Launched this week

OpenDecider

Launched this week

Open decision models that route AI agents without LLM calls

2 followers

Open decision models that route AI agents without LLM calls

2 followers

Visit website

Agent workflows often call an LLM just to decide where a request goes next. OpenDecider answers in one forward pass with a calibrated confidence, sending unsure cases to a fallback. Works with LangGraph, CrewAI, Agno, LlamaIndex, ADK and MCP. Apache-2.0.

Free

Launch tags:Productivity•Open Source•Artificial Intelligence

Launch Team / Built With

TinyFishSearch, fetch and browse for AI agents. Claim 30% bonus!

Promoted

Maker

📌

Hi! I'm Manjunath, and I built OpenDecider. Why I built it: in most agent workflows I looked at, a large share of the LLM calls weren't generating anything. They were deciding: is this billing or support? Should this go to the research agent or a person? Is the answer good enough to send? Each of those decisions costs a full LLM call, adds a second or more of latency, and gives you no reliable confidence score. What it does: OpenDecider is a family of small open models trained for exactly those decisions. You give it the input, the options and a question. One forward pass returns a probability for every option, so when it's unsure it sends the case to your fallback instead of guessing. What's included: 🔹 Integrations for LangGraph/LangChain, LlamaIndex, Agno, CrewAI, Microsoft Agent Framework, Google ADK, PydanticAI and Strands, plus an MCP server for Claude Code, Cursor and Mastra. 🔹 Models from nano (milliseconds, runs on a CPU) to larger GPU models, with GGUF and MLX builds for Ollama, LM Studio and Apple Silicon. 🔹 Production pieces: a batching server with auth and Prometheus metrics, Docker images, audit hooks, fallback on errors and OpenTelemetry spans. 🔹 A runnable example for every framework, with no API key needed. Everything is Apache-2.0 with open weights, and it all runs on your own hardware, so your data never leaves it. Coming next: guardrails against jailbreaks and prompt injection in each framework's own hook, a TypeScript client, and nano running in the browser. I'd love to hear where you'd use it, and which decisions in your agents you'd hand over first. Honest feedback on what's missing is the most useful thing you can give me.

Report

1d ago

Maker

Update: two of the three "coming next" items have shipped.

✅ Guardrails (0.5.0). A prompt guard against jailbreaks and prompt injection, plugged into each framework's own hook. The default guard model, opendecider-small-td, catches as many attacks as Laya's guard with about half the false alarms.

✅ TypeScript client (0.6.0). npm i @opendecider/client. Typed decisions, the router and the guard, with tools for the Vercel AI SDK and Mastra. It has no runtime dependencies and publishes with npm provenance, and each version goes live only after a maintainer approves it with 2FA.

🔜 nano in the browser. We're building this now: opendecider-nano as ONNX, with WebGPU and WebAssembly backends. It already runs in Chrome; we're checking every quantised build against the full benchmarks before it ships.

If you're putting guardrails or routing into a TypeScript or browser app, what would you want from it? Feedback is very welcome.

Report

5h ago