OpenSmith: Local LLM pipeline tracing without cloud

OpenSmith is an open-source, local-first tracing tool for LLM pipelines inspired by Ollama's relationship to OpenAI — it offers privacy and independence as its core value. Unlike LangSmith, which requires a cloud account, sends traces to a hosted service, and is most natural inside the LangChain ecosystem, OpenSmith installs with pip, works with any Python LLM code, and stores all trace data locally in SQLite (or optionally Postgres) so no data leaves the developer's machine. The tool provides three ways to instrument code: a @trace decorator that can be applied to synchronous and asynchronous functions with optional tags and token budgets; a context manager for manual tracing within specific code blocks; and autopatch(), which automatically patches supported backends without requiring code changes. Configuration is done via an optional opensmith.json file in the working directory, where developers can set the database path, enable console mode (printing traces to the terminal as they complete), and specify which backends to autopatch. A local web dashboard runs on localhost:7823 and automatically selects an alternative port if 7823 is in use. The command-line interface includes commands to initialize a config (opensmith init), start the dashboard (opensmith ui), list and filter traces by query, tags, and status (opensmith traces), view aggregate statistics (opensmith stats), export traces to JSON or CSV (opensmith export), and clear all stored traces (opensmith clear). Supported backends include OpenAI, Anthropic, LiteLLM, Qdrant, Chroma, and Pinecone. Traces are stored by default at ~/.opensmith/traces.db but can be redirected to Postgres by installing the opensmith[postgres] variant and setting the OPENSMITH_DB_URL environment variable. OpenTelemetry export is available as an optional feature via the opensmith[otel] install variant, allowing traces to be sent to an external OTLP HTTP endpoint at /v1/traces and metrics at /v1/metrics. The tool is licensed under MIT and is free and open-source.

OpenSmith positions itself as a local-first, privacy-preserving alternative to LangSmith in the LLM debugging and monitoring space. The core trade-off is explicit: LangSmith offers a hosted, cloud-based service tightly integrated with the LangChain ecosystem, while OpenSmith trades some convenience for full local control, zero vendor lock-in, and guaranteed data privacy. The tool is designed to lower the friction of getting started — no cloud account setup, no Docker required, and integration via simple Python decorators or context managers. The comparison table in the body makes this positioning clear: whereas LangSmith requires a cloud account and sends traces to hosted infrastructure, OpenSmith requires only pip and runs entirely on the developer's machine. For teams building LLM applications that are privacy-sensitive, cost-conscious, or working across multiple frameworks (not just LangChain), the local-first model removes significant friction. The supported backends (OpenAI, Anthropic, LiteLLM, vector databases like Qdrant and Chroma) cover the most common LLM stack components, suggesting the tool is aimed at a broad audience rather than a single framework.

No comments yet. Be the first to share your thoughts!