AI Governance

Take Control of AI Agent Uncertainty

LangWatch gives you a continuous understanding of how your agents behave, with the evidence to approve deployments confidently, and the controls to govern them as they scale.

The governance gaps teams bring us.

As AI tools spread across every team, the same control problems come up in almost every customer conversation. The gateway is how we close them.

• No visibility into AI usage: Nobody can see who is using which models, through which tools, or at what cost. Spend stays invisible until the bill arrives.
• Provider keys everywhere: Keys are copied into laptops, CI, and dotfiles with no revocation story. One leak means rotating keys across the whole org.
• Sensitive data leaks out: Data quietly leaves through AI tools and their connectors, with no way to see or stop where it goes.
• Policy lives in documents: Model, tool, and budget rules sit in policy docs and contracts you cannot enforce technically.
• No audit trail: When compliance asks who did what, there is no signed, exportable record to hand over.
• Cost runs unchecked: Without per-team budgets, spend spikes and off-hours usage go unnoticed until they get expensive.

Same data, two surfaces. No change to how people work.

LangWatch gives every developer a personal home for the AI tools they use, and gives admins one place to govern spend, policy, access and audit across all of it.

A governance home for the org

A bird’s-eye view of org-wide spend, top spenders, open anomalies, ingestion-source health and recent activity. Behind it: author routing policies, anomaly rules and the tool catalog, with a full audit log and CSV export.

Governance

• Spend, users and anomalies (Last 30 days)
  • Spend (30d): $2.05
  • Active users: 5
  • Open anomalies: 1 (0 critical · 1 warning)

Spending over time, by team (last 30 days)

• Customer Support: 68.3%
• Engineering: 24.6%
• Marketing: 4.7%
• Org-wide: 2.5%

A personal home for every tool

Sign in once with SSO and wrap any approved tool with one command. The portal shows the tools you are entitled to, issues per-provider virtual keys, and tracks your spend vs budget, model mix and daily cost. Every request lands in your own trace explorer.

My Usage

• Your AI usage this month
  • Spent this month: $8.47 of $1.00 budget
  • Requests: 274 this month
  • Most-used: sonnet-4-7 (45% of usage)

Spending over time

• By tool:
  • claude-opus-4-7: $6.20
  • claude-sonnet-4-7: $2.12
  • gemini-2-5-pro: $0.06
  • claude-haiku-4-5: $0.06
  • gpt-5-mini: $0.02

One control plane for every AI tool you run.

From the coding assistants on every laptop to the closed SaaS you can only ingest, it all routes through one gateway and lands in one place.

Coding assistants

Claude Code, Codex, Cursor, Gemini CLI and opencode. The wrapper signs the user in, mints a personal virtual key, and routes every request through the gateway.

Provider virtual keys

Per-user keys for Anthropic, OpenAI, Bedrock and Gemini. Drop into app config: same gateway, same budget, no secret in dotfiles.

External AI platforms

Workato, Microsoft Copilot Studio, OpenAI Enterprise, the Anthropic Compliance API and S3 audit drops all land in the same trace store as everything else.

Anomaly detection

Admin-defined rules for spend spikes, geo mismatch and off-hours activity, all folded from one event stream.

Compliance and SIEM

Every event is OCSF v1.1 mapped and replayable to Splunk, Datadog Security, Sentinel and Elastic.

In the request path

What the gateway enforces on every call.

The same drop-in endpoint carries attribution, budget enforcement, guardrails, caching and fallback on every request. It runs in front of every call, at about 11 microseconds of gateway-side overhead at 5k requests per second.

Virtual keys

lw_vk_live_keys scoped to org, team and project. Show-once, HMAC-hashed, rotatable, revokable within 60 seconds.

Hierarchical budgets

Caps at org, team, project, key or principal. Windowed, with soft-warn or hard-block semantics.

Inline guardrails

Run your evaluators on the request, response or each streaming chunk: allow, block or redact, with no app code change.

Tool, MCP and URL policy

Allow or block tool calls, MCP servers and outbound URLs by regex before the request ever leaves.

Caching passthrough

Anthropic cache_control is forwarded byte-for-byte, protecting your cache discount.

Automatic fallback

Per-key chains across providers on 5xx, timeout, 429 or open circuit. Client errors never trigger it.

Gateway vs SDK

Both ship in LangWatch. They solve different problems.

• Need: Trace my existing provider calls without changing credentials

  • Use: SDK
  • How: LangWatch SDK — instrument in place, keep provider keys.

• Need: Give the marketing team an LLM key they can’t leak, capped at $500/mo

  • Use: Gateway
  • How: Mint a virtual key with a budget. Share the VK.

• Need: Flip all internal apps from OpenAI to Anthropic for a week

  • Use: Gateway
  • How: Update model_aliases on the VKs. Zero code change.

• Need: Hand Claude Code or Cursor a personal, scoped key for every dev

  • Use: Gateway
  • How: langwatch login --device auto-issues a Personal VK.

• Need: OpenTelemetry traces from arbitrary frameworks already in production

  • Use: SDK
  • How: LangWatch SDK is OTLP-native, no proxy in the hot path.

Complete, not partial

Closed SaaS is ingested, not proxied. Still governed.

Tools that cannot route through a gateway, like Microsoft Copilot, OpenAI Enterprise and the Anthropic Compliance API, are pulled in via OTLP or audit-log drops into the same trace store. Proxied or ingested, everything is attributed, costed, retained on your policy, and exported to your SIEM the same way.

Control all agents in your organisation.

Virtual keys, budgets, fallback, caching and traces come with it. Cloud or self-hosted, your call.