Part 10: Automation and agentic coding for OSINT

AI coding agents, scripts and automation for OSINT work, and the rules that apply when a tool reads everything you show it.

Most investigators now write small scripts. A scraper for a registry. A screen of a supplier list against OpenSanctions. A notebook that joins trade data to company records. AI coding agents make that work much faster. They also read everything you show them, so the same rules apply as for chat assistants. The rubric for a coding tool follows the model behind it and the account you use.

10.1 Agentic coding tools

Tool What it is Pricing (October 2026) Where your code goes
Claude Code (Anthropic) Terminal and IDE agent that reads, edits and runs code Included with Pro, Max, Team and Enterprise; also via the API, Bedrock, Vertex AI and Foundry1 On Free, Pro and Max accounts, sessions fall under the consumer training setting (Part 2.1). Team, Enterprise and API are excluded from training
OpenAI Codex Cloud sandboxes plus a command-line agent Included with ChatGPT plans Follows the ChatGPT plan’s data terms; Enterprise and zero-retention API exclude training
Gemini CLI (Google) Open-source terminal agent Free tier with a Google account; paid via Vertex AI or AI plans Consumer use follows Gemini’s terms; Vertex AI does not train on your data
Cursor AI-first code editor Hobby free. Pro $20. Pro+ about $60. Ultra about $200. Teams $40 a user2 Privacy Mode stops Cursor and its model providers training on your code. SpaceX exercised an option to buy Cursor’s parent company in June 2026; reports differ on whether the deal has closed
Cline Open-source agent extension for VS Code Free, Apache 2.0. Teams about $20 a user Sends code and prompts to whichever model API you configure. Pointed at a local Ollama model it is fully local; pointed at a cloud API it inherits that vendor’s terms
GitHub Copilot Code completion and agents in the IDE and on GitHub Free. Pro $10. Pro+ $39. Max $100. Business $19 and Enterprise $39 a user. Usage billed in AI credits since June 20263 Business and Enterprise exclude training; individual plans have an opt-out
Replit Browser-based development environment with an agent Core $20 a month. Pro $1004 Everything runs and is stored on Replit’s cloud

Verified 2026-10-09. Continue.dev, which earlier editions recommended, was acquired by Cursor in June 2026 and wound down; its cloud data was deleted after 15 July 2026.5

How we use them safely:

  • Generic variable names and comments. The code can describe a technique without naming a target.
  • No case data in prompts or test fixtures. Use synthetic examples.
  • Read every line the agent wrote before running it, and run it in a virtual machine or container.
  • Keep secrets in environment variables the agent cannot read.

10.2 Workflow automation

Platform Pricing (October 2026) AI in the product Our view
n8n Self-hosted is free under the Sustainable Use licence. Cloud Starter €20 a month, Pro €50, Business €6676 Seventy or more AI nodes, including an AI Agent node that can call any model7 Our default. Self-host it and point its AI nodes at a local model. Licensed self-hosted editions contact a licence server daily and send telemetry by default; turn both off. Earlier editions said n8n had “no AI”, which is wrong
Zapier Free with 100 tasks a month. Professional $19.99 a month billed yearly. Team $698 AI agents, Copilot, chatbots and an MCP server Cloud only; everything passes through Zapier
Microsoft Power Automate Premium about $15 a user a month Copilot throughout; administrators can restrict it Fine inside a Microsoft 365 tenant you control
IFTTT Pro about $3 to $5 a month Little Simple triggers only
Airtable Team $20 an editor a month billed yearly ($24 monthly). Business $45 ($54) AI fields and summaries Cloud database; keep case data out
Notion Plus about $10 to $12 a month. Business about $20 to $249 Notion AI is included only from the Business plan up; Enterprise adds zero data retention with model providers Good for course notes, not for case files

We use self-hosted alternatives where we can. Obsidian for notes: free, including at work, since February 2025, with an optional $50 commercial licence.10 Baserow (MIT core) or NocoDB (moved to a Sustainable Use licence in January 2026) instead of Airtable. n8n instead of Zapier.

10.3 Private search over investigation corpora

Once a case has a few hundred documents, you want to ask questions across all of them. The usual pattern is retrieval-augmented generation (RAG). You split documents into chunks and turn each chunk into an embedding, a numeric fingerprint of its meaning. You store the embeddings in a vector database. When you ask a question, the system retrieves the closest chunks and hands them to a model along with your question. The model never needs the whole corpus, and if the model is local nothing leaves your machine.

One rule covers the OPSEC. An embedding API is a data channel like any other. Sending a document to a cloud embedding service exposes it in the same way as sending it to a cloud chat model. For sensitive corpora, embed locally.

