Inputs
Rendered from the workflow itself — exactly the fields it declares.
Runs the real Spaces in order; a sleeping brick can take a minute on its first call. Repairs appear inline.
Live log
Artifacts
Everything the run produced, served straight from this Space.
Step results
| step | time | brick | repairs / error | |
|---|---|---|---|---|
| — | ||||
app.py — real daggr code
Deterministic codegen: no model wrote this. The file that ran here is the file
you download. pip install daggr && daggr app.py runs it locally.
Build a workflow to see its code.
Publish this workflow
Commits your workflow to a public Hugging Face Dataset — auditable, forkable, no server involved. Metadata is shared under CC-BY-4.0; the assets your workflow produces keep the licence you chose.
Load a workflow
Pick one from the community leaderboard, or paste a spec.
Community leaderboard
| workflow | author | industry | bricks | steps | votes | licence | published |
|---|---|---|---|---|---|---|---|
| — | |||||||
The brick registry
Every brick was introspected against its live API; proven means we made a real call and it produced output. Dead spaces stay listed so the Medic can fail over instead of retrying forever.
| id | source | modality | outputs | licence | commercial | status | probe | industries | notes |
|---|---|---|---|---|---|---|---|---|---|
| — | |||||||||
What this is
Daggr Studio turns a described goal into a running daggr workflow assembled from bricks that were verified — introspected against their live API, and for cheap modalities actually executed — before being offered to you.
- Plan — a small model (default
deepseek-ai/DeepSeek-V4.1-Flash) chooses bricks; it can only pick ids from the catalogue, and endpoints/params are filled in from the registry, never invented. - Validate — every endpoint and parameter is checked against the live Space. Findings carry stable codes.
- Heal — deterministic repairs first (renamed parameters, dead Spaces, non-commercial swaps), then a model that may only emit ops from a closed vocabulary; rejected ops are reported with reasons.
- Run — executed in-process with daggr's own executor, node by node, with repairs at run time (sleeping → retry, quota → your token, changed API → remap, deleted → fail over).
- Share — published as a commit to a public dataset; leaderboard filtered by modality, industry, licence and compute.
No LLM writes the generated code: the spec is model-authored, app.py
is deterministic codegen, so what you run here is what you download.