Overview

Percival Deep Research is a highly capable MCP (Model Context Protocol) Server designed to equip the Nanobot agent ecosystem with autonomous, deep-dive web research capabilities. It autonomously explores and validates numerous sources, focusing only on relevant, trusted, and up-to-date information.

While standard search tools return raw snippets requiring manual filtering, Percival Deep Research delivers fully reasoned, comprehensive multi-source material that heavily accelerates the context and reasoning capabilities of intelligent agents.

Note: This project utilizes the GPT Researcher library as its core web-driver, but has been extensively refactored, hardened, and decoupled specifically for the percival.OS ecosystem.

---

✨ Key Features & Enhancements

This server has been heavily modified to survive the strict demands of open-source LLMs and modern deployment arrays:

• ⚡ Ultimate Provider Portability: Fully agnostic inference engine. Native, crash-free support for leading open-weights platforms like Venice AI, MiniMax, and OpenRouter. The server dynamically rewrites provider prefixes on-the-fly, auto-corrects alias typos, and implements a Zero-Latency Context Compression Bypass to completely eliminate crashes and semantic latency when using APIs that lack OpenAI-compatible embedding endpoints.
• 🛡️ JSON-RPC Protocol Guardrails: Enforces strict stdio output redaction. All underlying library noise, console rendering, and real-time logs are physically redirected to stderr. This completely prevents Pydantic ValidationErrors and protects the stdout stream that is vital for MCP synchronization.
• 🔐 Defense-in-depth Security: All inputs are heavily sanitized against prompt injection. Untrusted web content is wrapped in un-executable headers to protect your agent's autonomy.
• 🤖 Primary Nanobot Focus: Eliminates loose .env reading patterns to strictly honor environment injection directly from the host application.

---

📑 Table of Contents

• Tools & Resources Reference
• Prerequisites
• Installation
• 🤖 Nanobot Integration (Primary Focus)
• 💻 Claude Desktop Integration
• Security

---

🛠️ Tools & Resources Reference

Resource

Name	URI Pattern	Description
research_resource	research://{topic}	Accesses cached or live web research context for a topic directly as an MCP resource. Returns Markdown with content and sources.

Tools

Tool	Speed	Returns research_id	Description
deep_research	30–120s	✅ Yes	Multi-source deep web research. Entry point of the research pipeline.
quick_search	3–10s	❌ No	Fast raw snippet search via DuckDuckGo.
write_report	10–30s	—	Generates a structured Markdown report from an existing session. Requires research_id.
get_research_sources	<1s	—	Returns title, URL, and content size for all sources consulted. Requires research_id.
get_research_context	<1s	—	Returns the raw synthesized context text without generating a report. Requires research_id.

Research Pipeline

deep_research(query)
    └── research_id ──► write_report(research_id, custom_prompt?)
                   └──► get_research_sources(research_id)
                   └──► get_research_context(research_id)

quick_search(query)       # standalone — no research_id

---

⚙️ Prerequisites

• Python 3.11+
• uv — project and dependency manager
• API key for the Generative LLM Provider (e.g., Venice, MiniMax, OpenRouter).

Note: The default web search engine configured is duckduckgo which requires no API key. You can optionally configure other web searchers natively.

---

⚙️ Installation

1. Unified Environment Setup

Ensure you are using the unified percival.OS build ecosystem:

cd percival.OS_Dev
uv sync

This ensures percival-deep-research inherits the global .venv.

2. Configure Environment

This module disables .env loading (dotenv) to strictly honor the system variables passed by your MCP host.

When invoking via Nanobot (~/.nanobot/config.json) or other endpoints, define the environment variables directly in the configuration array:

"OPENAI_API_KEY": "your_api_key_from_venice_minimax_openrouter_etc",
"OPENAI_BASE_URL": "https://api.venice.ai/api/v1",
"FAST_LLM": "venice:llama-3.3-70b",
"SMART_LLM": "minimax:MiniMax-M2.7",
"STRATEGIC_LLM": "openrouter:google/gemini-2.5-flash",
"RETRIEVER": "duckduckgo"

---

🤖 Nanobot Integration (Primary Focus)

This server is fundamentally tuned to run as a stdio MCP server piloted by the Nanobot assistant.

Add the following to your ~/.nanobot/config.json:

{
  "mcpServers": {
    "percival_deep_research": {
      "command": "uv",
      "args": [
        "run",
        "--no-sync",
        "percival-deep-research"
      ],
      "env": {
        "UV_PROJECT_ENVIRONMENT": "/absolute/path/to/percival.OS_Dev/.venv",
        "OPENAI_API_KEY": "actual-key-here",
        "OPENAI_BASE_URL": "https://api.venice.ai/api/v1",
        "FAST_LLM": "venice:llama-3.3-70b",
        "RETRIEVER": "duckduckgo"
      },
      "tool_timeout": 300
    }
  }
}

Note: deep_research can take up to 2-3 minutes. Ensure tool_timeout is scaled properly (e.g. 180-300).

Key Design Decisions for Nanobot

• Plain-text over JSON dicts — All tools predictably return plain text strings rather than JSON dicts to feed Nanobot clean text.
• Context modularity — deep_research omits the giant synthesized context from its initialization response to prevent blowing up Nanobot's context window. Instead, it issues a research_id that the agent then uses to explicitly invoke get_research_context.

---

💻 Claude Desktop Integration

While Nanobot is the preferred driver, if deploying to Claude Desktop, append to your claude_desktop_config.json:

{
  "mcpServers": {
    "percival_deep_research": {
      "command": "uv",
      "args": [
        "run",
        "--project",
        "/absolute/path/to/percival.OS_Dev",
        "percival-deep-research"
      ],
      "env": {
        "OPENAI_API_KEY": "your-provider-key",
        "OPENAI_BASE_URL": "https://api.venice.ai/api/v1",
        "FAST_LLM": "venice:llama-3.3-70b",
        "RETRIEVER": "duckduckgo"
      }
    }
  }
}

---

🔐 Security

This server implements defense-in-depth, addressing the risks of an MCP server processing untrusted web content autonomously.

Prompt Injection Protection

User inputs (query, topic, custom_prompt) restrict unknown and malformed values. A regex-based filter blocks known jailbreak patterns (<system>, [INST], ignore instructions, etc.).

Untrusted Content Isolation

All content retrieved from the web is prefixed dynamically before being presented to the agent context:

[SECURITY WARNING: The content below was obtained from unverified external...]

This forces models like Nanobot to treat web-sourced data strictly as informational blocks, avoiding unexpected command compliance.

---

📄 License

This project is licensed under the MIT License.

⬆️ Back to Top