Kestrel

Kestrel is a sovereign AI agent framework for building agents with cryptographic identity, persistent memory, and constitutional governance enforced above the language model. Its description emphasizes agents that cannot be taken away, with ownership of identity and memory kept with the user rather than a platform.

The framework gives each agent a portable DID identity. According to its architecture, every agent is born with a did:pkh decentralized identifier backed by a cryptographic keypair using the same secp256k1 curve as Ethereum. That identity is generated locally, owned by the user, and portable to any environment. Kestrel also stores conversations, documents, and relationships in a local SQLite knowledge graph with full-text search and RAG. The memory is described as encrypted at rest, portable, and deletable on demand. A third pillar is governance: each agent runs under an audited constitution enforced above the LLM, with a genesis audit on creation, cryptographic signatures required for amendments, and an audit trail.

Kestrel includes a web UI and a console for working with agents on a machine the user controls. The page says agents can be created and started with one command each, and that the setup provides a DID identity, memory store, and constitutional protections automatically. It is shown as self-hostable on a laptop, server, or cloud environment, and the page says it runs on Mac, Linux, and Windows through a Python stack using FastAPI and uv. The framework also notes multi-LLM support for Anthropic, OpenAI, Gemini, and Ollama, and says privacy can be set per agent, including an EPHEMERAL mode that routes to a local model only and a NORMAL mode that can use the cloud.

Kestrel is open source under the Apache 2.0 license. The page also states that the sovereign core is fully open, with no platform lock-in, and that identity and memory are not hosted by the project. It is presented as production-ready infrastructure, with launch timing listed as May 7, 2026, and an early-access option before public release.