Assistant profiles¶

A profile is how the assistant works on this machine: Safe analysis, Full Claude Code, In-app assistant or Off. You choose one in Set up AI…, or switch with the mode chip in the assistant panel's header (Safe ▾, ◆ Full ▾, In-app ▾, Off ▾), and it is stored for this user on this machine only. While nothing is set up the chip reads Set up AI….

Set up AI¶

Set up AI… is the one place to choose how AI works with the analyser. It is the last item of the mode chip's quick switch, first in the AI menu (which opens it at Your own AI client when one is set up), at the top of Settings ▸ Assistant, and on the start page's AI line. On a first run the start page asks the same question inline, as part of the page: How do you want AI to work with the analyser? It is offered once — to a new installation and to one upgraded with nothing set up alike; after that the page shows the one AI line, which keeps the same door.

No AI: everything works without AI — records, topology, charts, reports, walks and the tour.
Your own AI client connects: Claude Code, Codex or another AI app that can connect to the analyser, in your terminal, drives the analyser (see Your own client, step by step below).
Claude inside the analyser: the same Claude Code, run by the analyser in this panel.
Both: Claude in the panel, and your terminal client too.

For Claude inside, the window first checks this machine by running claude --version and claude auth status --json. No question and no evidence is sent; the checks run in an empty private folder, and no e-mail or organisation from the answer is kept. Each check says what failed and has Check again. An API key is reported as uses an API key and never as signed in: Claude Code checks a key only when it starts. You then choose Safe analysis (recommended) or Full Claude Code. Full asks for the same confirmation as the Settings switch, with Cancel as the default. The in-app assistant with your own API key is behind Other options on the first screen; it is no longer one of the first choices.

Your own client, step by step¶

For Your own AI client (and the second half of Both, after Claude inside) one screen takes four steps, each saying what it checked:

Which client? Claude Code, Codex, or another MCP client (Claude Desktop included: copy the configuration).
Let your client reach this window — the External MCP/REST access switch. Turning it on asks first, with Cancel as the default: only programs on this computer that have this run's key can connect, the endpoint listens on 127.0.0.1 only, and a Full Claude Code session in the panel can use it too. The AI menu's checkbox, Settings ▸ Assistant and the Connect an AI client dialog ask the same question; turning access off never asks. Beside the switch, the Where data goes picture shows the path it opens: your client's arrow is solid only while access is on. Check connection, under the switch, starts the bridge your client will start and asks it what is open here: Bridge works now proves this window can be reached through it (not that your client has registered it). With access off, it asks the same question first, and confirming turns access on and runs the check in one step.
Register the analyser with your client. The exact command is shown. Register for me… runs it only after a confirmation that shows the full command (Cancel the default); I'll run it myself copies it; Check registration asks your client. Nothing runs before you press one of them, and the analyser never reads or edits your client's settings file.
Can your client see the analyser? Start a new session of your client and ask it Use analyser_context and tell me what is open. When a call with this window's key arrives, the step says when, and the name the client gives itself — its own claim, shown as such. If another analyser window holds the connection, or access is off, it says so.

Finish is enabled once access is on (the last check is never required). More options… opens the Connect an AI client dialog for the rest (project-scope command, remove, Claude Desktop's note); it is the only way to that dialog. If you Cancel after this window turned access on, access stays on and the closing line offers Turn it off. Choosing No AI while access is on offers a ticked box, Also turn off External MCP/REST access.

Nothing is sent from the window. If a question is waiting — you asked from a record, the start page or the panel before anything was set up — the last button reads Finish, then send my question: it closes the window and puts the question in the composer of the assistant you chose, so pressing Send is the only step left. Otherwise a short summary suggests one first step. Cancel applies nothing further; a question waiting in the start box stays there. Not now on the start page's block hides it and leaves one line, AI: not set up — Set up AI…. Both the block and the line are remembered on this machine only.

Choose how the assistant works¶

The mode chip in the assistant's header (it names the profile in force, e.g. Safe ▾) is a quick switch — a short menu, not a set of cards:

✓ Safe analysis · ◆ Full Claude Code… · In-app assistant · Off · ─── · Set up AI…

The profile in force is ticked, and said in words to a screen reader (Safe analysis, in force). Each row's tooltip says what it sets up and where your words go. With nothing being answered, a switch that Claude Code must restart for (Safe analysis → Full Claude Code) restarts it at once; while an answer runs, the panel asks Settings changed — restart Claude Code now? with Restart / Later, as for a change in Settings.

