias
+# --------------------------------------------------------------------------
+def _capture_provider(backend: str) -> dict[str, Any]:
+    captured: dict[str, Any] = {}
+
+    class FakeLiteLLM:
+        def __init__(self, provider: str, region: str | None = None) -> None:
+            captured["provider"] = provider
+            captured["region"] = region
+
+    registry.create_proxy_backend(
+        backend=backend,
+        anyllm_provider="openai",
+        bedrock_region="us-east5",
+        logger=logging.getLogger("test-vertex"),
+        litellm_backend_cls=FakeLiteLLM,
+    )
+    return captured
+
+
+def test_litellm_vertex_aliases_to_vertex_ai() -> None:
+    # The documented `litellm-vertex` must reach the `vertex_ai` provider, not
+    # a generic pass-through (which would drop the region and mangle the model).
+    assert _capture_provider("litellm-vertex")["provider"] == "vertex_ai"
+
+
+def test_litellm_vertex_ai_unchanged() -> None:
+    assert _capture_provider("litellm-vertex_ai")["provider"] == "vertex_ai"
+
+
+def test_litellm_bedrock_not_aliased() -> None:
+    assert _capture_provider("litellm-bedrock")["provider"] == "bedrock"
+
+
+# --------------------------------------------------------------------------
+# Native :rawPredict route → compression handler with region-derived host
+# --------------------------------------------------------------------------
+def _default_vertex_app() -> Any:
+    # No vertex_api_url override -> default us-central1 host -> route derives
+    # the regional host from the request path.
+    return create_app(ProxyConfig(optimize=True, cache_enabled=False, rate_limit_enabled=False))
+
+
+def test_vertex_rawpredict_anthropic_runs_compression_handler(monkeypatch) -> None:
+    captured: dict[str, str] = {}
+
+    # The route calls handle_anthropic_messages(request, base_url, provider, model).
+    async def fake(self, request, base_url, provider, model, *rest):  # type: ignore[no-untyped-def]
+        captured.update(base_url=str(base_url), provider=str(provider), model=str(model))
+        return JSONResponse({"ok": True})
+
+    monkeypatch.setattr(HeadroomProxy, "handle_anthropic_messages", fake)
+
+    with TestClient(_default_vertex_app()) as client:
+        resp = client.post(
+            "/v1/projects/p/locations/europe-west1/publishers/anthropic/models/"
+            "claude-sonnet-4-6:rawPredict",
+            json={"anthropic_version": "vertex-2023-10-16", "messages": []},
+        )
+    assert resp.status_code == 200
+    # The anthropic publisher is routed to the compression handler (not the
+    # verbatim passthrough), with the region-derived upstream host (exact match,
+    # not a substring check).
+    assert captured["provider"] == "vertex:anthropic"
+    assert captured["base_url"] == "https://europe-west1-aiplatform.googleapis.com"
+    assert captured["model"] == "claude-sonnet-4-6"