Due Diligence Agents
Open-source forensic M&A due diligence: 13 AI agents read your data room across 9 domains, connect findings no single reviewer catches, and trace every one to an exact page and quote.
- Source repo
- zoharbabin/due-diligence-agents
- Stars
- ★ 139
- Last updated
- 3d ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 69/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A corp dev team screening hundreds of targets runs
--quick-scanfor a GREEN/YELLOW/RED red-flag signal in minutes before committing to full analysis. - Not a fit if
- Users expecting it to replace qualified legal/financial advisors' conclusions
- Organizations that cannot run Python 3.12+ or adapt their runtime
- Website
- zoharbabin.com/due-diligence-agents ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-17latest release v1.22.3 (2026-11-17)
- Source review
- 69/100 · Some gaps
What does this agent do, and when should you use it?
Due Diligence Agents is an open-source M&A due diligence system driven by the dd-agents CLI. It reads an entire data room (PDF, Word, Excel, PowerPoint, images) and analyzes every document through 9 parallel domain specialist agents: Legal, Finance, Commercial, ProductTech, Cybersecurity, HR, Tax, Regulatory, and ESG. A 38-step Python pipeline with 5 blocking quality gates runs cross-domain analysis using symbolic trigger rules, plus 4 orchestration agents including a Judge and an Executive Synthesis agent. Every finding carries severity, an exact page-and-quote citation, and cross-references, producing an interactive HTML report, a 16-sheet Excel report, and per-subject JSON. Analysis runs locally; documents leave your machine only as API calls to your configured LLM endpoint, with no telemetry and a read-only data room. Built by Zohar Babin, the author reports it has been battle-tested across multiple completed acquisitions.
The core flow: dd-agents run deal-config.json starts the pipeline, where a Python orchestrator runs 38 steps and 9 domain specialists analyze every data-room document in parallel, each tracing findings to exact page, section, and quote. A cross-domain analysis layer uses symbolic trigger rules to detect inter-domain dependencies (e.g., a termination clause in one contract plus revenue concentration risk on the same subject); an optional Judge agent adversarially spot-checks findings; merge & audit applies dedup, numerical checks, and citation verification across 31 QA checks; an Executive Synthesis agent calibrates severity and emits a Go/No-Go signal. Output lands in {data_room_path}/_dd/forensic-dd/runs/latest/ including report/dd_report.html (interactive report with a clause library and market-norm comparisons), report/dd_report.xlsx (16 sheets), audit.json, and per-subject JSON. Companion tools: dd-agents search prompts.json --data-room ./data_room answers targeted contract questions with citations and verification scores in Excel; --quick-scan gives GREEN/YELLOW/RED triage; and post-run commands include chat, query, cost, portfolio compare, diff, and lineage. Agent personas are customized via dd-config/agents/*.md markdown files with no code changes.
- A corp dev team screening hundreds of targets runs
--quick-scanfor a GREEN/YELLOW/RED red-flag signal in minutes before committing to full analysis. - A PE firm runs the full pipeline on portfolio targets and uses
dd-agents portfolio addandportfolio compareto track and compare risk across deals. - A legal team uses
dd-agents searchwith a JSON prompts file to ask specific questions across hundreds of contracts (e.g., is consent required on change of control) and get an Excel report with answers, citations, and verification scores. - A deal team preparing an IC memo or negotiation checklist filters findings by severity in the interactive HTML report and verifies each to its exact page and quote.
- An M&A advisor on a compressed timeline needs the cross-domain risk picture — legal, financial, and commercial flags on the same subject connected — that siloed advisor reports never assemble.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install 'dd-agents[pdf]'brew install zoharbabin/due-diligence-agents/dd-agentsdocker pull zoharbabin/due-diligence-agents:latestpip install dd-agents # core only (no PDF extraction)
- Environment variables
ANTHROPIC_API_KEY- Config files
deal-config.json- Runtime
- Python 3.12+
- Install notes cover
- macOS · Docker
How do you install or deploy this agent?
