ThinkWatch

For a whole organization, or for one developer's machine.

Website · Docs · ThinkWatch Enterprise · ThinkWatch Lite · ThinkWatch Core

Claude Code, Codex, Cursor, and your own agents talk to model providers directly. API keys end up scattered, nobody sees what was sent, and the bill is a surprise. ThinkWatch puts a gateway in that path, so every model request and tool call is routed, checked, and accounted for.

It comes as two products, built for different people.

Underneath both sits ThinkWatch Core, a set of MIT-licensed Rust crates for routing, failover, cost accounting, and redaction.

Just as an SSH bastion host is the single gateway through which all server access flows, ThinkWatch Enterprise is the single gateway through which all AI access flows.

Every model request. Every tool call. Every token. Authenticated, authorized, rate-limited, logged, and accounted for.

                    ┌──────────────────────────────────────┐
 Claude Code ──────>│                                      │──> OpenAI
 Cursor ───────────>│    Gateway  :3000                    │──> Anthropic
 Custom Agent ─────>│    AI API + MCP Unified Proxy        │──> Google Gemini
 CI/CD Pipeline ───>│                                      │──> Azure OpenAI / AWS Bedrock
                    └──────────────────────────────────────┘
                    ┌──────────────────────────────────────┐
 Admin Browser ────>│    Console  :3001                    │
                    │    Management UI + Admin API         │
                    └──────────────────────────────────────┘

• AI API gateway. OpenAI, Anthropic, Google Gemini, Azure OpenAI, and AWS Bedrock behind one endpoint, with scoped virtual keys per team, project, or developer.
• MCP gateway. A central tool proxy with namespace isolation, tool-level RBAC, and an audit trail for every invocation.
• Security and compliance. SSO through any OIDC provider, role-based access control, PII redaction, and encryption at rest.
• Rate limits and budgets. Request and token limits per key or per user, with spend budgets and alerts.
• Observability. Searchable audit logs, cost analytics, and Prometheus metrics.

Start with the Quick Start, or read the docs.

• What it cost, and how far to trust that number. Measured, estimated, and unpriced are reported separately and never added together; usage served by a subscription is counted apart from billed usage.
• Where each request went, and why. The rule it matched, the group it went through, and every upstream attempt with its status and duration. A dry run answers the same question before any traffic.
• What went out with it. Secrets replaced on their way to an untrusted upstream and restored in the response, dangerous tool calls cut off mid-stream, and client configuration files scanned for hidden characters and dangerous commands.
• Cost at a glance. Today's cost and output rate, or the quota left on a subscription account, right in the menu bar.

Released for macOS 12 or later on Apple Silicon. The gateway ships inside the app, and the app updates itself:

brew install --cask thinkwatchproject/tap/thinkwatch-lite

Or download the disk image from the releases page. The interface is currently in Simplified Chinese; an English interface is in development.

Rule-based routing, mid-flight failover, cost accounting against a price snapshot, outbound secret redaction, and inspection of tool calls coming back from upstream. twcore is a complete, self-contained gateway binary built from these crates:

cargo run -p twcore -- init     # write a commented config.yaml
cargo run -p twcore -- serve    # start the gateway and control plane

ThinkWatch Lite and ThinkWatch Core are released under the MIT License.

ThinkWatch Enterprise is source-available under the Business Source License 1.1. Non-production use is free. Production use is free up to 10,000,000 Billable Tokens and 10,000 MCP Tool Calls per UTC calendar month. See LICENSING.md for the full terms, the tiering model, and the changeover to GPL-2.0-or-later.

Made with ❤️ for AI-native engineers and teams