Experiential: Open Source Model Gateway with Smart Routing
Experiential is a new open source model gateway from experientiallabs that provides a unified OpenAI-compatible API for hosted, BYOK (bring-your-own-key), and local models. The Rust-native gateway adds under 1ms for BYOK requests and under 2ms when Experiential supplies the provider key, and comes with every major inference provider and 1000+ models refreshed daily via a Codex agent that opens a PR.
Key Features
- One control plane for closed, open-source, local, and custom models — mix and match as needed.
- No markup: Unlike other routers that charge 10% token fees, Experiential takes zero markup on their hosted gateway, and the source is fully open for self-hosting.
- Opt-in traffic-based optimization: It uses standardized OpenTelemetry traces to build a router that picks the optimal model per request based on cost/quality.
- Text world models simulate rollouts for various models, an LLM judge scores them, and a nearest-neighbor classifier on prompt embeddings decides which model to call.
- Cache optimization and new model suggestions come out of the same simulation pipeline.
- Budget controls per user, agent, and use case.
Getting Started
Install and run the local gateway:
pip install experiential
exp
On first run, the setup wizard connects providers and assigns a one-time key. Then you can call any model through the gateway:
export EXP_GATEWAY_KEY=...
curl http://127.0.0.1:8000/v1/chat/completions \
-H "Authorization: Bearer $EXP_GATEWAY_KEY" \
-H 'Content-Type: application/json' \
-d '{"model":"opus-5","messages":[{"role":"user","content":"Help me"}]}'
Optimizing from Your Traffic
To build a custom router from your existing OTel traces:
exp build support-agent
This command walks you through providers, models, budget, and asks for your trace file. You can also fine-tune an open source model from collected traces using exp optimize model support-agent. If you just want to try it, fetch the public terminal-tasks dataset:
curl -L -o traces.otel.jsonl https://huggingface.co/datasets/experiential-labs/wmo-terminal-tasks-traces/resolve/540883e451dc13d34fb50fdd36b143cb0f1fb0db/traces.otel.jsonl
The project is under active development (450 commits, 655 stars) and includes an AGENTS.md for repo conventions. Telemetry is anonymous and can be disabled via exp config telemetry disable.
Experiential is positioned as a direct alternative to proprietary gateways like OpenRouter, with the promise that simple routing doesn’t justify a 10% token markup. It’s an option worth evaluating if you’re paying per-token fees on a commercial router or juggling multiple provider SDKs.
📖 Read the full source: HN AI Agents
👀 See Also

Strale.io offers free IBAN and email validation API for AI agents with no signup
Strale.io provides a free API with five capabilities including IBAN validation, email validation, DNS lookup, URL-to-markdown conversion, and JSON repair. No signup or API key is required, and it includes an MCP server for Claude or Cursor integration.

Krasis: Hybrid CPU/GPU Runtime for Large MoE Models Achieves 3,324 tok/s Prefill on RTX 5080
Krasis is a hybrid CPU/GPU runtime that runs large MoE models by handling prefill on GPU and decode on CPU, achieving 3,324 tokens/second prefill on an RTX 5080 with Qwen3-Coder-Next 80B Q4. It requires ~2.5x model size in system RAM but enables running models too large for VRAM.
Cue AI Uses Gemma 4 for Faster Voice Dictation: 44% Latency Drop, 30% More Usage
Cue AI replaced a cloud-based text polish step with Google DeepMind's Gemma 4 E4B running locally via Ollama, cutting median latency from 876ms to 488ms and increasing dictation usage by 30%.

Decision Passport: An Audit Layer for AI Agent Execution Governance
The Claude Code leak highlights a gap in AI agent governance. Decision Passport addresses this with append-only execution records, portable proof bundles, and offline verification for tamper-evident audit trails.