Temet

mike has two kinds of settings. Boot settings decide where mike listens and which database it uses. They come from a configuration file, environment variables and command-line flags, and they take effect when mike starts. Everything else, including every feature switch, is changed in the web UI at http://localhost:9990/settings and stored in the database. This page covers both, and says what a fresh install has on.

Open http://localhost:9990/settings, or run mike features and mike print-config from a terminal.

The settings page with the capability panel and the session features

This is a fictional example: the docs session in ~/projects/site is being captured, and nothing here is a real person’s setting.

What a fresh install does

A fresh install captures sessions and forwards your requests and Anthropic’s answers as they are. Most features that rewrite a request, change an answer or send captured text to another service are off. A few background jobs that only work on your own machine are on; they are listed below.

A setting you have saved is never replaced by a newer default. After an upgrade, open the settings page if you want to be sure what is on.

Turn features on or off

The example: Claude Code kept putting the Linux install first on the docs page, and you corrected it with “Put Homebrew first on the install page.” You want mike to keep that correction as a lesson and offer it to later sessions. That needs the lesson switches on.

Open http://localhost:9990/settings.
Read the capability panel at the top. It says which features are working, which are waiting for something such as a key or a login, and links to the page where you fix it.
In Session features, switch on recall_injector, agent_lessons and reminders. Each is marked Applies to new sessions.
Start a new session. The docs session you already have keeps the behaviour it started with.

You should now see a <system-reminder> note with your Homebrew correction in what Claude Code sends to Anthropic on the next session’s requests. Memory shows how to read it back.

Or do the same from a terminal with mike features recommended, described next.

The three starting points

mike features <choice> sets every feature switch at once, reads them back and reports any that did not change. The setup step called features starts on recommended.

Choice	What it sets
`watch-only`	Every switch at its default.
`recommended`	Watch only, plus `recall_injector`, `agent_lessons` and `reminders`, and the setting `reminders.inject_lessons`. All four add a `<system-reminder>` note to what Claude Code sends to Anthropic. None sends text anywhere else.
`everything`	Every switch on, plus `reminders.inject_lessons`. That includes switches that rewrite requests and ones that call a model. The ones that call a model do nothing until you have set up a provider (`mike setup models`).

The switches live in the database, so the command needs one. When mike is not running, the choice is saved in the configuration file as setup_features and applied the next time mike starts.

The settings page

The page has:

A capability panel at the top.
Session features. The switches that change what a session sends or receives. A session already running keeps the behaviour it started with.
Global settings and the settings of each other page.
Every setting, a searchable list. Each row says when a change applies. Type in Find a setting to narrow the page.

Page	Path	What it holds
Setup	`/settings/setup`	The setup steps in the browser. Steps that need your terminal show a Run this step in a terminal button, because the browser does not run them.
Steering	`/settings/steering`	The steerers that watch a session for drift.
Workers	`/settings/workers`	The background jobs, such as the turn and session judges, and their evaluations.
Assistants	`/settings/assistants`	The assistants mike can start, with their sign-in state.
Embedding	`/settings/embedding`	Search by meaning: its on switch, address and model.
Flagging	`/settings/flagging`	Flagging of weak answers, and the forbidden-word intercept.
Integrations	`/settings/integrations`	Connections to outside services and what they have cost.
Configuration history	`/settings/history`	Every change to a setting, filterable by kind and by where it came from.

The setup page in the browser, with steps that open in a terminal

Setup walks through the setup steps. Memory, Steering, Judge, Filters and Compression explain the features behind these pages.

What is on and off by default

Defaults below are the ones in this build. “Rewrites” means the request mike forwards, or the answer you receive, differs from what the other side sent. The original request is still captured as Claude Code sent it.

Off by default: rewrites what mike forwards to Anthropic

Switch	What it does when on
`reminders`	Adds a note built from the weak-answer findings on the previous request.
`recall_injector`	Adds a note of lessons that match the topic.
`agent_lessons`	Adds lessons for a helper session started inside another.
`self_identity_reminder`	Adds a note saying which session this is.
`mailbox_redelivery`	Delivers messages between sessions again when they were delivered but not read.
`proxy_tools`	Gives a session mike’s own tools and answers their calls itself.
`repeated_notices`	Removes repeats of Claude Code’s standard notices from the copy sent on.
`file_read_compression`, `shell_output_compression`	Shorten older file reads and command output in a conversation’s history. Each starts in `shadow` mode, which only records what it would cut. See Compression.
`reminder_on_token`	Replaces a marker some hooks add with a note.

