AnonRouter

AnonRouter is a privacy-first LLM router whose core positioning is to connect to multiple model providers through an OpenAI-compatible endpoint while not storing prompts or responses.

Purpose and positioning: It is aimed at users, teams, and developers who want to use multiple AI models without sacrificing privacy, unifying multi-provider access, routing, billing, and monitoring behind a single gateway.

Main capabilities:

1. Unified OpenAI-compatible endpoint: Change the base URL to access the entire model catalog. The official OpenAI libraries work as-is in TypeScript, Python, and curl, with request structure and streaming format remaining unchanged.
2. Automatic routing: Send /auto to select the lowest-cost model with the appropriate capabilities within privacy boundaries.
3. Privacy tiers and filtering: Each route is labeled with the level of privacy guarantee it actually provides. The platform claims to include four clearly labeled tiers (for example, the provider can see prompts while the service is running but identity is hidden; prompts and responses are not retained after the request completes; prompts run inside a hardware-isolated enclave that the operator cannot read; the client encrypts prompts so that only a verified enclave can decrypt them). Users can filter the catalog by the required tier.
4. Confidential computing: Prompts are sealed on the browser side and only opened inside the enclave. Usage is metered based only on metadata. The platform claims that any keys it holds cannot read prompts, and that inputs are not stored or used for training.
5. No moderation by default: Moderation is treated as a model attribute rather than a gateway policy. Selecting an unmoderated model allows operation without filtering.
6. Workspace controls: Create scoped keys, review prefixes, and revoke access; set /auto policies, model access, provider ordering, and which providers to use; set privacy boundaries and fallbacks; confirm that prompts or responses are not retained. View request metadata, token counts, costs, and latency.
7. Anonymous accounts: No email, phone, or identity required; use a generated username and password, plus a recovery phrase for resets; a standard account can also be chosen for recovery. The platform claims that sessions remain opaque and that identity does not enter inference requests.
8. Web-based Chat and Studio: Use the same private routing as the API. Chat can converse with any text model or run multiple models side by side. Rooms and history are stored in the browser and do not land on the gateway. Studio can generate images and speech, and results are stored in a library local to the device.
9. Open source and verifiable: The confidential content plane is public, allowing inspection of enclacode running inside ve, and compare the attestation quote against the already-read build.
10. Agent-oriented: any client, agent, or IDE that supports the OpenAI API can point to this endpoint.

How to use:

• Create a private account (standard account or anonymous account), add credits, and connect through an OpenAI-compatible endpoint.
• Create scoped inference keys, point any OpenAI-compatible SDK to AnonRouter; switching only requires changing the base URL and key.
• Browse 200+ models, each labeled with privacy level, pricing, and capabilities before sending, and filter by the desired guarantees or use /auto.
• Command-line example: git clone anonrouter/confidential-content-plane; anonrouter attest --route tee (output includes measurement, signer intel-tdx, quote verified, matches published build).

Use cases: production routing and usage management for teams that need privacy protection, individual users who want to use models anonymously, developers connecting any agent or IDE to a unified private endpoint, and compliance-sensitive scenarios that need to choose models by privacy level.

Limitations and information not provided: the material does not explain which providers and model lists are specifically supported, pricing and billing standards, free credits, rate limits, available regions, performance metrics, or comparisons with other routers. The platform claims to provide 200+ models, but the specific list is not yet provided. Before use, please refer to the latest information on the official website.