Temet

Setup is the first thing you run after installing mike. It checks your machine, asks where mike should keep history, connects Claude Code to mike and lets you choose which optional features are on. It changes nothing without showing the exact command first, and it never runs sudo: a command that needs administrator rights is printed for you to run.

Run it in a terminal:

mike setup

Setup needs a terminal. Without one it stops with a message that points you to mike doctor. The same checks and the commands that fix them are also in the browser at http://localhost:9990/settings/setup, where you can change the settings a browser can and copy the command for the rest.

The steps

Setup shows one screen per step. Each lists what it found, then the choices that fit this machine, with the recommended one first. ⏎ runs the highlighted choice, and f leaves setup and keeps the answers so far. Run mike setup <step> to reopen it at one step: database, service, claude, features, models, embeddings or terminal.

Step	What it settles
Welcome and machine	Whether setup should list the tools it can use on this machine. It also repeats the warning that mike stores every prompt and answer in full, on this machine.
Database	Where history is kept.
Background service	Whether mike keeps running after you log out.
Claude Code	How Claude Code reaches mike, and what is registered in it.
Features	Which optional features are on.
Background models	Which model service mike’s own workers use.
Search by meaning	Whether recall can match by meaning as well as by text.
Terminal	Which terminal tool `mike spawn` and fleets use.
Ready	A summary, and shortcuts to start Claude Code, open the web UI or take the tour.

You can skip most of these on the first run. The database and Claude Code steps are the ones that decide whether anything is captured.

Choose where history is kept

mike stores captured sessions in PostgreSQL 19 with the vector extension. The Database step offers what fits your machine:

Choice	Pick it when	What happens
Run PostgreSQL in Docker	Docker is installed and you want history that lasts, with mike managing the database.	Starts a database in a container on your own machine only, keeps the data in a Docker volume and generates a password into mike’s config file.
Install PostgreSQL 19 and pgvector with Homebrew	You are on a Mac with Homebrew.	Runs the Homebrew commands it shows, then creates the database.
Install PostgreSQL with your package manager	Neither of the above fits.	Prints the commands for this machine and runs none of them. Run them, then choose the next option with the URL they end with.
Use a PostgreSQL I already run	You have a server.	Checks the URL you type and saves it only if the version, the extensions and the migrations pass.
Keep captures in memory only	You want a first look.	Nothing is stored between runs, and the web UI shows only what happened since mike started.

If a database is already set, the first choice is Keep the configured database, which re-checks it.

To make the database choice without the screens:

mike setup --mode docker --accept-trial-loss

--mode takes trial, docker or existing. With existing, pass --database-url-file with the path of a file only you can read; mike never takes a password as an argument. --port sets the database port (default 6666). --accept-trial-loss is required every time, because switching storage can restart mike and lose captures that were only in memory.

To manage the Docker database yourself, mike database check validates the prerequisites without changing anything, mike database provision creates the container, and mike database stop stops it and keeps the data.

Keep mike running

The Background service step installs mike as a login service, so the proxy and its forwarding sidecar start at login and restart after a crash. On Linux it writes two systemd user units; on macOS, two launchd agents. It refuses if something already answers on mike’s ports.

mike service install

mike service status reads the state of both, and mike service restart restarts them. mike service install --dry-run shows what it would write. If you skip the service, mike claude starts mike by itself when it is down, and nothing keeps running after you log out.

Connect Claude Code

The Claude Code step makes four choices, each skippable:

Register mike-mcp. Lets a Claude Code session query what mike captured. mike mcp register does the same; mike mcp probe checks it and mike mcp unregister removes it. See MCP.
Route. Through the forwarding sidecar (recommended) sends Claude Code through the sidecar, so mike also captures the traffic of Claude’s Remote Control. It needs a certificate authority file that mike forward creates the first time it starts. Straight to the proxy needs no certificate and captures model traffic only. Change it later with mike claude-code route sidecar or mike claude-code route plain. mike claude and the shell function both follow this choice.
Folder trust. Whether mike claude marks the folder as trusted before it starts. Off means Claude Code asks about each new folder itself. mike claude-code trust on and trust off change it.
Mods and a shell function. The mods add panes and buttons inside Claude Code. The shell function, cc by default, starts Claude Code through mike from any folder. Setup writes it into your shell’s startup file and replaces it in place if you run the step again.

Whichever you choose, mike claude starts Claude Code wired to mike.

Choose features

The Features step offers three presets, and mike features <preset> applies one from the command line:

Preset	What it turns on
`watch-only`	Nothing that rewrites a request or sends captured text to another service. Every other switch keeps its default.
`recommended`	Watch only, plus recall, lessons and reminders. They add notes to what Claude Code sends to Anthropic; nothing leaves your machine for them. See Memory.
`everything`	Every switch on the settings page, including those that send captured text to OpenRouter and those that rewrite requests and responses. The ones that call a model do nothing until a provider is set.

Choose watch-only to keep everything optional off. The step shows which switches a preset would turn off. If mike is not running, the choice is saved and applied the next time it starts. Settings lists every switch with its default, and Privacy says what each one sends and where.

Models and search by meaning

Background models are what mike’s own workers use for judging and for finding lessons. Choose “Use” plus a provider’s name “for every job” for a provider the step found on your machine, set up an OpenRouter key, or leave the models alone. Features that need a model stay idle until one is set. Skip this step unless you turned those features on; the judge and steering say which need one.

Search by meaning lets recall match a request to lessons by meaning, not only by words. The step can pull a small embedding model into Ollama on your machine, point at another embedding server, or leave it off, in which case recall uses text matching. Skip it on a first run.

Check the result

mike doctor

mike doctor lists each check as ok, warn or FAIL, and each line that is not ok ends with the setup step that fixes it, for example fix: mike setup claude. It exits with an error when a required check fails. mike doctor --json and mike doctor --support-bundle print a JSON summary of the installation instead, with values left out; use the support bundle to preview what you would attach to a bug report. Neither changes anything.

When a step fails

What you see	What to do
`FAIL database url`: persistence is on but no database URL is set	Run `mike setup database`, or choose Keep captures in memory only.
`FAIL inspect port`: nothing answers	mike is not running. Run `mike claude`, or install the service with `mike setup service`.
`warn mike-mcp in claude`: not registered	Run `mike setup claude`, or `mike mcp register`.
`warn claude route`: no route is chosen	Run `mike setup claude` and choose one.
`warn terminal backend`: none chosen	`mike spawn` and fleets are unavailable until you run `mike setup terminal`.
Setup says it needs a terminal	Run it in a terminal window, or use `mike doctor` and the commands it prints.
The Docker choice is missing	Docker is not installed or not running. Pick another database choice.

Troubleshooting covers the rest, and Capture your first session is the next step.