Open source Open infrastructure for AI you can own. Adaptive agents, real-time voice, and deterministic data workflows on code that can be inspected, self-hosted, extended, and governed.
self-hosted human-governed no lock-in open code
open adaptive stack inspect · own · extend
01 substrate Areev governed adaptation → 02 voice runtime Flowcat real-time voice → 03 deterministic tools Agent Tools reproducible work →
01 / 03 Areev · governed adaptation Context, execution history, and evidence-backed improvement under human authority.
open source The stack Three foundations. One operating philosophy. Own the runtime, keep authority visible, and move probabilistic intelligence behind deterministic boundaries wherever correctness matters.
01 · Areev 02 · Flowcat 03 · Agent Tools
01 / Adaptive substrate Rust · MIT · Apache-2.0
Areev Self-improving agents, governed.
An embedded substrate for adaptive agents: knowledge and execution history in one content-addressed file, with every improvement kept under explicit human authority.
View on GitHub Propose → review → apply → verify—evidence-cited, undoable, and re-measured Budget-shaped context via CAL · in-process recall · no server in the recall path Local-first, no telemetry · provenance, encryption, and subject erasure built in ~30 µs in-process recall
13 deterministic analyzers
1 file knowledge + run history
02 / Real-time voice Rust · Apache-2.0
Flowcat A self-hosted, native-Rust runtime for real-time voice agents.
One self-contained binary runs VAD → STT → LLM → TTS—or speech-to-speech—with in-process SIP/RTP, provider choice, and no control plane phoning home.
View on GitHub ContextRelay preserves long-call context by reseeding sessions as compact text Sub-1ms p99 latency at 2,000 concurrent calls per process Self-host in your VPC · broad STT, TTS, and LLM provider support <1 ms p99 runtime latency
2,000 calls per process
1 binary voice runtime
03 / Deterministic work TypeScript · MCP · A2A · REST
Agent Tools Deterministic tools for work agents cannot afford to guess.
Large files, binary formats, and strict transformations stay outside the LLM, handled by reproducible tools that can be tested and verified.
View on GitHub 18 categories across JSON, CSV, PDF, XML, SQL, crypto, and more MCP, A2A, and REST from one server Self-host with Docker, Kubernetes, and Terraform 136+ deterministic tools
18 tool categories
3 agent protocols
Open by design Control is a feature, not a deployment detail. Open code makes the operating boundary visible—from context and voice to every deterministic tool call.
01 Inspect every layer Runtime behavior, context assembly, providers, and policies remain visible.
02 Deploy where needed Cloud, private VPC, on-premise, or edge—the boundary stays a choice.
03 Bring provider choice Models, speech providers, data systems, and channels stay interchangeable.
04 Extend without permission Fork, integrate, replace, and evolve the stack around real operational needs.
Run it. Inspect it. Extend it. Start with the open source, or bring a real-world problem that needs an accountable production system.