How to get Pragmatic, what ships in the repository, and how to use what exists today. If this site and the repository ever disagree, believe the repository.
Pragmatic is in public beta and the source is public. The core example runs without a model or an API key; the tool-loop example records a real Claude agent end to end, then resumes and replays it.
your terminal
# the source, on GitHub git clone https://github.com/pragmaticco/pragmaticMain # the runtime, no model and no key required cargo run -p pragmatic --example research_agent # a real Claude tool-use agent, durable end to end ANTHROPIC_API_KEY=... cargo run -p pragmatic-anthropic --example tool_loop # the same runtime from Python pip install pragmatic-runtime
The repository carries the full usage guide and every example: github.com/pragmaticco/pragmaticMain.
The pragmatic crate, sync and async, with zero dependencies. It links into the agent you already run and journals every model call and tool effect to an append-only, hash-chained log; there is nothing to provision.
Claude as a journaled Oracle, from plain completions to full Messages API tool-use loops. The OpenAI adapter speaks to any OpenAI-compatible server: OpenAI, Ollama, vLLM. Or implement the Oracle trait for any model client of your own.
Five commands over your journal directory: runs, show, verify, export, serve. Serve a live console over your journals, or export one run as a self-contained HTML file. Loopback-only by default, since journals contain prompts and outputs.
The same runtime from Python via pip install pragmatic-runtime, with journals byte-compatible with Rust and the CLI. C++, Java, Go, and Node bindings sit behind one C ABI.
Pragmatic is licensed under the Business Source License 1.1: read it, modify it, self-host it. Each released version converts to Apache 2.0 four years after release, so nothing stays closed forever.
Free for non-production use, and for production use by organizations under one million dollars in annual revenue. Production beyond that takes a commercial license: email us.
Journals are single-process on local disk today, and the managed cloud backend is still in development. The replay guarantee is exercised by the runtime's determinism and recovery suites, including a full tool-use loop crashed and resumed over live sockets.