Off by default: changes what you receive

Switch	What it does when on
`weasel_intercept`	Watches a streamed answer for words you list. When one completes, mike cuts the answer before it, asks again with a correction, and continues the same message. It does nothing until you add at least one word. See Flagged responses.

Off by default: sends captured text to another service

All of these need a model provider, such as an OpenRouter key, except where noted. Privacy says what each sends.

Switch	What it does when on
`steerer`	Runs steerers that judge drift and queue corrections for the next turn. See Steering.
`judge_worker`	Judges each turn. See Judge.
`session_judge`	Judges whole sessions.
`judge_lessons`, `session_end_shadow`	Further judging jobs. The settings page describes each; treat them as sending captured text to your provider when on.
`turn_guard`	Checks a turn’s closing text against a few questions.
`steerer.correction.enabled`	Decides whether you were correcting the model.
`lesson_relations.enabled`	Compares stored lessons.
`embedding.enabled`	Embeds turns and lessons for search by meaning. Needs no key. Text goes to the address in `embedding.url`, `http://127.0.0.1:7997` unless you change it.

Off by default: other

These are off, and none of the first four sends text to another service:

Switch	What it does when on
`dossier_builder`	Builds a record of each closed turn from captured data: how it ended, checks run, session class. No model call.
`directive_proposer`	Proposes standing directives from corrections you restated across sessions. Proposals wait on Memory; nothing is injected. No model call.
`change_record`	Keeps a record of what happened in each checkout a session works in, so fleets can ask what others changed without git. Nothing leaves the machine.
`change_record_tail`	Adds a few lines saying what others changed to some answers inside a fleet. Needs `change_record`.

workspace_briefing, context_focus_question, launch_brief, native_peer_mail, mod_lesson_delivery and the client_serves_* switches other than client_serves_self_identity are also off. They relate to fleets, messages between sessions and mods; the settings page describes each.

On by default

Switch	What it does
`lesson_worker`	Tidies stored lessons. Local.
`harvester`	Saves lessons that a session writes into its own answers as `<mem:update>` tags. No model call.
`retention_sweep`	Deletes sessions older than `default_ttl`, 365 days by default.
`channel_link_sweep`	Keeps Remote Control channels linked to their sessions.
`workspace_capture_registration`	Lists captured sessions in the roster used for messages between sessions.
`db_resync`	Copies captures written to a backup database back to the main one. Only matters if you set up a backup database, which a normal install does not have.
`semantic_context`	Uses search by meaning, only while the embedding address is on your machine.
`fleet_auto_start`	Starts a fleet when its definition is ready. Off means it waits for `mike fleet start`.
`prompt_assistant`	The prompt assistant; its sign-in and limits are on `/settings/assistants`.
`client_serves_self_identity`	When a client such as a Claude Code mod supplies a session’s identity note, mike stands aside for that session.

A switch can be on in this list and still do nothing. semantic_context needs an embedding service, and embedding.enabled is off in the boot defaults. If the settings page shows embedding.enabled or lesson_relations.enabled as on in a fresh install, that is the page’s own default and the capability panel says whether it works.

Boot settings and environment variables

From lowest to highest, a boot setting comes from:

Built-in defaults.
The configuration file.
MIKE_* environment variables.
Command-line flags.

An exported variable in your shell therefore beats a value saved in the file, and a flag beats both. mike print-config shows what mike would start with, which variables it found, and then, after a line reading LIVE OVERLAY, the settings you changed in the web UI. The second part is what a running mike uses.

Settings that change where mike listens, which database it uses, or where it writes files take effect when mike restarts. Settings in the web UI say when they apply.

mike does not read a .env file from the current folder. To bring one in, use mike config import-env or mike import-dotenv; see Run mike from the command line.

Where the files are

mike print-config --paths prints the real folders for your installation as JSON. In general:

What	Where
Configuration file	`~/.config/mike/config.toml`, or under `$XDG_CONFIG_HOME` when that is an absolute path. `mike config path` prints it.
State: machine identity, database files, `service-token.<port>`, logs	`~/.local/state/mike`, or under `$XDG_STATE_HOME`. `mike config path --state` prints it.
Data folder	`~/.local/share/mike` on Linux; `~/Library/Application Support/mike` on macOS. `MIKE_DATA_DIR` moves it.

