BoundFlow

BoundFlow is the operational layer for LLM agents and workflows you run unattended. It is not a prompt framework, an inference provider, or an agent builder — it is the control plane around the agents you already build, answering the questions that matter the moment agents run without supervision: what if it loops, what if it spends $50, what if it is about to do something irreversible, and which model should it be using? You write agents and multi-step, stateful workflows against a clean async Python SDK; the self-hostable Go backend (Apache-2.0, one binary running server, scheduler, and worker modes over Postgres) schedules, dispatches, governs, and audits them over gRPC. Guardrails are declared as policy rather than code, in three layers: runtime policies enforce hard per-run caps like maxcostusd, maxllmcalls, maxtokensper_call, and per-tool limits during execution; agent lifecycle policies react to trends, for example automatically downgrading to a cheaper model like Haiku when cost spikes; and workflow lifecycle policies self-heal by pausing, cooling down, or rolling the workflow back to a known-good version after repeated failures. Approval gates park a workflow for human sign-off before sensitive actions such as issuing a refund, and every approval and policy decision lands in a durable, queryable audit log. Inference is bring-your-own: your worker calls Claude (or any tool-calling LangChain model — OpenAI, Google, Bedrock) with your own key, which the backend never sees. Observability is OpenTelemetry-native, shipping run traces to Jaeger, Tempo, Langfuse, or Phoenix. Public preview, pre-1.0; a managed BoundFlow Cloud is in early access. No MCP integration is documented.