Intel-box

IntelBox

Self-hosted, autonomous market intelligence and outreach research.

Under active development. Features, APIs, workflows, and architecture may change as the project evolves. Not yet production-ready.

---

What it does

Give IntelBox a company name and a category. It researches the company, maps its competitors, finds decision-makers, and drafts outreach — persisting everything to MongoDB so repeat runs are cheap and reports build on prior context.

Rather than a manually-wired workflow where you configure which providers run for every request, IntelBox runs an autonomous agent that decides which research a given company actually needs. A quick brand-perception check might only trigger a web search; a full outreach-ready profile also pulls in competitor mapping and decision-maker discovery — one tool call at a time, each one informed by what the last call turned up.

Two more things shape the project:

• Self-hosted, MIT-licensed. No hosted product, no accounts, no telemetry. You run it, your data stays on your infrastructure.
• Bring your own LLM. Anthropic Claude, OpenAI, or Groq (free tier available) today, behind a provider-neutral interface built so adding another one doesn't touch the orchestrator.

---

Table of contents

• How it works
• Features
• Quick start
• Configuration
• Repository structure
• Demo targets
• Deployment
• License

---

How it works

flowchart LR
    A[Company + category] --> B{Cached profile\nunder 30 days?}
    B -->|Yes| F[Return cached report]
    B -->|No| C[Agentic research loop]
    C -->|picks tools one at a time| D[web_search / competitor_search\nlinkedin_search / scrape_url]
    D --> C
    C --> E[Synthesis: report +\ndecision-makers + outreach drafts]
    E --> G[(MongoDB via MCP)]
    G --> F

1. Cache check. If a fresh company profile (< 30 days old) already exists in MongoDB, it's returned immediately — no research is re-run.
2. Agentic research loop. Otherwise, the configured LLM decides which tools to call — web_search, competitor_search, linkedin_search, scrape_url — one at a time, reacting to each result before deciding the next call. Every call is emitted as a live status step.
3. Skills. Procedural guidance for the agent's two jobs — market research and decision-maker outreach — lives in agent/skills/ as separate markdown files loaded into its system prompt.
4. Synthesis. Once research is sufficient, the same BYO-LLM interface produces the final report, a campaign playbook, decision-maker records, and outreach drafts.
5. Persistence. Everything is written to MongoDB through a separate MCP server (JSON-RPC), not an in-process driver — mcp/operations.py is the only place that talks to it directly.

Full diagram and notes: docs/architecture.md.

---

Features

Implemented

Area	What's there
Orchestration	Agentic tool-calling loop — the model decides which tools to call and when, not a fixed fan-out
LLM	BYO-LLM interface (agent/llm/) — Anthropic Claude, OpenAI, or Groq, via LLM_PROVIDER or auto-detected from whichever API key is set
Skills	Procedural research guidance (agent/skills/) loaded into the agent's system prompt
Web search	Self-hosted SearXNG, structured/parsed JSON results — no per-query API cost
Decision-makers & outreach	Exa-backed LinkedIn search and outreach draft generation, with a fallback path
Scraping	Firecrawl, with a Jina Reader fallback
Persistence	MongoDB via a separate MCP server, with a 30-day company-profile cache
Status tracking	Live run-status stepper — the frontend polls while the backend streams step-by-step updates
Frontend	React + Vite, no router/state library

In progress / known gaps

• The outreach tracker (POST /track) accepts manual status updates from the frontend; automated follow-up detection (IntelBoxRepository.find_follow_ups_due) exists but nothing calls it yet
• Report and campaign-playbook synthesis is a single LLM call producing a flat markdown report — no structured market analysis (Porter's Five Forces, TAM/SAM/SOM, etc.)
• Test coverage is thin; the SearXNG integration test requires a live local instance
• google-genai is listed as a dependency but isn't wired into any code path — vestigial from an earlier iteration

Planned

• Multi-user support
• Authentication and authorization
• Additional research providers and LLM adapters
• CRM integrations

---

Quick start

Backend

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env   # then fill in at least one LLM key

uvicorn api.main:app --reload

Frontend

cd frontend
npm install
npm run dev

Local dependencies for a full end-to-end run

• SearXNG for web search:

  docker compose -f deployment/docker-compose.yml up searxng
  

  The compose file mounts deployment/searxng/settings.yml, which enables the JSON output format web_search.py requires. A SearXNG instance run some other way needs the same setting.
• A MongoDB MCP server for persistence (MONGODB_MCP_SERVER_URL).

Without an LLM provider configured, the orchestrator falls back to calling all four research tools once instead of letting the agent decide.

Tests

pytest                          # full suite
pytest tests/test_api.py        # single file
pytest tests/test_api.py::test_healthcheck   # single test

---

Configuration

# LLM -- bring at least one. LLM_PROVIDER forces a choice; otherwise Anthropic is preferred,
# then OpenAI, then Groq, based on whichever key is set.
ANTHROPIC_API_KEY=
OPENAI_API_KEY=
GROQ_API_KEY=
LLM_PROVIDER=

# Research tools
EXA_API_KEY=
FIRECRAWL_API_KEY=

# Database
MONGODB_URI=
MONGODB_DATABASE=intelbox

Optional:

SEARXNG_URL=http://localhost:8080
MONGODB_MCP_SERVER_URL=http://localhost:8001/mcp

GOOGLE_API_KEY / GOOGLE_CLOUD_PROJECT also appear in .env.example but aren't read by any current code path.

---

Repository structure

agent/        Agentic orchestrator, BYO-LLM interface, procedural skills, and tool adapters
api/          FastAPI routes and service endpoints
mcp/          MongoDB MCP client integrations
models/       Shared Pydantic data models
pipeline/     Orchestration and persistence workflow
output/       Report and playbook generation
frontend/     React application
docs/         Architecture and documentation
deployment/   Deployment configurations
tests/        Test suites

---

Demo targets

Current development testing uses:

Company	Category
Nike	Sportswear
Zepto	Quick commerce
Taj Hotels	Hospitality

Expected output: a market intelligence report, competitor analysis, decision-maker discovery, outreach drafts, and MongoDB-backed persistence — with no code or schema changes needed between categories.

---

Deployment

Self-hosted only — there's no hosted IntelBox product. Deployment assets are included but should be considered experimental until the first stable release.

• Docker Compose configuration (includes SearXNG)
• Cloud Run deployment configuration
• Architecture documentation (docs/architecture.md)

---

License

MIT — see LICENSE.