# manyside: agent quick start

This is documentation, not a hosted MCP endpoint. Requires a project checkout,
Python 3.10+, and Codex or Claude Code CLI. Request project access through
manyside.contact@gmail.com.

From the checkout root run ONE:
python3 scripts/setup_chat.py --client codex
python3 scripts/setup_chat.py --client claude

Setup changes the selected client's local configuration, installs the MCP runtime
and companion skill, and makes no model calls. --dry-run previews the actions.
Keep the checkout in place. Start a new chat after setup.

The local compatibility registration and skill ID is `user-review`.
Find the installed tools; your host may prefix their names.
1. review_options: discover supported counties, questions, fields and execution availability.
2. start_review: preserve the user's exact content; default to 20 reviewers and seed 42.
3. get_review: follow the returned run_id until completed, partial or failed.
4. list_reviews: retrieve saved history. Check history after a lost response before retrying.

Offline example arguments for start_review:
{"content":"A grocery app with shared household lists.","kind":"idea","county_ids":["06001"],"sample_size":20,"seed":42,"engine":"mock"}

Mock results are arbitrary offline workflow demonstrations, not useful audience feedback.
Never derive audience conclusions or recommendations from mock scores.
For live configuration consult docs/chat-setup.md in the checkout and discover
the installed tool schema. Require explicit paid-call/data-transfer authorization.
There is no hard dollar-budget cap. Keep credentials server-side, never in chat.
Never silently substitute mock answers for failed live calls.

Resolve counties with review_options(topic="counties", search=...). The local
integration supports all 3,144 counties and county equivalents in the 50 states and DC.
Do not silently substitute a county for an unsupported or national audience.
Do not infer preferences or abilities solely from demographics.

For evidence use get_review(section="responses", question_id=...) and,
when needed, section="profiles", profile_id=.... Include run_id in every lookup.
Cite run_id, question_id and profile_id with claims. Responses are structured
records, not interview transcripts. Do not invent quotes or motives.
Score indices are not percentages. Report missing or excluded responses.

Compare revisions using the same audience, sample size, seed, questions,
context and engine; retain both run IDs. Live synthetic responses are exploratory,
unvalidated estimates, not real interviews or representative polls.
This integration runs locally; website accounts and runs are separate.