ashlar

Ashlar is an Apache-2.0, local-first .NET runtime for building and embedding auditable AI workflows. According to the README, it can run as a CLI, an HTTP API, containers, or inside a host application through AddAshlar(). A submitted task is stored under an id and the trust log records an entry whose sourceId is that id, so both the output and the record of what ran are available. Proposed code from Ashlar or a model becomes trusted only after the certification gate: analyzer fence, witness, mutation testing, and determinism. The gate is the required CI check cert-gate on master, and admitted or rejected outcomes are tracked in the certification evidence ledger. The README is careful about status: the in-process autonomy loop is experimental and ships in hold mode by default, admitting nothing until enabled. The supported operator-governed self-extend path uses policy controls, a build course, and a post-apply canary with auto-rollback; it ships sealed. Local model routing via Ollama is the default, with mock or offline available behind an explicit ASHLAR_ALLOW_MOCK flag, while OpenAI or Azure are opt-in targets. The API refuses to start on a network exposure profile without auth, and defaults are HTTP-only with no auth for local dev. It also documents MCP and A2A protocol support, gRPC transport, AWS SNS and DynamoDB ingress, peer federation and mesh sharing, and deployment through GHCR images and compose files. Ashlar is not a hosted SaaS or chatbot, and it is not cloud-dependent; air-gapped and local-only deployments are first-class.