II Deep Research
Search the web, assess findings, and produce referenced research reports for difficult questions.
- Source repo
- Intelligent-Internet/ii-researcher
- Stars
- ★ 505
- Last updated
- 4mo ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 47/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude.aiOpenAI API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Researchers who need to investigate a complex question across web pages and produce a sourced synthesis.
- Not a fit if
- Users unwilling to configure model, search, and scraping provider credentials
- Teams that need research to run without external network access
- Website
- ii.inc/web/blog/post/ii-researcher ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-07-02latest release v0.1.5 (2025-05-09)
- Source review
- 47/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
II-Researcher is a deep-search agent from Intelligent Internet that searches the web, visits pages, and reflects on evidence through a multi-step reasoning loop. Its stack includes a Python library and CLI, a FastAPI backend with SSE streaming, a Next.js web UI, and an MCP server. Search and scraping providers are configurable, while LiteLLM routes model calls to OpenAI, DeepSeek, Gemini, OpenRouter, or OpenAI-compatible self-hosted endpoints. The report builder produces referenced answers in basic or advanced form and supports multiple languages. It can be run locally or with Docker Compose, and requires credentials for the chosen model, search, and scraping services.
A user submits a question through the CLI --question option or the web UI. II-Researcher runs a think, act, and reflect loop, searches with SerpAPI, Tavily, or Jina, and reads pages through Firecrawl, Browser, BeautifulSoup, Tavily Extract, or Jina; the README also describes handling for PDFs and YouTube. Retrieved material can be filtered by embedding-based compression and an optional LLM compressor before the model continues reasoning. Once it has gathered enough evidence, the report builder returns a referenced answer in basic or advanced form. The CLI can stream output and save a Markdown report; the FastAPI /search SSE endpoint and web UI stream the research process, while the MCP server connects to Claude Desktop and other MCP clients.
- Researchers who need to investigate a complex question across web pages and produce a sourced synthesis.
- Analysts who want configurable search and scraping providers to gather material from pages, PDFs, or YouTube.
- Individuals who want to ask questions in a CLI, follow streamed output, and save reports as Markdown.
- Teams deploying a FastAPI backend and Next.js UI locally to provide a streaming research interface.
- Claude Desktop users who want to connect web research as an MCP tool.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install ii-researcherpip install -epip install litellmdocker compose up --build -d
- Environment variables
OPENAI_API_KEYTAVILY_API_KEYSERPAPI_API_KEYFIRECRAWL_API_KEYNEXT_PUBLIC_API_URL- Config files
litellm_config.yamlfrontend/.env.env.example- Runtime
- Python 3.10+
- Install notes cover
- Docker
How do you install or deploy this agent?
Local installation requires Python 3.10+. Install from PyPI or clone the repository and install in editable mode:
pip install ii-researchergit clone https://github.com/Intelligent-Internet/ii-researcher.git
cd ii-researcher
pip install -e .Before running, configure model credentials and the LiteLLM endpoint, plus keys for the selected search and scraping providers. The README examples use OPENAI_API_KEY, TAVILY_API_KEY or SERPAPI_API_KEY, and FIRECRAWL_API_KEY when using Firecrawl; set SEARCH_PROVIDER and SCRAPER_PROVIDER as well. Install LiteLLM and prepare litellm_config.yaml before starting its proxy:
pip install litellm
litellm --config litellm_config.yamlFor Docker deployment, Docker and Docker Compose are required, and the environment variables must be configured first:
docker compose up --build -dHow do you use this agent?
After installation and configuration of model, search, and scraping services, run a first research question from the CLI:
python ii_researcher/cli.py --question "your question here" --streamUse --report-type basic or --report-type advanced to choose the report style, and --save-report to save a Markdown file. For the web mode, start the API:
python api.pyThe API runs at http://localhost:8000 by default, with /search streaming research steps over SSE. Set NEXT_PUBLIC_API_URL=http://localhost:8000 in frontend/.env, then run the frontend:
cd frontend
npm install
npm run devTo connect an MCP client, copy and fill in .env.example, then run:
mcp install mcp/server.py -f .envWhat are this agent's strengths and limitations?
- Offers SerpAPI, Tavily, and Jina for search, plus Firecrawl, Browser, BeautifulSoup, Tavily Extract, and Jina for page scraping.
- Routes models through LiteLLM, including OpenAI, DeepSeek, Gemini, OpenRouter, and OpenAI-compatible self-hosted endpoints.
- Provides several entry points: Python library, CLI, streaming FastAPI API, Next.js web UI, and MCP server.
- Uses embedding-based compression and an optional LLM compressor to limit how much long-page content enters the model context.
- The quick start requires several setup steps for the model proxy and search and scraping provider credentials.
- Model, search, and scraping calls depend on network access and the selected services; a fully offline mode is not documented.
