A powerful, autonomous research agent built with LangGraph and TinyFish AI. This agent takes a research query, plans a series of targeted web searches, executes them using TinyFish's headless browser automation, and synthesizes the findings into a comprehensive research document.

Available as a Claude Code Plugin or a standalone CLI.

• Autonomous Planning: Decomposes complex research queries into actionable tasks with specific URLs and goals.
• Headless Browser Execution: Uses TinyFish AI to interact with websites and extract content.
• State-Managed Polling: Implements a robust polling mechanism with custom backoff intervals ([2, 2, 4, 4, 4, 6, 6, 7, 8, 8]) to efficiently wait for browser tasks to complete.
• Multi-LLM Support: Supports OpenAI, Anthropic, and Google Gemini via LangChain.
• Structured Synthesis: Aggregates results from multiple sources into a unified Markdown report.
• Intermediate Results: Saves raw per-source results as JSON for debugging and transparency.

The agent is built using a directed acyclic graph (DAG) via LangGraph:

1. Planner: Generates research tasks from the query.
2. Submitter: Submits tasks to TinyFish as asynchronous "fire-and-forget" runs.
3. Poller: Periodic checks for completed runs.
4. Wait Node: Implements the polling delay with backoff logic.
5. Synthesizer: Finalizes the research document once all data is collected.

From within Claude Code:

/plugin marketplace add likhitjuttada/tinyfish_web_research_agent
/plugin install tinyfish-research@likhitjuttada-claude-marketplace

From the command line:

claude plugin marketplace add likhitjuttada/tinyfish_web_research_agent
claude plugin install tinyfish-research@likhitjuttada-claude-marketplace

Manual installation (for development):

git clone https://github.com/likhitjuttada/tinyfish_web_research_agent.git
cd tinyfish_web_research_agent
pip install -r requirements.txt
# Configure .env (see below), then in Claude Code:
claude --plugin-dir /path/to/tinyfish_web_research_agent

/tinyfish:research How does RLHF work?

The plugin will:

1. Validate prerequisites and create a research workspace
2. Submit browser tasks to TinyFish AI
3. Poll for results with increasing backoff intervals
4. Synthesize findings and present them directly in Claude Code

Cancel an active session:

/tinyfish:cancel-research

Monitor progress:

tail -f research/<research_id>/progress.log

User → /tinyfish:research → setup-research.sh (validates + creates state)
                                    ↓
                          Claude finishes response
                                    ↓
                          Stop Hook fires automatically
                                    ↓
                          run-research.sh → python main.py
                                    ↓
                          Results saved → Hook returns synthesis prompt
                                    ↓
                          Claude reads results and presents summary

• Python 3.9+
• A TinyFish API Key
• An LLM API Key (OpenAI, Anthropic, or Google)

1. Clone the repository:

   git clone https://github.com/likhitjuttada/tinyfish_web_research_agent.git
   cd tinyfish_web_research_agent

2. Install dependencies:

   pip install -r requirements.txt

3. Configure environment variables — create a .env file:

   TINYFISH_API_KEY=your_tinyfish_key
   LLM_PROVIDER=google  # or openai, anthropic
   GOOGLE_API_KEY=your_google_key
   LLM_MODEL=gpt-4-turbo  # optional, if using openai

python main.py "Your research topic here"

The final document will be saved in the output/ directory as a Markdown file.

├── .claude-plugin/          # Plugin manifest files
│   ├── plugin.json          # Plugin metadata
│   └── marketplace.json     # Marketplace listing
├── commands/                # Slash commands
│   ├── research.md          # /tinyfish:research
│   └── cancel-research.md   # /tinyfish:cancel-research
├── hooks/                   # Lifecycle hooks
│   ├── hooks.json           # Hook registration
│   └── orchestrator-stop-hook.sh
├── scripts/                 # Shell scripts
│   ├── setup-research.sh    # Validation + workspace setup
│   └── run-research.sh      # Runs the Python pipeline
├── nodes/                   # Python pipeline nodes
│   ├── executor.py          # TinyFish submission + polling
│   ├── planner.py           # LLM query planning
│   └── synthesizer.py       # Result synthesis
├── models/                  # Pydantic schemas
├── prompts/                 # LLM prompt templates
├── main.py                  # CLI entry point
├── graph.py                 # LangGraph workflow
├── config.py                # Configuration
├── AGENTS.md                # Agent instructions
├── CLAUDE.md                # Points to AGENTS.md
└── LICENSE                  # MIT License

Feel free to open issues or submit pull requests to improve the agent!

MIT