When you pass --config with another file, mike keeps that installation’s data, state and logs in a folder named .mike-<id> beside the file. Keep these folders private to your account: mike refuses a configuration folder that other users can enter.

The configuration file

The file is TOML. Each key is the name of a MIKE_* variable without the prefix, in lower case: database_url for MIKE_DATABASE_URL. Values are read literally, so a $ in a password stays a $. Setup and mike claude-code also keep a few notes of their own in the file, such as setup, setup_features, claude_route and launch_auto_trust.

To keep a key out of the file, write <name>_from_env = "VARIABLE": mike then reads the key from that environment variable. For the OpenRouter key, steerer_openrouter_api_key_from_env = "OPENROUTER_API_KEY" does this, and setup stores only the variable’s name when you type one.

Environment variables you may need

Only MIKE_ names are read. A variable still named PROXY_... is from an older version and is ignored.

Variable	Flag	Effect and default
`MIKE_DATABASE_URL`	`--database-url`	The PostgreSQL address. Setup writes it for you.
`MIKE_PERSIST_ENABLED`	`--persist-enabled`	`0` runs without a database; captures are kept in memory and lost on restart. Default on.
`MIKE_LISTEN_ADDR`	`--proxy-listen-addr`	Where the capture listens. Default `127.0.0.1:9988`.
`MIKE_INSPECT_ADDR`	`--inspect-listen-addr`	Where the local API listens. Default `127.0.0.1:9989`.
`MIKE_SPA_BIND`	`--spa-bind`	Where the web UI listens. Default `127.0.0.1:9990`.
`MIKE_SPA_ENABLED`	`--spa-enabled`	`0` turns the web UI listener off. The API stays on port 9989, so this does not reduce what a local program can reach.
`MIKE_UPSTREAM_URL`	`--upstream-url`	Where requests are forwarded. Default `https://api.anthropic.com`.
`MIKE_DATA_DIR`	`--data-dir`	Moves the data folder.
`MIKE_DEFAULT_TTL`	`--default-ttl`	How long a session is kept before the retention sweep removes it. Default `365d`.
`MIKE_MAX_REQUEST_BODY_BYTES`, `MIKE_MAX_RESPONSE_BODY_BYTES`	same names with `--` and dashes	Largest request (16 MiB) and response (32 MiB) mike will carry.
`MIKE_LOG_FORMAT`	`--log-format`	Log layout. Default `pretty`.
	`--log-level`	Log detail. Default `info`. Also the configuration key `log_level`.
`MIKE_LAUNCH_BACKEND`	`--launch-backend`	The terminal program `mike spawn` opens sessions in. Default `none`, which makes `mike spawn open` exit 69.
`MIKE_STEERER_OPENROUTER_API_KEY`		An OpenRouter key for the features that call a model. `OPENROUTER_API_KEY` works as a fallback. mike never writes the key to the database.
`MIKE_INSPECT_ALLOWED_HOSTS`		Extra host names the local API accepts. Leave it empty; adding a name does not add a login.
`MIKE_INSPECT_URL`		For `mike-mcp`: where to find the local API. Default `http://127.0.0.1:9989`.
`MIKE_MCP_PROFILE`		For `mike-mcp`: `core` (default) or `full`; see Run mike from the command line.

mike --help lists every other flag, including the --launch-* and --mailbox-* ones.

mike refuses to start with a listener that is not on a loopback address. Do not work around this; Privacy explains why.

When something is off

You see	Why	Do this
A feature is on but nothing happens	It needs a provider key or an embedding service	Read the capability panel; run `mike setup models` or `mike setup embeddings`.
A switch you turned on is not in effect	It says Applies to new sessions	Start a new session.
`mike features` says it saved the choice but applied nothing	mike is not running	Start mike; the choice is applied then.
A variable you exported has no effect	A flag overrides it, or it starts with `PROXY_`	Check `mike print-config` for the source of each `MIKE_` variable.
mike will not start and names the configuration folder	Other users can enter it	Make the folder private to your account.
Your settings look different after an upgrade	Saved settings are kept; new defaults apply only to settings you never saved	Open the settings page and check.