Library / SDK
Guardrails & StructureInference & Serving
Microsoft AICI
1Wasm 2controllers 3steer 4LLM 5token 6generation
the six Ws · specification
W1
Who
Microsoft Research built AICI as a research prototype.
W2
What
A framework where WebAssembly-compiled Controller programs constrain and direct LLM output token by token in real time.
W3
Where
Runs alongside an LLM inference server, with controllers written in Rust, Python, or JavaScript compiled to Wasm.
W4
When
Open-sourced in 2024 as a research prototype.
W5
Why
Enables fine-grained, programmable control over decoding, prompt editing, and multi-agent generation strategies beyond simple grammars.
W6
With
Integrates with local inference runtimes and ships example jsctrl and pyctrl controllers.
W7
Watch
Open-source (MIT license) on GitHub from Microsoft Research, self-hosted, experimental prototype status with no guaranteed maintenance cadence.