The admin panel that ships with AI In Your Pocket. Add or remove team members, switch AI models, attach client folders the bot can read — all from http://localhost:5680 on the office machine. No terminal. No reinstall. No restart needed.
Most private-AI tools assume the IT person who installs them is the same person who lives in a terminal. That's rarely true in a 5-50 person firm. Office managers, partners, and fractional IT need a web page, not a command prompt.
The admin UI is built around that. Everything that changes over time — who's on the team, which client folders the AI can see, which model it's using — lives on a single web page on the office machine. Bookmark it, get on with the day.
Localhost-bound (127.0.0.1:5680) — never exposed to your LAN, never reachable from outside the office. The security boundary is “can you sit at the machine?” — if you can, you're already trusted.
Paste Telegram numeric IDs from a spreadsheet column (one per line, comma-separated, semicolons — anything sensible). Optional shared label like "Sales team" or "Partner" applies to every new user in one go.
Remove a leaver with one click. The bot stops replying to them within ~1 second — no restart, no re-config.
Every team member has their own private memory file. The panel shows turn count, file size, and last-active timestamp per user — so the admin can spot stale users, oversize histories, or wipe a colleague's memory on their request (GDPR right-to-erasure).
If you're using local Ollama as your AI engine (the privacy-first default), this panel lists every model installed on the machine. Pull a new one by typing its tag — progress bar appears in real time. Swap which model is active with one click; the bot picks it up on the next reply (no restart).
Honest experimentation: start with llama3.2:3b on a modest machine, pull qwen2.5:14b when you upgrade hardware, drop the smaller one when it's not earning its disk space.
Attach a folder on your office machine (or a mapped network drive) and the bot can read its files when your team asks. Read-only — the AI can never modify, move, or delete anything. Files stay on your machine; the AI just gets to see them.
Add a folder by typing an alias ("Acme matters", "FY25 audits") and the path. Behind the scenes a host-side helper service creates a Windows junction inside a mounted parent dir; the folder is searchable + readable from Telegram within ~5 seconds.
When a project finishes or a matter closes, remove the folder with one click. The files themselves are untouched on disk — only the junction unlinks.
The indexer turns every text file across every attached folder into searchable chunks the bot can retrieve. When your team asks a broad question, relevant snippets from across all folders are prepended to the AI's context with file-path citations.
Force a re-walk with one button when you've added a lot of new files. Re-index time is roughly 5 minutes per 1,000 small files (one-off; subsequent walks only re-embed changed files).
Attach the right folders and the AI stops being a generic chatbot. It becomes the colleague who already read every file. Some real examples, by sector:
Folder attached: D:\Matters\Conveyancing. Bot retrieves from Smith-Riverside-Drive subfolder, finds the latest exchange + completion notes, replies with current chain status, outstanding actions, and the next deadline — with file-path citations the fee-earner can click in Explorer.
Folder attached: D:\Clients\Acme. Bot retrieves the Q3 management-accounts.xlsx commentary, the bank reconciliation notes, the email thread about an unusual transaction — produces a partner-ready summary citing each source file. Trainee saves an hour of skim-reading before the meeting.
Folder attached: D:\Blocks\Riverside-Court. Bot finds the latest S20 stage (consultation notice issued 12 Mar, observation period closes 18 Apr), summarises leaseholder objections received, lists the contractor estimates being compared. Property manager opens the cited PDFs in one click.
Folder attached: D:\Roles\Acme-Infrastructure-Engineer. Bot lists candidates in the candidate-submissions.md file with submission date, current stage, last note from the consultant. Saves the recruiter from opening 20 individual CV folders.
Folder attached: D:\Projects\Mill-Conversion. Bot retrieves the BCO correspondence, the structural engineer's report, the planning officer's questions — replies with what's outstanding and what's been answered, citing each document. Architect prepares the response email with one paste.
In every case the folder content stays on your machine. The bot reads chunks at query time and (with local Ollama) does the reasoning entirely on your machine too — nothing leaves your office.
Most “private AI” tools punish the IT person who installed them. Every change — new joiner, new project folder, model upgrade — means SSH, a YAML file, a `docker compose down`, and a restart that sometimes wipes state.
This is the opposite of that. After install, the IT person never needs to touch a terminal again to do anything operational. Web form, click, done. The product gets out of their way and stays out.
That's why managers buy this for IT teams who've already pushed back on three other AI tools. The setup is honest about being technical; everything after install is honestly easy.