Runsight — YAML-first workflow engine for AI agents

Runsight is an open-source YAML-first workflow engine for AI agents that lets developers design, commit, and track agent pipelines from a local terminal or web canvas.

What is Runsight?

Runsight is a developer tool that turns AI agent workflows into version-controlled YAML files. It takes a YAML workflow definition (with block types like linear, gate, or code) and executes agent steps on your machine using your own API keys. The tool outputs per-block cost and latency data, plus evaluation assertions. It is built and maintained by cubic.dev.

Key Features

• Dual canvas/YAML editor — Edit YAML in the Monaco editor and see the canvas update; move nodes on the canvas and the YAML stays clean. Both views stay in sync.
• Per-run cost tracking — Every block records cost to the cent. Set hard budget caps (max_cost, max_tokens) that kill execution before overspending.
• Built-in eval framework — Define assertions on block outputs and use transform hooks for structured extraction. Supports regression testing across runs.
• Pause and kill mid-run — Stop a running agent before it wastes budget. Pause execution, inspect state, then resume or kill.
• Git-diffable workflows — Every workflow is a plain YAML file on disk. Branch, review, and merge agent changes like any other code.
• Self-hosted with your keys — Runs on your machine. Your API keys, your models. Nothing leaves your network unless you choose.
• Loops and sub-workflows — Iterative refinement with break conditions. Nest workflows as child steps for hierarchical orchestration.

Who is it for?

• AI engineers who need to design and iterate complex agent pipelines without scattering logic across Python scripts.
• DevOps and MLOps teams that want version-controlled, reviewable workflow definitions in their Git repository.
• Cost-conscious developers who need per-run billing visibility and hard budget caps to prevent surprise API charges.

What can you do with Runsight?

• Design agent pipelines in YAML — Write a workflow with sequential research, gate review, and publish blocks. The YAML file is your source of truth; run it from the terminal with runsight run research-pipeline.
• Commit workflows to Git — Every edit is a file change. Run git diff to see what changed, then commit and deploy like any code.
• Evaluate and iterate — Attach assertions to block outputs. If a review block fails, the workflow loops back to research. Regression test across runs to catch regressions.

How does Runsight work?

Run $ uvx runsight in your terminal — it scaffolds a project if none exists and starts a local web UI at localhost:8000. No signup or cloud account required. Design your workflow by writing YAML in the editor or dragging blocks on the canvas. Click Run to execute; each block shows real-time cost and latency.

FAQ

Is Runsight free?

Yes. The tool is open source and self-hosted. You run it on your own machine with your own API keys. There is no signup, no cloud account, and no vendor lock-in.

Do I need a cloud account to use it?

No. Runsight runs entirely on your local machine. Your API keys and data never leave your network unless you explicitly connect an external service.

What happens if Runsight stops being maintained?

Your workflows are plain YAML files on disk — standard, human-readable, and committed to your Git repository. You keep full control of your agent definitions and can migrate to another tool or run them manually at any time.

Can I use my own AI models and API keys?

Yes. You supply your own API keys during onboarding. Runsight uses your keys to call the models you specify in each block's soul_ref. No external service proxies your requests.