Coding CLI Integrations

The fastest-growing surface for enterprise LLM usage in 2026 is coding CLIs: every engineer has at least one running locally, and the gap between personal and corporate use has become a governance crisis. The AI Gateway solves it, every engineer keeps their preferred CLI, the org controls cost, guardrails, and visibility. Setup is always the same two steps:

Mint a LangWatch virtual key (see Quickstart).
Set the CLI’s base URL and API key to the gateway + VK.

The gateway exposes both OpenAI-compatible and Anthropic-compatible endpoints on the same port, so any CLI that speaks either dialect works unchanged.

Debugging a CLI sign-in or fold pipeline? See AI Governance → CLI Debug: langwatch governance status, langwatch ingest tail, the langwatch login error catalog (incl. 409 no_default_routing_policy), and OCSF probe.

At a glance

The langwatch CLI ships wrappers for the 4 most common coding CLIs, they auto-inject the right env vars from your governance config so you don’t have to. Native env-var setup still works everywhere if you prefer.

CLI	Wrapper	Endpoint	Native env vars	Notes
Claude Code	`langwatch claude`	`/v1/messages`	`ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`	Tool-call deltas preserved byte-for-byte. Pin to dated model name (e.g. `claude-haiku-4-5-20251001`).
Codex CLI	`langwatch codex`	`/v1/responses`	`~/.codex/config.toml` `model_providers`, `wire_api = "responses"`	Codex 0.122+ requires `wire_api = "responses"` (chat dropped). Pin model to a Bifrost-registered name.
Gemini CLI	`langwatch gemini`	`/v1beta/models/{model}:generateContent`	`GOOGLE_GEMINI_BASE_URL`, `GEMINI_API_KEY`	Native Gemini API passthrough. Works with all Gemini models supported by your VK’s Vertex/Gemini credential.
Cursor	`langwatch cursor` *	`/v1/chat/completions`	Custom “OpenAI API base URL” in settings	* Cursor is a GUI app, its AI panel reads from settings, not env vars, so the wrapper is mostly cosmetic for terminal launches. The canonical setup is the in-app paste. Agent mode benefits most from budgets.
opencode	,	`/v1/chat/completions` (or `/v1/messages`)	Per-provider config in `opencode.json`	No wrapper today, set the per-provider config explicitly. Pin opencode `1.13.x`, `1.14.x` has a known regression with custom providers.
Aider	,	`/v1/chat/completions`	`OPENAI_API_BASE`, `OPENAI_API_KEY`	No wrapper today, manual env-var path. Confirmed working with model aliases.

Why this matters for the enterprise

Before the gateway, governance of coding CLIs was a choice between:

Ban them: kills productivity, drives shadow usage.
Allow them with personal provider keys: no visibility, no cost control, leaked credentials in dotfiles and CI logs.

The gateway adds a third option: allow every CLI, centrally governed.

Cost. Each engineer’s CLI spend debits the org → team → project → principal budgets you’ve set.
Visibility. Every CLI call shows up in LangWatch traces, scoped to the project the VK belongs to.
Policy. policy_rules can deny shell-exec tools or untrusted MCPs at the gateway level, even if the CLI would otherwise enable them.
Portability. An engineer on Claude Code and a co-worker on Codex hit the same gateway with different VKs but the same budget, you don’t need to pick a winner.
Revocation. Rotate or revoke a VK and the CLI stops working globally within 60 seconds. No more “which laptops still have the old key?”

Recommended VK layout for coding CLIs

A workable pattern used by several early customers:

One personal-access VK per engineer ({engineer}-cli) bound to the engineer’s principal.
Attach a principal-scoped monthly budget (e.g. $200/month for engineers, $1000/month for staff+). on_breach: block.
policy_rules.tools.deny: ^shell\\.exec$, ^filesystem\\.write$ (or your org’s list).
Fallback chain: Anthropic → OpenAI → Azure OpenAI. CLI autoswitches on outage.
cache.mode: respect so Anthropic prompt caching keeps saving 90%.

Then every engineer gets a one-time setup (env vars in their shell rc) and never touches provider keys again.