- Local web development requires Node.js and npm, while container deployment requires Docker and Docker Compose.
- The reported 84.12% Frames score is specifically tied to DeepSeek-R1-0528 and that dataset, so it does not establish results for other models or tasks.
How does this agent compare with similar options?
The README says the project draws prompt inspiration from node-DeepResearch and gpt-researcher, and uses a scraper tool from gpt-researcher; it does not provide a direct feature or performance comparison with either project.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| II Deep Research This agent | 47 · Major gaps | CLIFree + model costs | ★ 505 | 4mo ago | Python | Claude.ai |
| Firecrawl CLI | 72 · Some gaps | — | ★ 694 | 2d ago | TypeScript | Codex · Claude Code |
| Create Context Graph | 62 · Some gaps | CLIFree + model costs | ★ 750 | 1mo ago | Python | ChatGPT · Claude Code · Claude.ai · OpenAI API · Claude API |
| wigolo | 71 · Some gaps | CLIFree + model costs | ★ 5.6k | 11d ago | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The code shows the agent searching the web, fetching pages, and sending the question and research process to configured models and external search or scraping services. The README configures API keys through environment variables; SECURITY.md says scraped content is untrusted and warns that the API has permissive default CORS. Per-action permission controls, user confirmation, and clear data-retention details are not shown, so least privilege, confirmation, data-flow transparency, and sensitive-data handling score low. Results can optionally be saved as Markdown, offering limited recoverability. Reports are intended to include references, but the packet does not establish that every citation supports a specific claim. Dependencies have lower bounds and one exact pin, but no security-audit evidence is supplied; external effects are only partly described.
The README, CLI, and Agent constructor differ on the default report type: the CLI defaults to basic while the Agent defaults to advanced. The project supports several model, search, and scraping services, but ordinary use depends on external credentials and availability. The Agent catches some timeouts, cancellations, parsing errors, and tool errors and returns error text. Other exceptions lack a demonstrated uniform handling path, and the main loop has no explicit step limit in the supplied code, so error handling and availability receive partial scores.
The materials target developers and research use, and describe CLI, Web, MCP, Docker, self-hosted models, and multiple search, scraping, and report options. The search tool is explicitly framed around research questions requiring current web information. Unsuitable scenarios, capability boundaries, and search-quality limitations are not fully described. Python requirements, installation, environment variables, and deployment options are documented, though setup depends on external services.
The README is organized around overview, architecture, installation, usage, and configuration, with CLI, Web, MCP, and Docker examples; naming and module structure are generally clear. The LICENSE file contains the full Apache 2.0 terms and earns full marks. News entries and the package version provide limited version information, but no complete changelog is shown. SECURITY.md notes deployment hardening needs and provides a vulnerability-reporting path, though limitations are not comprehensively documented. Contact information, contribution guidance, and security reporting help identify maintenance responsibility; release workflow files provide some evidence of an update path.
The framework combines multi-step search, page reading, context compression, report generation, and streaming in a self-hostable package, with useful potential for research tasks. Results are accessible through CLI, Web, or MCP and can be saved as Markdown. Quality depends heavily on the selected model, search services, and retrieved content, while setup requires several services and credentials, limiting the cost-benefit score.
The architecture, configuration variables, CLI arguments, interfaces, and dependencies have concrete counterparts across the README, source, and project configuration, making the main feature claims partly traceable. The supplied tests cover report generation, the client, and action parsing, but the sample is limited and uses mocks; it does not independently substantiate the benchmark or end-to-end performance claims. The README states a benchmark result, but the supplied materials do not include its method or replication evidence. Fact and inference are not consistently distinguished, so these criteria do not receive high scores.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Research questions and fetched page content are sent to configured models and external services; review those services' data handling and restrict API access, CORS, and network exposure before deployment.
- The supplied static materials do not independently verify report citations, benchmark results, or end-to-end performance; actual quality depends on model and service configuration.
- The project relies on many external packages and services, and the packet provides no dependency security audit; review dependencies and credential configuration before deployment.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
First Release * New open-source framework for complex, multi-step search with reasoning tasks— a powerful, transparent alternative to proprietary systems * Two methods were implemented: finite state machines and tool-augmented thought processes with advanced reasoning models (R1 & QwQ) * Consistently delivers strong performance on GAIA and FRAMES benchmarks * MCP support * ii-researcher Pypi package * CI/CD for testing, linting and pypi publication What Changed * init release v0.1 by @PhungVanDuy in https://github.com/Intelligent-Internet/ii-researcher/pull/1 …View on GitHub ↗
FAQ
Do I need paid APIs to use it?
Can it run entirely on my machine?
Which models and search providers can I connect?
Are sources included in the reports?
Is II Deep Research open source?
Is II Deep Research safe to use?
Is II Deep Research still maintained?
Who makes II Deep Research?
What is II Deep Research's official website?
Thanks — we use this to improve the page.