Profile	What it sets up
Safe analysis (the default)	Claude Code that reads your evidence; it changes what is open or writes a file only after you approve each call: Read, Grep and Glob in the project folder and the analyser's own views. No shell, no edits, no web, none of your own skills or servers — safe with logs you did not write. Not local-only: what you ask, and the evidence Claude reads through the analyser, is sent to the model provider (Anthropic, on your Claude plan or API key), as in any Claude Code session.
◆ Full Claude Code	Your own Claude Code — your settings, skills, plugins, hooks and MCP servers, and your own tools under your rules, in the mode Settings ▸ Assistant holds (`default`, `acceptEdits`, `auto`, `plan`). Choosing it opens the same confirmation as the Settings switch (the warning, Cancel the default) and nothing changes until you confirm; while it is showing, the menu's rows are disabled. When Claude Code asks for approval, the analyser asks you (Approvals). Like Safe analysis, it sends what you ask, and the evidence Claude reads, to the model provider.
In-app assistant	The analyser's own assistant: a provider, a model and an API key. Not local-only: what you ask, and the evidence the assistant reads through the analyser, is sent to the model provider you chose, on your API key.
Off	No assistant in the analyser. No question is sent anywhere from the analyser; every view, chart, report and walk still works. A stored key is kept, unused. An external MCP or REST client you connected is not affected.

A switch applies at once — on this machine only, never in a project, a shared settings file or a bundle — and puts the tab in that profile's mode. It writes the same settings Settings ▸ Assistant holds; Settings ▸ Assistant still has every option (the models, Full's mode, the spend limit, the provider and key), grouped the same way.
A switch that cannot work as it stands opens Set up AI at the screen that finishes it, and changes nothing: Safe analysis or Full Claude Code while Claude Code was not found on this machine opens Claude inside, which checks for it; In-app assistant with no API key stored opens Other options, at the key field.
With nothing set up, a Send, a Start or an Ask opens Set up AI, every time, and the question waits for it. A question asked on the start page — or from a record, a chart or the topology with Ask the assistant — is placed in the composer of the assistant you set up when you press Finish, ready for Send — you do not ask again, and nothing is sent until you press Send. Cancel and the start page's question stays in the box. A machine that was already set up (Claude Code allowed, or an API key) counts as set up and is never asked.
A change while an answer is in progress is refused, by name — "Claude is still answering — Stop or wait, then change how the assistant works." — in the status bar, as a mode switch is.
External MCP/REST access is not part of the switch. The switch chooses the assistant in the analyser, never the external loop: that is Set up AI's Your own AI client screen, the AI menu's checkbox, or Settings ▸ Assistant ▸ External MCP/REST access.

(Earlier releases had four cards in the panel that opened by themselves on the first question; Set up AI and this switch replaced them.)

Where profiles are stored¶

Set up AI, the quick switch and Settings ▸ Assistant write the user's own ~/.fluxtion-analyser/config, on this machine. A profile is never read from a project, a shared settings file or an evidence bundle, so a cloned repository or a received bundle cannot turn an assistant on. No switch travels: a shared settings file's Assistant limits category carries only the in-app assistant's two action limits, and External MCP/REST access and the in-app assistant's view switch are set on each machine (Projects, profiles and sharing). What an administrator can and cannot enforce today is on Managed policy (planned).

Settings ▸ Assistant groups its switches by profile: Safe analysis, Full Claude Code, In-app assistant, and Every assistant for what all of them share (the bridge, files and projects).

Set up the in-app assistant¶

Set up AI's Other options sets the provider, model and key in one step (the quick switch's In-app assistant opens it there when no key is stored). Or open Settings ▸ LLM (provider, model, key and base URL; Settings ▸ Assistant ▸ In-app assistant holds its action switch and limits):

Provider / model — Anthropic (Claude) or OpenAI, and a model id.
API key — stored locally in cleartext (single-user tool), and sent to the configured provider endpoint to authenticate requests. It is never included in a shared settings file.

For OpenAI, leave Base URL blank to use its standard API endpoint and enter an OpenAI API model ID in Model. OpenAI HTTP 404 can mean the model ID or a custom endpoint is wrong. OpenAI HTTP 429 means a rate, credit or usage limit was reached; check your API account's billing and limits guidance. The analyser does not show the provider's raw error body because it may contain private data.