Layer Cloud options Local options
Embeddings Voyage AI (owned by MongoDB; the voyage-4 family launched January 2026).11 MongoDB’s documentation says data sent to its Atlas embedding API may be used to improve Voyage models unless an organisation administrator opts out, and we found no zero-retention option.12 OpenAI text-embedding-3, Cohere Embed, Gemini Embedding, Mistral Embed, each under that vendor’s API terms nomic-embed-text, BGE-M3 (multilingual), Qwen3-Embedding, EmbeddingGemma (Google, on-device), IBM Granite Embedding, jina-embeddings-v4 (text and images). All run through Ollama, LM Studio or a Python library
Vector store Managed services from the database vendors pgvector (inside PostgreSQL), Qdrant, Chroma, LanceDB, sqlite-vec (a single file)
Reranking Voyage and Cohere rerankers Open rerankers from BGE and others
Document conversion Vendor OCR APIs Docling and Granite-Docling (IBM, open source) for PDFs, scans and tables13
Groundedness check Vendor guardrail APIs Granite Guardian can check whether an answer is supported by the retrieved text14

A minimal local setup needs four pieces. Docling converts the files. A local embedding model runs through Ollama. sqlite-vec or Chroma holds the vectors. A 20 to 30 billion parameter local model answers. AnythingLLM and Open WebUI (Part 3.3) package most of this behind a user interface. Whatever you build, keep the retrieved passages visible in every answer so the analyst can check them, and never let the model answer from memory alone.

Model and product names verified 2026-10-12 for existence; test any of them on your own documents before you trust them.

10.4 Safe development checklist

  1. Develop in a virtual machine or container, with a local Git repository.
  2. Use AI coding tools on an account whose terms you have read. Prefer a local model or an enterprise account for anything that touches case data.
  3. Never include real targets, identifiers or credentials in prompts, code, comments or test data.
  4. Review all generated code. Pay attention to anything that makes a network request.
  5. Respect rate limits and terms of service for the sites you collect from. Automated collection can breach a site’s terms and, in some jurisdictions, the law.
  6. Encrypt data at rest, log access, and back up securely.

Footnotes

  1. Anthropic. “Pricing.” https://claude.com/pricing. Accessed 2026-10-09.↩︎

  2. Cursor. “Pricing.” https://cursor.com/pricing ; SatNews (13 August 2026) on the SpaceX acquisition, https://satnews.com/2026/08/13/spacex-finalizes-regulatory-procedures-to-close-60-billion-acquisition-of-ai-platform-cursor/. Accessed 2026-10-09.↩︎

  3. GitHub. “Copilot plans.” https://github.com/features/copilot/plans. Accessed 2026-10-09.↩︎

  4. Replit. “Pricing.” https://replit.com/pricing. Accessed 2026-10-09.↩︎

  5. The New Stack (June 2026). “Cursor acquires Continue.” https://thenewstack.io/cursor-acquires-continue-coding/. Accessed 2026-10-09.↩︎

  6. n8n. “Pricing.” https://n8n.io/pricing/. Accessed 2026-10-09.↩︎

  7. n8n Docs. “LangChain in n8n.” https://docs.n8n.io/build/integrate-ai/langchain-in-n8n. Accessed 2026-10-09.↩︎

  8. Zapier. “Pricing.” https://zapier.com/pricing. Accessed 2026-10-09.↩︎

  9. Notion. “Pricing.” https://www.notion.com/pricing. Accessed 2026-10-09.↩︎

  10. Obsidian. “License overview.” https://obsidian.md/license. Accessed 2026-10-12.↩︎

  11. Voyage AI (15 January 2026). “voyage-4.” https://blog.voyageai.com/2026/01/15/voyage-4/ ; Microsoft Azure AI model catalog, https://ai.azure.com/catalog/models/voyage-4-embedding-model. Accessed 2026-10-12.↩︎

  12. MongoDB Docs. “Model training data” (Voyage AI). https://www.mongodb.com/docs/voyageai/management/model-training-data/. Accessed 2026-10-12.↩︎

  13. IBM Granite model collection on Hugging Face, https://huggingface.co/ibm-granite ; I Programmer (October 2025), “IBM launches Granite version 4.0 and Granite-Docling”, https://www.i-programmer.info/news/80-java/18409-ibm-launches-granite-version-40-and-granite-docling-.html ; DataNorth (August 2026) on Granite 4.2, https://datanorth.ai/news/ibm-releases-granite-4-2. Release details for 4.1 and 4.2 are from secondary reporting; check the model cards. Accessed 2026-10-12.↩︎

  14. IBM Granite model collection on Hugging Face, https://huggingface.co/ibm-granite ; I Programmer (October 2025), “IBM launches Granite version 4.0 and Granite-Docling”, https://www.i-programmer.info/news/80-java/18409-ibm-launches-granite-version-40-and-granite-docling-.html ; DataNorth (August 2026) on Granite 4.2, https://datanorth.ai/news/ibm-releases-granite-4-2. Release details for 4.1 and 4.2 are from secondary reporting; check the model cards. Accessed 2026-10-12.↩︎