AstrolabeSelf-hosted OpenAI-compatible routing gateway for OpenClaw agents with cost and safety policy

4.4 (5)

Reviewed by Daniel Nikulshyn·Updated June 2026

Routing Proxy OpenAI-Compatible Policy-Driven Cost Optimization Safety LLM Ops

Overview

Astrolabe is an open-source AI gateway designed to sit between OpenClaw agents and OpenRouter. It acts as a routing proxy that classifies each request, resolves an appropriate model lane from a static checked-in roster, executes the call against OpenRouter, and applies safety policy around tool use and untrusted inputs. The goal is to let self-hosted agents avoid hand-tuning providers and model IDs on a turn-by-turn basis. The project exposes a set of virtual models such as astrolabe/auto, astrolabe/coding, astrolabe/research, astrolabe/vision, astrolabe/strict-json, astrolabe/cheap, and astrolabe/safe. These map to concrete underlying models from providers like DeepSeek, OpenAI, Anthropic, MiniMax, Moonshot, xAI, Qwen, Google, and Mistral, which are maintained in static manifests rather than a hardcoded configuration object. Astrolabe centralizes four concerns for OpenClaw agents: routing flexibility, reliability and fallback behavior, cost control, and safety policy for tool use. It is intended to deliver these without adding a database, a hosted control plane, or any SaaS dependency. The OSS version is stateless and self-hosted; the operator supplies their own OpenRouter API key and an Astrolabe API key, then points OpenClaw at the Astrolabe instance. At runtime, OpenClaw sends a request to Astrolabe's POST /v1/responses endpoint (with POST /v1/chat/completions retained as a compatibility adapter). Astrolabe classifies category, complexity, and modifiers, resolves a lane and candidate model set, runs the request, verifies non-streaming responses, applies tool policy checks, and may escalate once to a stronger model. It returns the upstream response along with x-astrolabe-* headers and inline metadata. As of version 0.3.0 Beta, the project is early-stage and small. It is purpose-built for the OpenClaw ecosystem rather than as a general-purpose LLM gateway, so users outside that workflow may find more mature alternatives in tools like LiteLLM or OpenRouter's own routing. Its static, checked-in model roster gives reproducibility but requires manual updates as models change.

Key features

OpenAI-compatible /v1/responses and /v1/chat/completions endpoints
Static checked-in model manifests across multiple providers
Virtual model lanes (auto, coding, research, vision, cheap, safe, strict-json)
Request classification by category, complexity, and modifiers
Tool-use safety policy checks with single escalation
Response verification and x-astrolabe-* metadata headers

Pricing

Model: Free
Category: AI Model Serving Platforms
Rating: 4.4 / 5 (5)

Use cases

Cost-optimized LLM routing

Automatically route OpenAI-compatible requests to the lowest-cost model that satisfies policy, reducing inference spend without changing client code.

Safety-gated AI gateway

Apply policy-driven safety gates in front of model calls so prompts and responses are checked before reaching downstream applications.

Single-escalation fallback

When the cheapest model falls short, escalate once to a stronger model to balance reliability with cost control.

OpenClaw integration layer

Serve as the routing proxy for OpenClaw deployments, centralizing model selection and policy enforcement across services.

Pros & Cons

Pros

Self-hosted and stateless with no database or SaaS dependency required
Virtual model lanes abstract away provider and model ID selection
Built-in safety policy for tool use and untrusted inputs
Cost-aware routing with single-escalation fallback behavior
OpenAI-compatible endpoints ease integration

Cons

Early beta (0.3.0) with a very small user base
Tightly focused on the OpenClaw ecosystem rather than general use
Static model roster must be maintained manually as models change
Requires the user to supply and manage their own OpenRouter API key

Reviews

4.4

Average from 5 ratings.

Rina Desai

May 16, 2026

Years in this space

I've evaluated a lot of these over the years. What stands out here is the core workflow — handled better than most — and it is genuinely easy to set up. Worth the time if this is your use case.

Carlos Mendoza

Mar 25, 2026

Compared a few options

Evaluated this against two competitors. Where it wins: the core workflow and it is genuinely easy to set up. Where it lags: the docs could be deeper. On balance the feature set — especially the dashboard — justifies the 4 stars for our use case.

Ahmed Saleh

Jan 26, 2026

Does the job

Pretty happy overall. The onboarding just works and it saves real time. A few rough edges remain can be annoying, but no dealbreakers — I'd recommend it to a friend without hesitating.

Elena Rossi

Sep 6, 2025

Use it every day

Honestly didn't expect to like it this much. The automation is exactly what I needed, and it saves real time. I do wish a few rough edges remain, but I reach for it almost every day now and it just clicks.

Joanna Kowalski

Jun 18, 2025

Does the job

Pretty happy overall. The integrations just works and it saves real time. The docs could be deeper can be annoying, but no dealbreakers — I'd recommend it to a friend without hesitating.

Q&A

How does Astrolabe choose which model to use for a request?

Astrolabe uses policy-driven routing to automatically select the lowest-cost model that meets your requirements. It also includes safety gates and can escalate to a more capable model once if the initial response doesn't meet criteria.

Is Astrolabe compatible with existing OpenAI client libraries?

Yes. Astrolabe is an OpenAI-compatible proxy, so applications built against the OpenAI API format can route through it with minimal changes. It's designed specifically for use with OpenClaw.

What safety controls does Astrolabe provide?

Astrolabe adds safety gates into the routing pipeline, allowing policies to govern requests and responses. Combined with single-step escalation, this helps balance cost, quality, and safety on each call.