Python scripts become data apps by rerunning top to bottom.
streamlit.io ↗repo ↗docs ↗st.fragment (run_every) ↗2026 release notes ↗Build Streamlit apps with agent skills (Snowflake) ↗AppTest (headless testing) ↗
Streamlit owns browser dashboards authored as code; scene targets live agent-operated terminal UIs, and should learn from its agent-skills packaging.
Complement it: it does a neighbouring job; interoperate.
Streamlit made the dashboard a twenty-line Python script, and that is exactly the shape an LLM writes best. The rerun model means there is no callback graph to get wrong: the UI is a pure function of the script and session state, which is why generated apps usually work on the first try. It is the default target for 'make me a dashboard' across ChatGPT, Claude and Snowflake's own Cortex Code, and Streamlit now ships agent skills inside the package so coding agents learn current APIs. Fragments with run_every give live dashboards without websockets plumbing, AppTest gives headless verification, and Snowflake's backing means a two-week release train and enterprise hosting. Nearly 5M weekly downloads say the model won.
| UI as data | 3 scene 5 | Declarative Python; the wire form is a protobuf element tree, not an authored value. |
|---|---|---|
| Agent can drive it | 2 scene 5 | Official agent skills for code generation plus AppTest; no live-control protocol. |
| Agent can see it | 3 scene 5 | Structured via AppTest; screenshots for a live browser session. |
| Small, targeted edits | 2 scene 5 | Hot reload on script edit; coarse. |
| Time travel | 0 scene 0 | |
| Live data rate | 2 scene 5 | run_every polling reruns; seconds-scale. |
| Terminal native | 0 scene 5 | |
| Pixel graphics | 4 scene 4 | |
| Teaching | 2 scene 3 | |
| Maturity | 5 scene 1 | |
| Openness | 4 scene 4 | Apache-2.0 under Snowflake control. |
Streamlit scene now scene planned
Code generation is the main path: Streamlit bundles 'developing-with-streamlit' agent skills (17 sub-skills, app and theme templates) in the pip package; Snowflake Cortex Code generates Streamlit-in-Snowflake apps. Community MCP wrappers exist (e.g. dkedar7/streamlit-mcp, 2 stars) but nothing official. AppTest lets an agent run and inspect an app headlessly.
Low for generation (write a .py, `streamlit run`); higher for driving a live session (browser automation or AppTest).
AppTest exposes elements and values; live sessions need screenshots/DOM.
Edit the script and the app hot-reloads; no id-addressed patches to a running UI.
None for the UI; reruns are not recorded.
The most LLM-friendly code target for dashboards, with official agent skills, but agents author code rather than operate a live UI.
An app is a Python script; every interaction reruns it top to bottom, and the server streams a delta-encoded element tree (protobuf over WebSocket) to a React front end that reconciles it. State lives in st.session_state and caches (st.cache_data/resource). Fragments (@st.fragment) rerun only part of the page, on widget interaction or on a timer via run_every, which is the mechanism for live dashboards. Since 1.57 the server is Starlette/Uvicorn, exposed programmatically via st.App. Custom components extend the front end.
No published update-rate or throughput numbers. Live updates are polling reruns (run_every), suited to seconds-scale refresh, not streaming ticks.
The de facto Python data-app framework; homepage claims 'over 90% of Fortune 50' and ~19.4M monthly downloads. The usual target when an LLM is asked for a quick dashboard.
Used by: Snowflake (Streamlit in Snowflake), Cortex Code CLI users
Very active, corporate-funded, fast release train.
Releases: ~every 2 weeks: 1.61.1 2026-08-05, 1.62.0 2026-08-19, 1.63.0 2026-09-01, 1.64.0 2026-09-15, 1.65.0 2026-10-02. · Contributors: ~338 (GitHub, incl. anonymous) · Recent: 1,002 PRs merged since 2026-07-11 (GitHub search). · Bus factor: high: funded Snowflake team, many committers.
Agent-built dashboards bound to data, refreshed on a timer, with a pure 'state to view' model.
The 'dashboard from a few declarative lines' experience and its huge catalog of chart/table/metric widgets; Streamlit is what an agent reaches for by default and it already works.
A long-lived UI an agent patches by id at high rate, reads back as structured state and screen text, in the terminal; Streamlit regenerates code and reruns scripts in a browser.
| 1–7 | home · features · use cases · docs · changelog · under the hood · blog |
| t | light or dark |
| j k | scroll |
| g G | top · bottom |
| ? Esc | this · close |