Skip to content

🤖 On-Premise Custodian Agents

A custodian agent is not an assistant: you do not talk to it to ask questions. You assign it recurring or event-driven tasks — keeping a vault up to date, detecting obsolete code, proposing sourced fixes — and it works autonomously while leaving the final decision to a human.


🧭 Key concept: Human-in-the-loop

A sovereign custodian agent does not “commit”, “merge”, or “publish” without human validation.

The typical cycle is:

  1. Trigger — scheduled (cron) or event-driven (new file, open PR)
  2. Execution — the agent reads, analyzes, generates proposals
  3. Branch + diff — changes are isolated in a dedicated Git branch
  4. Report — the agent produces a readable summary (PR description, email, message)
  5. Human validation — you merge, or you don’t
  6. Publication — only if approved

Autonomy levels vary: from “report only” up to “automatic commit on a feature branch” — but never automatic merge or publication to main without explicit agreement.

Human validation protects the output (merge, publication), not the input: what the agent reads and executes before review remains the weak link. OWASP now ranks “Excessive Agency” third in its Top 10 for LLM applications (2026 edition, per its 2026-09-01 announcement), the CNIL and the Conseil de l’IA et du Numérique published an exploratory note on agentic AI and personal data in July 2026, and an Ollama vulnerability (CVE-2026-102697, fixed in 0.31.2) showed that a human-approved shell command could be extended by the model. See Vision and Excessive agency123.


📋 Reference pages

Understand

Go further


🛠️ Solution pages

ToolRole in the stackSovereignty
AiderCode agent, terminal-first, supports Ollama — development frozen (see its page)✅ if local, ⚠️ unmaintained
OpenHandsDocker/sandbox agent, Agent Canvas (scheduled automations, third-party agents via ACP), local models supported4⚠️ configurable
LiteLLM / lexiconUnifying proxy (Ollama, vLLM, cloud); vulnerabilities exploited in 2026, patch every month5✅ if local-only and up to date
SearXNGSelf-hosted meta-search, no API key✅ web privacy
Cursor CLIPowerful MVP, but Cursor cloud routing (SpaceX since August 2026)6❌ strict

🔗 See also

📚 Sources and References

Footnotes

  1. OWASP GenAI Security Project, OWASP GenAI LLM Top 10 — 2026 Edition (published 2026-08-03) and 2026-09-01 announcement (“Excessive Agency, now number three”). https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/ · https://genai.owasp.org/2026/09/01/owasp-genai-security-project-unveils-2026-top-10-for-llm-applications-new-agent-control-standard-and-sponsors-as-community-tops-30000-members/ ↩

  2. CNIL and Conseil de l’IA et du Numérique, IA agentique et données personnelles : note exploratoire (increased decision-making autonomy, diluted responsibilities, cyber risks extended to every connected service), 2026-07-20. https://www.cnil.fr/fr/ia-agentique-cnil-cianum-note ↩

  3. MITRE, CVE-2026-102697 (Ollama 0.14.0 → < 0.31.2: shell control operators appended to an approved command in agent mode, CVSS 3.1 7.8 / 4.0 8.5), published 2026-09-29. https://cveawg.mitre.org/api/cve/CVE-2026-102697 ↩

  4. OpenHands GitHub README (repository presented as “Agent Canvas (beta)”, MIT), accessed 2026-10-10; OpenHands, Introducing Agent Canvas (2026-06-16) and Use any coding agent in OpenHands with ACP (2026-06-18). https://github.com/OpenHands/OpenHands · https://www.openhands.dev/blog/introducing-agent-canvas · https://www.openhands.dev/blog/use-any-coding-agent-in-openhands-with-acp ↩

  5. CISA, Known Exploited Vulnerabilities Catalog (CVE-2026-42208, CVE-2026-42271, CVE-2026-59822 LiteLLM), catalog dated 2026-10-08; LiteLLM, Version Support Policy (four minor lines maintained since 2026-06-29), 2026-06-20. https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json · https://docs.litellm.ai/blog/version-support ↩

  6. Cursor, Cursor is now a part of SpaceX, 2026-08-14. https://cursor.com/blog/joining-spacex ↩

Give my opinion

Everything is optional: one answer is enough. This site sends nothing: you send it, the way you choose.

Your noteNo note
Your opinion is about
About what? (several choices possible)
What will be sent
Send by e-mail

Private, no account: opens your mail program with the message ready, for ia-on-prem@damien.becherini.fr.

Open a GitHub issue ↗

Public, with a GitHub account. You read the issue again before you publish it.

To paste it wherever you want.

Give at least a note, a subject or a message to be able to send.