Open Research Agent (ORA)

Open Research Agent (ORA) is an open-source multi-agent research CLI. ORA plans research, searches and scrapes web sources, synthesizes findings, and optionally uses an adversarial reviewer for higher-intensity research.

Current release: 0.3.0

What ORA does

ORA turns a research question into a sourced markdown report:

1. A supervisor drafts a research plan.
2. The researcher searches and scrapes web sources.
3. At intensity 3+, an LLM extraction layer pulls key claims, data, and entities from each source.
4. The writer synthesizes findings into a report.
5. For intensity levels 3 and above, an adversarial reviewer audits the draft.

Current backend support

ORA 0.3.0 supports two LLM backends:

• DeepSeek API (default), models like deepseek-v4-flash and deepseek-v4-pro
• OpenRouter, an OpenAI-compatible gateway to many models, e.g. anthropic/claude-3.5-sonnet via openrouter:anthropic/claude-3.5-sonnet

Search and scraping use Firecrawl.

Installation

Install from PyPI:

pip install open-research-agent

The primary CLI command is open-research-agent. The shorter ora command is also installed as a convenience alias.

Install from source for development:

git clone https://github.com/cameronmpalmer/open-research-agent.git
cd open-research-agent
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Configuration

Set the required API keys:

export DEEPSEEK_API_KEY="your-deepseek-api-key"
export FIRECRAWL_API_KEY="your-firecrawl-api-key"
export OPENROUTER_API_KEY="your-openrouter-api-key"

Create a default config file:

open-research-agent config --init

Show the active configuration and intensity levels:

open-research-agent config --show

The config file is stored at:

~/.ora/config.yaml

See CONFIG.md for the complete reference of the config.yaml format, the available settings, and their environment-variable equivalents.

Using a provider prefix

Any model name can carry a provider:model prefix to select the backend:

open-research-agent research "..." --model openrouter:anthropic/claude-3.5-sonnet

Without a prefix, calls use the default provider (deepseek by default). The default can be changed in config.yaml under provider.default. An unknown prefix (e.g. openai:gpt-4.1) warns and falls back to the default provider. Each provider reads its API key from the environment variable named in its api_key_env (DEEPSEEK_API_KEY for deepseek, OPENROUTER_API_KEY for openrouter), and can be configured under providers: in config.yaml:

provider:
  default: deepseek
providers:
  deepseek:
    base_url: https://api.deepseek.com
    api_key_env: DEEPSEEK_API_KEY
  openrouter:
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    headers:
      HTTP-Referer: https://github.com/cameronmpalmer/open-research-agent
      X-Title: ORA

The legacy top-level deepseek_base_url key is still honored when no providers: section exists.

Quick start

Preview a research plan without running the full pipeline:

open-research-agent plan "What are the tradeoffs between Rust and Go for backend services?"

Run a standard research task:

open-research-agent research "What are the tradeoffs between Rust and Go for backend services?" --intensity 2

Run deeper research with adversarial review:

open-research-agent research "What are the tradeoffs between Rust and Go for backend services?" --intensity 4

Save to an explicit file:

open-research-agent research "AI memory systems" --output ai-memory-systems.md

Print only to stdout and do not save a report file:

open-research-agent research "AI memory systems" --no-save

Intensity levels

ORA supports five research intensity levels:

Levels 3, 4, and 5 use the adversarial reviewer by default.

CLI flags

Output files

By default, open-research-agent research saves a timestamped markdown report in the current directory. Generated research reports are local outputs and should not be committed to the repository.

Use --output to choose a specific path, or --no-save to print the report without writing a file.

Development

ORA requires Python 3.10. Later versions (3.11+) may encounter incompatibilities with the LangChain/LangGraph ecosystem.

The repository ships a Makefile that manages a local virtual environment automatically. From a clean clone:

make setup

Run the test suite (pass T=tests/path for a focused file, ARGS="-x -q" for pytest options):

make test

Lint and auto-format:

make lint
make format

Build the package (wheel and sdist into dist/):

make build

Run a research task from the Makefile. Reports are saved to reports/ by default:

make research QUERY="What are the tradeoffs between Rust and Go for backend services?"

make research QUERY="..." INTENSITY=4 runs at a higher intensity level; OUTPUT=path.md overrides the output path. make plan QUERY="..." previews a research plan without running it. make check runs lint then tests. See make help for the full target list.

Install the git pre-commit hook (run once per clone):

make install-hooks

The hook runs make precommit (lint, tests, and build) before every commit. Bypass it for a quick commit with git commit --no-verify.

Run the CLI locally:

open-research-agent --help
ora --help
python -m ora --help

See CONTRIBUTING.md for contributor setup and repository hygiene expectations.

License

MIT License. See LICENSE.