Built (with love) by Devoteam Snowflake Partner, brings Snowflake CoCo into Streamlit — streaming agent UI, tool cards you can actually read, and approval gates that fit governed data apps.
You own the page. CoCo owns the session. panel() streams the transcript; copilot_rail() wraps that panel as a right-rail Copilot for multipage apps. Your app keeps st.chat_input, metrics, and forms. Approvals pause Write / Edit / Bash / SQL until someone clicks Approve once, Always allow, or Deny.
Alpha
0.1.7— API may still move. Star / watch the repo if you plan to ship on it.
Repo: github.com/lletourmy/streamlit-coco (temporary PyPI source) · Dev: streamlit-coco-dev
SDK docs: Cortex Code Agent SDK
| Without streamlit-coco | With streamlit-coco |
|---|---|
| Wire CoCo yourself across Streamlit reruns | Session + fragment polling that keeps streaming |
| Raw JSON tool dumps | Meaningful cards (Glob, Grep, Read, Write, SQL, AskUser…) |
| Hope the agent behaves | require_approval_for + HITL UI |
| Chat-only demos | Structured callbacks into your own widgets |
Also: headless query() for scripts and CI, plus a legacy all-in-one chat() if you want built-in input.
- You want CoCo Desktop, the CLI, or an IDE extension. This is an embed in your Streamlit app — no file tree, multi-tab workspace, or full IDE.
- The Streamlit host cannot run CoCo. The agent is server-side (CLI on the host today; remote API is still on the roadmap). Typical Streamlit Community Cloud without that setup will not work.
- You are not building in Python / Streamlit. There is no TypeScript package; use the Cortex Code Agent SDK directly.
- You need Slack, a hosted CoCo SaaS, or a product MCP server. Out of scope. MCP passthrough via
mcp_serversalready works. - You would not type the SQL yourself on this role. The agent uses the Snowflake role in the connection — do not wire
ACCOUNTADMINinto a web UI.
Alpha 0.1.7 — APIs may still move. Prefer panel() + your own input; chat() is the legacy all-in-one.
uv add "streamlit-coco[sdk]"
# or: pip install "streamlit-coco[sdk]"From a clone (editable + tests):
make install- Python 3.10+ · Streamlit ≥ 1.53
- CoCo CLI on
PATH(cortex --version) - Authenticated Snowflake connection (
~/.snowflake/connections.tomlor equivalent) cortex-code-agent-sdk(pulled in by thesdk/devextras)
Full local setup: doc/deployment/local.md (CLI install, Snowflake connections.toml, running examples, troubleshooting).
API: doc/api.md.
Preferred pattern — you own the input; CoCo owns the session and output:
import streamlit as st
import streamlit_coco as st_coco
opts = st_coco.CocoOptions(
connection="analytics",
cwd=".",
allowed_tools=["Read", "Glob", "Grep"],
require_approval_for=["Edit", "Write", "Bash"],
)
env = st_coco.check_environment(connection=opts.connection)
if not st_coco.render_start_gate(opts, session_key="copilot", env=env):
st.stop()
session = st_coco.get_or_create_session(opts, key="copilot")
st_coco.panel(session=session, warm_up=True, show_status=True, run_every=0.25)
st_coco.chat_input_bar(session, placeholder="Ask CoCo…")allowed_tools— may auto-run (enforced in Python when approvals are configured)require_approval_for— pause for Approve once · Always allow · Deny
Legacy all-in-one component (built-in input):
st_coco.chat(session=session, key="coco_chat", height=560)make chat # panel + chat input + tool cards + approvals
make cwd-upload # upload files into agent cwd + chat
make approval # legacy CCv2 chat
make structured # custom structured-output panel
make headless # asyncio query() pipeline
make backlog # Product Backlog Desk (multipage business demo)
make bi-semantic # Tableau / Power BI → semantic view + RAP (screens 1–6)
# make tableau-semantic is an alias for bi-semanticExploratory prompts: examples/testdata/prompts.json.
Backlog desk: examples/backlog_desk/README.md.
