The deliberate design choices, and the reason for each.

Archie's name, tone and rules are set by Archie's own server. They do not come from the page you are looking at.

Anything Archie reads on screen is treated as reference. It is never treated as an instruction. A web page or a document that says “ignore your rules” is just text on a page.

Archie Vault searches your team's own documents: procedures, manuals, the way your team does things. Every answer that uses one names it, for example “From: Valuation procedure, 4.2”, so you can open the source and check.

The search is local, full-text keyword search (SQLite FTS5, with stemming). It is not a vector database.

Why. Procedure documents use exact terms, and keyword search finds exact terms. It is free, it runs on the PC, and it can always show the source of an answer.

Archie's own MCP server is the only toolset the AI gets. If a tool is not on Archie's list, the AI has no way to use it.

An Admin Console sets who can do what: roles and permissions, team spaces, and approvals.

Shared tasks go through approval. Nobody can approve their own task.

Archie defends against prompt injection by limiting what the AI can do, in several layers.

• Three layers of lockdown. Only Archie's tools are available. Claude Code runs in its restricted mode. If anything else appears, the session is ended.
• A yes before anything that cannot be undone.
• Never passwords or security codes.
• Risk ratings by fixed rules that the AI cannot change.
• Masking. ID numbers, phone numbers, emails, IBANs and secrets are masked before anything leaves the PC.
• A window blocklist for programs Archie must not read.

Why. Limiting what injected text could do is stronger than trying to detect it. It also needs no extra model.

An approved task replays with no AI at all. For repeated work that beats any cache, because there is no model call left to cache.

Answers use a fast model by default. Look deeper asks again with the strongest one.

What happened is recorded in a run log, a journal, an audit trail, a history of changes and cost records.

Archie is tested through a test lab, simulated-user runs, a reliability suite and programs tested live. A mistakes log records every mistake, and every mistake gets a rule.

Privacy. Screen text and prompts are deliberately not kept in readable logs.

• A test for every fix. Over 2,400 automated tests.
• Review before every merge.
• Versioned releases and safe updates.
• A written decision record for every design choice.
• A security review, and privacy checks on every commit.
• Crash reports.