# LLM Workbench LLM Workbench is a model-agnostic control plane for LLM-powered products. Every run becomes a tamper-evident, human-gated, replayable bundle of trace events, artifacts, gates, and cost telemetry. The runtime is headless and provider-agnostic; the host application owns prompts, models, and tools, and calls explicit APIs to record what happened. Protocol v1.0.0. ## Key concepts - run bundle — signed JSON of run, trace, artifacts, ruleSets, optional engine snapshot. - trace event — typed, time-ordered fact (step_started, model_io, artifact_written, human_gate_resolved, span_started, etc.). - artifact — versioned, schema-validated structured output keyed by artifactKey + typeId. - gate — PAUSE_BEFORE, PAUSE_AFTER, CHECKPOINT human-review hold points on workflow steps. - telemetry — derived ledger summarizing token usage and cost per provider, model, step, user, tenant, and plan. ## Important links - robots.txt: https://www.llmworkbench.io/robots.txt - Sitemap: https://www.llmworkbench.io/sitemap.xml - Blog: https://www.llmworkbench.io/blog (RSS feed: https://www.llmworkbench.io/feed.xml) - Topic landings: https://www.llmworkbench.io/blog/tags/ (e.g. /blog/tags/run-bundles) - Protocol overview: https://www.llmworkbench.io/docs/protocol - OpenAPI spec: https://www.llmworkbench.io/api/openapi.json - MCP descriptor: https://www.llmworkbench.io/.well-known/mcp.json - Agent contract: https://www.llmworkbench.io/agents.md - Humans behind it: https://www.llmworkbench.io/humans.txt - Public demo run: https://www.llmworkbench.io/runs/demo - Long-form (full): https://www.llmworkbench.io/llms-full.txt - Source repository: https://github.com/llmworkbench/llm-workbench ## License Proprietary. Source-available, not OSI open source. Free to read, fork, modify, redistribute (noncommercial), and use for personal projects, research, education, and public-benefit work. Commercial use requires a separate paid license. ## Contact Open issues at https://github.com/llmworkbench/llm-workbench/issues. For commercial inquiries, see https://github.com/llmworkbench/llm-workbench/blob/main/COMMERCIAL.md.