BI → Semantic: examples/bi_to_semantic/README.md.
File upload: doc/features/file-upload/file-upload.md.
output = st.container()
def render(data: dict, result: st_coco.CocoChatResult) -> None:
with output:
st.dataframe(data.get("selected_features", []))
st_coco.panel(session=session, on_structured_output=render)import asyncio
import streamlit_coco as coco
async def run():
async for event in coco.query("Profile ANALYTICS.CUSTOMERS"):
if event.type == "result":
print(event.structured_output)
asyncio.run(run())The agent is server-side. The browser only sees Streamlit widgets. panel() (or copilot_rail() around it) polls a CocoSession worker via @st.fragment; the session talks to the Cortex Code Agent SDK, which runs the cortex CLI against your Snowflake account. Destructive tools pause in Python (can_use_tool) until someone clicks Approve / Deny. Headless query() skips the UI and uses the same session/SDK path.
Browser ──► panel() / copilot_rail() / chat_input_bar()
│ @st.fragment poll (app page does not rerun)
▼
CocoSession (thread + asyncio, transcript + pending approval)
│ can_use_tool → render_approvals()
▼
cortex-code-agent-sdk ──► cortex CLI ──► Snowflake CoCo + RBAC
Legacy chat() is the same session, with a CCv2 frontend instead of native widgets.
| Capability | Entry points |
|---|---|
| Native panel + approvals | panel(), chat_input_bar(), render_approvals() |
| Copilot rail (right-column Copilot) | copilot_rail(), transcript_view_pills() — doc/features/copilot-rail/ |
| App viewer (child Streamlit iframe) | app_viewer(), default_fix_prompt() — doc/features/app-viewer/ |
| Tool cards & AskUser / plan UI | see doc/features/tools-display/ |
| Session & options | CocoSession, CocoOptions, get_or_create_session |
| Headless events | query() |
| Legacy CCv2 | chat() |
streamlit_coco/
├── ui.py # panel(), send_prompt(), render_approvals()
├── rail.py # copilot_rail(), transcript_view_pills()
├── viewer.py # app_viewer()
├── app_preview.py # child Streamlit process helpers
├── session.py # CocoSession worker + transcript
├── permissions.py # HITL can_use_tool gates
├── query.py # headless query()
├── component.py # legacy chat() CCv2 mount
└── frontend/ # static CCv2 assets
examples/ # chat, backlog desk, BI → Semantic, …
doc/ # PRD, roadmap, feature specs
Full diagram and runtime notes: doc/prd.md §5 · threat-model topology: doc/security/threat-model.md.
API: doc/api.md. Feature checklists: doc/features/README.md.
make install # uv sync --extra dev
make check # ruff + unit/smoke (ignores tests/e2e)
make e2e-install # Playwright + Chromium (once)
make e2e # UX e2e vs examples/e2e_ux_harness.py
make test-all # check + e2e + audit
make audit # pip-audit
make format # ruff format + fix
make build # sdist + wheel
make sync-release # copy tree → public clones (see doc/deployment/publish.md)
make help # all targetsCI runs lint, tests, and pip-audit on every PR to main. Full local gate: make test-all (doc/testing.md).
Releases: develop here (streamlit-coco-dev), sync + tag on lletourmy/streamlit-coco → PyPI (guide).
Docs: PRD · API · Roadmap · Training · Deployment · Changelog · AGENTS.md
| Role | Name | Contact |
|---|---|---|
| Asset Owner | Laurent Letourmy | laurent.letourmy@devoteam.com |
| Contributors | DevoteamSP / streamlit-coco contributors | streamlit-coco-dev |
Snow Builders level: N0 (alpha), targeting N1. Identity sheet: ID.md.
Apache-2.0