Cleaning up a machine

langwatch logout reverses the setup. It revokes and clears your device session, then removes the telemetry wiring the wrappers installed so a plain claude, codex, gemini, or opencode stops sending traces. It only touches the blocks LangWatch wrote (the OTEL env keys in ~/.claude/settings.json, the [otel] block in ~/.codex/config.toml, and the scoped shell functions in your shell rc), leaving the rest of each file untouched. It scans first, shows you what it found, and asks before removing anything.

langwatch logout --yes skips the confirmation, for scripts and machine teardown.
langwatch logout --keep-credentials removes the telemetry wiring but stays signed in.

Handing a laptop back or switching accounts is one command.

Real-time feedback

Each CLI’s trace lands in LangWatch live. You can:

Pin a filter “where langwatch.vk.tags contains cli” on the project dashboard.
Page on-call if any engineer crosses 80% of their monthly personal cap.
Run an offline eval comparing Claude Code vs Codex quality on tickets of a given type.

See the per-CLI pages for exact setup commands.

Verified smoke output

Lane A ran the gateway locally against pnpm dev on 2026-04-19 to confirm the response shape CLIs will see. Pinning the transcripts here so integrators can diff their actual output against known-good. Start the gateway pointed at a running LangWatch control plane on :5560:

cd services/gateway
GATEWAY_CONTROL_PLANE_URL=http://localhost:5560 \
  GATEWAY_ALLOW_INSECURE=1 \
  go run ./cmd/gateway

`/healthz`, always 200 once the process is up

HTTP/1.1 200 OK
Content-Type: application/json
X-Langwatch-Gateway-Version: dev
X-Langwatch-Request-Id: req_f5e4fd9f33a8861af4ab328aa00c45

{"status":"ok","version":"dev","uptime_s":0.988610959}

Kubernetes liveness probe target. X-Langwatch-Gateway-Version is set from the binary’s main.Version build-arg, production deploys carry the commit SHA so operators can answer “which pod served this” straight from the response.

`/v1/models` and `/v1/chat/completions` with no auth, 401 OpenAI-compat envelope

HTTP/1.1 401 Unauthorized
Content-Type: application/json
Traceparent: 00-e8aeda507fb93fafd3d2c20bbef283d6-53e32e8f6030ef05-01
X-Langwatch-Gateway-Version: dev
X-Langwatch-Request-Id: req_8f1250b94a959c1b2502792c757264
X-Langwatch-Span-Id: 53e32e8f6030ef05
X-Langwatch-Trace-Id: e8aeda507fb93fafd3d2c20bbef283d6

{"error":{"type":"invalid_api_key","code":"missing_api_key","message":"missing API key; supply Authorization: Bearer vk-lw-... or x-api-key"}}

The error message names both accepted auth headers, Authorization: Bearer vk-lw-... (OpenAI SDK, Claude Code, Cursor, Aider) AND x-api-key (Anthropic SDK). Either works against either endpoint. Traceparent + X-Langwatch-Span-Id + X-Langwatch-Trace-Id are present even on unauth 401s, observability of probe-abuse, misconfigured CLIs is available without inspecting access logs.

`/startupz` and `/readyz` behavior at cold boot

# /startupz, 503 only during the brief window before MarkStarted fires
# (initialisers + optional GATEWAY_STARTUP_NETCHECK_HOSTS probe), then 200
HTTP/1.1 503 Service Unavailable
{"status":"degraded","checks":{"control_plane_reachable":"ok"}}

/startupz and /readyz go to 200 as soon as the gateway has finished its startup initialisers and (if configured) the network-check probe has succeeded. They do NOT block on the auth cache observing a VK, a cold pod with no traffic and a fresh control-plane install with zero VKs will still go ready, and auth resolution happens on demand at request time.

Graceful drain on SIGTERM

INFO msg="gateway_draining" pre_drain_wait="5s"
INFO msg="gateway_shutting_down" timeout="15s"
INFO msg="gateway_stopped"

Matches the 4-phase drain documented in self-hosting/helm § Graceful drain. If your CLI sees one of these response shapes, the gateway is healthy. End-to-end completion through a real VK + provider requires the per-CLI config on each integration page.

Get Started

SDK Integration