What is stored where, what leaves¶
The analyser is a single-user desktop tool. It keeps its own files under your home folder, writes into a project only what a project holds, and sends nothing anywhere unless you choose an assistant or press a button that fetches.
On this machine¶
| Where | What | Written when |
|---|---|---|
~/.fluxtion-analyser/config |
Your machine settings, in cleartext: the in-app assistant's provider, model and API key; Claude Code's switches, mode, model, spend limit and the folders you confirmed; External MCP/REST access, file exchange, project creation and the exchange folder; theme, window, recent files and search history; your no-project settings; the bins of recently deleted reports, walks and charts (with the definitions a later call replaced), from which they can be restored | whenever a setting changes |
~/.fluxtion-analyser/sessions/ |
Session-recovery snapshots, one file per project: the paths of what was open and their SHA-256, and the view, including filters and flags | when you close a project or quit |
~/.fluxtion-analyser/rest-endpoint |
The listener's URL, its per-run token, the process id and start time. Mode 600 | while External MCP/REST access is on; removed when the window closes |
~/.fluxtion-analyser/demo/ |
The DEMO log, graph and source, unpacked from the jar, and the sample project's profile | the first time you open the sample project, the tour or the demo |
~/.fluxtion-analyser/bundles/ |
Disposable working copies of the evidence bundles you open | when you open a bundle |
~/.fluxtion-analyser/plugins/ |
Log-reader plugin jars, which you put there yourself | never by the analyser |
~/.fluxtion-analyser/local-agent/ |
A lock per Claude Code session, so two analysers cannot run one session | while a Claude Code session runs |
| your temporary folder | A copy of a log opened from S3 (fluxtion-s3-*.log), deleted when the analyser exits; and a private, empty scratch folder for Claude Code when no project folder applies (no project, the sample project, a bundle) |
when Claude Code starts there |
~/.fluxtion/fluxtion.apiKeyFile |
The Fluxtion build key, for regenerating a processor; the analyser writes it and reports only whether it is present | when you use AI ▸ Fluxtion API key… |
your OS credential store (macOS Keychain, or the Secret Service keyring through secret-tool); otherwise ~/.fluxtion-analyser/compiler-account.credential, owner-only |
The compiler-account device credential, with the address of the service that issued it. Never on a command line, in the config file, a project, an export, a report or a bundle | when you connect with AI ▸ Compiler account…; deleted by Disconnect or Revoke this device |
<project>/.analyser/project.fluxtion-settings |
The project's settings and context: source roots, processors, saved charts, reports, walks, runbook and glossary pointers, environments, analyses, report destinations (Projects, profiles and sharing). No key, ever | as you edit the project |
<project>/CLAUDE.md |
Links to the Fluxtion authoring docs, only if you tick that option when creating a project, and never over an existing file | once, on creation |
| your MCP client's own configuration | The registration Set up AI's Register for me… (or the Connect an AI client dialog) adds, by running claude mcp add or codex mcp add after you confirm; for any other client, nothing (you copy or save the record yourself) |
when you register a client |
| the exchange folder you choose | Screenshots, report PDFs and CSVs, reels and evidence bundles an assistant writes, never over an existing file | only with Allow assistant file exchange on |
Claude Code keeps its own conversation history and settings in its own folders, as it does in a terminal; the analyser never reads or stores your Claude Code credentials. The analyser keeps a conversation in memory, up to its last 500 entries, and saves turns only when you add them to a walk.
Protect the config file. It holds the in-app assistant's API key in cleartext, readable by anything running as you. A shared settings file, a project profile and an evidence bundle never contain it.
What leaves the machine¶
Nothing the analyser does on its own sends your evidence anywhere. What does, and when:
| What | Sends | To | When |
|---|---|---|---|
| Safe analysis or Full Claude Code | what you ask, and the evidence Claude reads through the analyser | Anthropic, on your Claude plan or API key, as in any Claude Code session | Start contacts Anthropic (the sign-in check and the bridge connection) before your first Send, with no record content; each Send then sends your turn |
| Full Claude Code's own tools and MCP servers | whatever they send | wherever your tools and servers connect | as your permission rules allow |
| In-app assistant | what you ask, the selected records' context and the evidence it reads | the provider you chose (or your base URL), with your API key | each Send |
| Off | nothing from the analyser | — | — |
| AI ▸ Compiler account… | a sign-in request naming this device ("Fluxtion analyser on" and your machine's name); then this device's credential, to read the account, usage and device list, or revoke a device you pick. No evidence | the compiler service address you entered | when you connect, open the dialog while connected, press Refresh, Revoke selected or Revoke this device |
| An outside client you connected | whatever it reads through the analyser | wherever that client sends it | as you use that client; Off does not affect it |
| Open log from S3… | your request for that object | your S3, through your local aws tool and its credentials |
when you open it |
| Starter templates | a request for the catalogue, then for one template's archive | the Fluxtion playground, over HTTPS | only when you press Show templates, Start building, or use New project from template… |
| An example entry | a request for one archive | GitHub or GitLab, over HTTPS | only when you press Download example… and confirm |
| Help and docs links | the page address | your web browser opens the docs site | when you click one |
No assistant profile is local-only, Safe analysis included: Safe is about what Claude Code may do on your machine, not where your words go (Security model ▸ Not local-only).
Does it work offline? Yes, everything except the assistant and the buttons above. With the assistant Off and External MCP/REST access off, nothing in the analyser goes on the network unless you press one of those buttons.
What you send yourself¶
- A report's PDF or CSV contains the evidence it shows, including record values and any narrative.
- An evidence bundle contains the whole log (or an excerpt), the graph, your saved charts, reports and walks, and your notes (Evidence bundles). It leaves out your source roots, keys and assistant settings.
- A reel contains pictures of your screen at each step.
- A shared settings file contains the categories you tick (Sharing settings with a file).
A log from a live system names real hosts, services and accounts. Treat anything made from it as a production artefact.