Vociro — Social-Media Search Assistant

A command-line tool that launches autonomous social-media search agents powered by OpenAI (o3 or o4-mini). Each agent can gather information from:

• DuckDuckGo Web Search (HTML scrape)
• Reddit posts (plus top comments)
• Bluesky posts

The collected evidence is summarised and passed to a report compiler model that produces the final analysis. If the compiler feels the results are insufficient, it can request another search round via an internal redo_search tool.

Why?

Quickly answer exploratory questions that benefit from perspectives across traditional web pages, social-media discussion (Reddit) and emerging networks (Bluesky) without juggling multiple APIs or manual browsing.

Architecture

┌──────────────┐   1. strategy_model (o3/o4-mini)
│ generate     │      • Produces 3-8 search queries
│ search       │
│ objectives   │
└──────┬───────┘
       │queries[]
┌──────▼───────┐   2. N search agents (agent_model)
│ each agent   │      • Picks one query
│ uses tools   │      • Calls search_web / search_reddit / search_bsky
└──────┬───────┘
       │summaries[]
┌──────▼───────┐   3. report_model
│ compile      │      • Writes final report
│ final report │      • May call redo_search to loop back
└──────────────┘

Installation

# (optional) create and activate a virtual environment
python -m venv .venv && source .venv/bin/activate

# install from PyPI
pip install vociro

Environment variables

Set the following variables in your terminal session before running Vociro (no .env file is used):

Examples (Unix shells):

export OPENAI_API_KEY="sk-..."
export REDDIT_CLIENT_ID="abc" REDDIT_CLIENT_SECRET="xyz"

Windows (PowerShell):

setx OPENAI_API_KEY "sk-..."

Usage

vociro init  # start an interactive research session

1. Clarification phase — the assistant asks follow-up questions until it proposes a final objective:

   READY: <concise objective>
   

   You must then confirm with y (accept) or n (explain why, loop continues). Press z at any prompt to skip the phase entirely.

2. Source selection
   • Reddit and Bluesky are always enabled (sentiment sources).
   • DuckDuckGo Web search is optional (default n).

3. Model selection / number of agents — same as before.

During execution you will see, for each generated search query:

Search — <query>
Tool calls:
  1. search_reddit(query='…')
  2. search_bsky(query='…')
  …
Total cost so far: $0.0123

The agent is encouraged to perform deep dives (many tool calls) on Reddit and Bluesky to surface real user sentiment. The report compiler will call redo_search automatically if it feels more evidence is required.

Skip everything quickly

If you want a totally non-interactive run you can feed inputs through stdin, e.g.

echo -e "My question\nz\n\n\no4-mini\no3\n" | vociro init | cat

(The first z skips clarifications.)

Extending functionality

• Add new search back-ends by:
  1. Implementing a simple Python function that returns JSON-serialisable results.
  2. Registering a matching schema in build_tool_specs().
  3. Handling the tool call in execute_tool().
• All OpenAI calls are centralised, so adding caching or async batching is straightforward.

Caveats

• DuckDuckGo HTML scraping is brittle and for light personal use only.
• The project is not production-grade: no rate-limit back-off, retries or robust error handling.
• Token counts rely on the usage field from the OpenAI response and may vary slightly from billing.

License

MIT – do what you like, just don't blame me.