How-to: your first run (zero to wired in minutes)¶

mokata is the memory + seatbelt for your AI coding agent. This is the fastest path from nothing to a wired, personalized setup — and a 60-second demo if you'd rather look before you leap.

Already have mokata installed? If not, start with Getting started (pip install mokata → mokata setup claude), then come back here.

See it work first — `mokata tour` (read-only)¶

mokata tour            # or /tour inside Claude Code

A 60-second, read-only walk through three things on a tiny sample — it writes nothing to your repo:

Graph query — ask the codebase a structural question (mokata query callers <symbol>) instead of grepping.
Memory recall — mokata remembers your project's decisions so the agent stops re-asking.
Gate catch — every durable write is scanned; a secret is a hard block approval can't override. Nothing is committed.

The guided first run — the wizard¶

Run mokata init interactively (or /setup inside Claude Code) and you get a guided Q&A wizard instead of flag-wrangling:

Pick a profile — minimal (engine only), standard (engine + graph + memory on lean local defaults), or full (every known provider).
mokata detects your integrations — graph backends, memory backends, Postgres / vector — and shows you exactly what's installed.
You choose what to wire. For something detected-but-not-installed, mokata recommends the install command (e.g. pip install 'mokata[postgres]') — it never installs a third-party tool for you. Detect → recommend → run with your approval.
It wires what you approved — scaffolds the config, wires the chosen integrations, and (optionally) wires mokata into your harness (slash commands + MCP server + hooks + status badge) — every durable step human-gated. Decline and nothing is written.
A 30-second recap — "here's what I just did": what was detected, what got wired, the graph and memory now standing, the 5 starter guardrails (your constitution), and the one next step.

Everything is local-first and reviewable — the config is committed plain JSON, and every write went through a gate you approved.

Non-interactive (CI / scripts)¶

The flag path never prompts:

mokata init --profile standard --yes      # scaffold, no wizard, no prompts
mokata init --mode seatbelt --yes         # or name an on-ramp instead of a profile
mokata setup claude --yes                 # wire the harness non-interactively

🔴 Changed in 0.0.19 — mokata init --yes now wires the harness too. --yes is read as consent to the whole init plan, and wiring the run-state gate is part of that plan, so a non-interactive init also runs setup claude at project scope: .claude/commands/, .claude/skills/, .claude/settings.json (hooks, the MCP permission grant, the status line) and .mcp.json, plus a short-lived mokata-mcp subprocess that checks the server answers. Existing JSON is merged, never overwritten, and a wiring failure never fails an init that already succeeded. If your CI times, sandboxes or diffs mokata init, expect those writes and that subprocess — mokata unsetup claude --scope project reverses them.

--mode {seatbelt,memory,full} is the graduated on-ramp (mutually exclusive with --profile): seatbelt = the gates + the AST code graph they need, memory = that plus typed persistent memory, full = everything the spine can wire. It resolves to a profile and only the profile is persisted. Under --yes the interactive extras offers never fire, so a CI init can never reach pip. Full detail: Getting started.

When you mistype a command¶

mokata helps instead of just erroring:

$ mokata statuss
mokata: 'statuss' is not a mokata command.
Did you mean 'status'?  (try `mokata status --help`)
Next: run `mokata init` (or `/mokata:setup` inside Claude Code) to set up this repo …

It suggests the closest real command (a difflib match over the command set) and the single most useful next step for where you are.

Change your setup later — `mokata reconfigure`¶

You're never locked into your first-run choices. Re-run the same guided Q&A any time on an already-set-up repo to change what's wired — it re-detects your tools, shows a current→proposed diff, and applies behind one gate:

mokata reconfigure                                   # interactive — or /reconfigure
mokata reconfigure --add postgres --yes              # wire a now-installed integration
mokata reconfigure --remove serena --yes             # cleanly unwire one (no residue)
mokata reconfigure --profile full --yes              # switch the profile
mokata reconfigure --set tools.sqlite.config.path=mem/custom.db --yes   # switch a backend

It's idempotent (no changes → a no-op, nothing written), human-gated (decline → nothing changes), and reversible (--remove leaves no residue — gone from the capability chain and the tools table). Like first-run, it detects → recommends → runs with approval — an absent --add tool is recommended, never installed. Integrations grow with your project instead of requiring a manual teardown.

Your brainstorm design, saved as a plan file¶

When you approve a brainstorm approach — before any spec — mokata saves the design write-up as an internal plan file under .mokata/temp_local/plans/<slug>.md, so the reasoning behind the approach isn't lost. Browse and share it with mokata plan:

mokata plan list                       # the saved plans
mokata plan show [<slug>]              # print one (defaults to the sole plan)
mokata plan export [<slug>]           # copy it to a committable plans/<slug>.md
mokata plan export <slug> --to docs   # choose the destination dir

export is user-initiated and never silently overwrites an existing copy — pass --force to replace one. (Full flags: the CLI reference.)

Next¶

Once you're wired, start your first governed change with /brainstorm, or read the pipeline & gates. To change what's wired later, use mokata reconfigure (above); to remove mokata entirely, mokata unsetup / mokata reset.