Prerequisites: Python 3.12+ and an Anthropic API key (or AWS Bedrock, Google Vertex AI, or an Anthropic-compatible gateway).
Recommended install (includes PDF extraction):
bash
pip install 'dd-agents[pdf]'Alternatives:
bash
brew install zoharbabin/due-diligence-agents/dd-agents
docker pull zoharbabin/due-diligence-agents:latest
pip install dd-agents # core only (no PDF extraction)
pip install 'dd-agents[vector]' # + ChromaDB semantic search
pip install 'dd-agents[ocr]' # + OCR for scanned docs (English)Note: the optional [pdf] extra installs pymupdf, which is AGPL-3.0 licensed — if you redistribute software bundling pymupdf, AGPL copyleft terms apply.
How do you use this agent?
Set your key and run:
bash
export ANTHROPIC_API_KEY="sk-ant-..."
dd-agents auto-config "Buyer Corp" "Target Inc" --data-room ./data_room
dd-agents doctor --config deal-config.json
dd-agents run deal-config.jsonResults appear at {data_room_path}/_dd/forensic-dd/runs/latest/report/dd_report.html — open it in your browser.
Quick triage without the full pipeline:
bash
dd-agents run deal-config.json --quick-scan --model-profile economyGenerate a config without any API calls:
bash
dd-agents init --data-room ./data_roomOrganize the data room into folders by subject or counterparty; PDFs, Word, Excel, PowerPoint, and images are supported, with OCR for scanned PDFs. The minimal config requires config_version, buyer.name, target.name, deal.type, deal.focus_areas, and data_room.path.
What are this agent's strengths and limitations?
- Cross-domain cross-referencing is the core differentiator: symbolic trigger rules automatically connect findings from different domains on the same subject (e.g., termination clause + revenue concentration risk) — exactly what siloed reviewers miss.
- A substantial trust layer: 31 QA checks, 5 blocking quality gates that halt the pipeline on failure,
numerical_manifest.jsontracing every financial figure to source, and every finding bound to an exact page and quote. - No vendor lock-in: model and provider are switchable via env config (Anthropic API, AWS Bedrock, Google Vertex AI, or any model behind an Anthropic-compatible gateway — GPT, Gemini, DeepSeek, local), verifiable with
dd-agents doctor. - Explicit privacy design: local execution, no telemetry, read-only data room, and API keys never stored in output artifacts; agent personas are markdown-customizable with safety rules that can never be removed.
- Requires Python 3.12+ and paid LLM API calls — a full pipeline over a large data room can incur significant token costs (
dd-agents costgives per-provider rollups, but the source gives no typical cost range). - The optional
[pdf]extra depends on AGPL-3.0 pymupdf; redistributing software that bundles it triggers copyleft obligations. - Explicit positioning statement: the tool does not replace professional advisors — legal, financial, and regulatory conclusions must be made by qualified professionals.
- Billing through a model provider means document content is sent as API calls to a cloud LLM endpoint; deals with sensitive data require reviewing the data-handling details in SECURITY.md.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Due Diligence Agents This agent | 69 · Some gaps | CLIFree + model costs | ★ 139 | 3d ago | Python | Claude API |
| SwarmVault | 75 · Good | CLIFree + model costs | ★ 737 | 4mo ago | TypeScript | Codex · Claude Code |
| Obsidian Wiki — A Digital Brain You Grow with Your AI Agent | 73 · Some gaps | Agent plugin / skillFree | ★ 3.8k | today | Python | Codex · Claude Code |
| YantrikDB Memory for Hermes | 72 · Some gaps | Agent plugin / skillFree | ★ 97 | 16d ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The docs describe local execution, document transmission only to the configured model endpoint, read-only data-room access, outputs under _dd/, and credentials excluded from artifacts. These provide concrete support for least privilege, data-flow transparency, and sensitive-data handling. Config validation and dry-run offer limited pre-run control, but no per-request confirmation mechanism is shown. Dependencies have lower bounds and optional extras, but no lockfile or supply-chain review evidence is included.
The README describes quality gates, citation checks, and resumable runs; troubleshooting lists exit codes, common errors, and recovery steps. The docs explain the 13-agent, 9-domain, 38-step design, but the packet lacks implementation detail to independently verify those claims, so full marks are not warranted.
The product targets corp-dev, private equity, legal teams, and advisors, with full diligence, quick triage, targeted contract search, and post-run Q&A. The packet describes configurable agents, models, domains, and providers, and says the tool does not replace professional advisors. Capability boundaries and trigger rules remain mostly asserted at a high level, limiting the score.
Documentation is organized into getting started, configuration, and troubleshooting topics, with concrete installation steps and configuration examples. Apache-2.0 is explicit, and the package has a version and release-history link. The packet shows neither an actual changelog nor a clear maintenance commitment. Some FAQs and limitations are covered, but not comprehensively.
Outputs include interactive HTML, a 16-sheet Excel report, per-subject JSON, citations, and audit information, with filtering, comparison, and follow-up queries. Cross-domain synthesis and citation checks offer plausible added value. Cost guidance is limited to model tiers and a cost-rollup command, with insufficient quality or time-savings evidence to assess return thoroughly.
The product claims page-and-quote citations and describes citation checks, deterministic rechecks for severe quotes, numerical audits, and audit artifacts, supporting traceability. Cross-domain links have described rules and review steps, but the packet does not include their implementation or sample results for independent verification. The docs distinguish tool analysis from professional judgment, but do not sufficiently show how outputs label facts versus inference.
- Diligence documents are sent to the configured model provider; review that provider’s data handling and retention terms before use.
- Treat reports as support for advisor review; the docs say legal, financial, and regulatory conclusions should be made by qualified professionals.
- This assessment uses only the supplied static file sample; no code was run and marketing claims were not independently verified.
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
What changed A small fix release for the desktop app. Desktop app - Remove now asks "Remove this project from the list?" before it acts. In v1.22.2 one click removed the project. Your data room files were never touched. - Findings show readable categories ("Renewal terms", not renewal_terms). - Settings and the setup card show readable provider names ("Amazon Bedrock", not bedrock). Quality - The upgrade check (desktop-upgrade.yml) now smoke tests the old app with its own release's script. Before, it failed older builds for checks they never had. …View on GitHub ↗ -
Release notes excerpt
What changed A fix and polish release for the desktop app, plus one security hardening that applies to the CLI too. Desktop app - The signed macOS app no longer changes itself after the first launch, so Gatekeeper keeps accepting it. Python files are precompiled and the service writes no new files inside the app. - Ask shows answers as formatted text (tables, bold, lists, code) instead of raw markdown. - Ask counts the review's findings before the first question. - The run panel shows readable agent names. - Old projects show the same risk label as the report. …View on GitHub ↗ -
Release notes excerpt
Changes since v1.22.0 - chore(release): bump version to 1.22.1 (#329) - docs: cask install line, first release status, tap repo in the release steps (#328) - fix: flatten downloaded release artifacts before naming the assets (#327) - Update Homebrew formula to 1.22.0 --- Install ```bash pip install dd-agents==1.22.1 ``` Docker ```bash docker pull zoharbabin/due-diligence-agents:1.22.1 ``` What's Changed * fix: flatten downloaded release artifacts before naming the assets by @zoharbabin in https://github.com/zoharbabin/due-diligence-agents/pull/327 …View on GitHub ↗
FAQ
Does my data leave my machine during analysis?
Am I locked into Anthropic models?
dd-agents doctor.What does running it cost?
--quick-scan with --model-profile economy offers a low-cost first read, and dd-agents cost gives a per-provider/per-model cost rollup for each run.What happens if quality fails mid-run?
dd-agents assess ./data_room checks data room quality before a run.Can it replace my lawyers and financial advisors?
Is Due Diligence Agents open source?
Is Due Diligence Agents safe to use?
Is Due Diligence Agents still maintained?
Who makes Due Diligence Agents?
What is Due Diligence Agents's official website?
Thanks — we use this to improve the page.