kikimimi/ kkmm

VISIBILITY FOR TEAMS BUILDING WITH AI

Your team’s agents.
Usage. Cost. Risk.

See who’s using which agents, models, and tools. Find what drives cost, and surface risky detours in how the work gets done.

Explore team setup ↗Try it on your Mac ↓

Start locally. Share metadata with your team when you choose.

YOUR TEAM’S ACTIVITY, IN PERSPECTIVE.Illustration · Sample data
WORKSPACE / TEAM

Team overview

Last 7 days ▾
Sessions128Claude Code + Codex CLI
Total tokens2.4MInput & output
Recorded cost$18.62Reported API usage cost

Token activity

Claude Code Codex CLI
MonTueWedThuFriSatSun
Shared team activity. Metadata only.Team workspace ↗
WORKS WITH YOUR WORKFLOWClaude Code+Codex CLIThe agents you already use.

01 / BUILT FOR TEAMS

Know who uses what. Understand the cost.

Give engineering leads a shared view of adoption, usage, and opportunities to reduce waste.

See who’s using what.

Review member usage, then drill into sessions to see the agents, models, and tools involved. Understand how AI coding is used across your team.

Explore team visibility ↗

Find the cost drivers.

Compare token usage and reported costs. Investigate retry loops, model choices, and unused MCP overhead to decide where to optimize.

Find optimization opportunities ↗

Share with clear boundaries.

Choose which repos sync to your team. Per-member usage and cross-member session drilldowns are restricted to admins and owners, with access recorded in an audit log.

See team access controls ↗

Start with the Members view, then investigate the sessions behind the numbers. Cost visibility depends on what each agent reports; missing cost stays unknown.

Open your team workspace ↗

02 / WHEN AGENTS FIND ANOTHER WAY

Spot warning signs
in how work gets done.

Agents can keep going after a tool fails or a permission is denied. kikimimi surfaces those detours so your team can review what happened along the way.

  • Permission detoursA denied tool followed by a Bash or browser call.
  • Repeated failuresRetry spirals and repeated permission denials.
  • Unexpected overheadContext growth and unusually slow MCP tool calls.
See the detection patterns ↗

ILLUSTRATIVE SEQUENCE · SAME SESSION

  1. 01
    WebFetch deniedA tool call hits a permission boundary.
  2. 02
    Bash called afterwardWithin the next five recorded events.
  3. 03
    Potential detour flaggedReview the session and the tool sequence.

A sequence-based signal: it does not establish intent or prove that a restriction was bypassed. kikimimi observes recorded activity; it does not block actions or detect every risk.

LOCAL FIRST. QUIET BY DESIGN.

Activity, recorded.
Your content, private.

kikimimi records metadata: tool names, durations, and token counts. It does not collect prompts, file contents, or command lines.

Data lands in local Parquet files first. Cloud sharing is opt-in. No proxy or TLS interception required.

Read about privacy ↗

03 / START LISTENING

Meet your Mac’s little listener.

An app or a terminal. Start the way you like to work.

Development preview

kikimimi for macOS

Manage collection from your menu bar and open your local dashboard in one app. CLI and DuckDB included.

Check GitHub Releases for signed macOS installers.

View GitHub Releases ↗macOS 11 or later · Apple Silicon / Intel
Move the app to Applications before enabling collection.
CLI

At home in your terminal.

Install with Homebrew. Use kkmm as a shorter name for the same CLI.

# Install
brew install isamisushi/tap/kikimimi
brew install duckdb

# Start listening
kkmm init
kkmm web
Installation guide ↗DuckDB is required for CLI queries and the local dashboard.

Give your team visibility as agents do more.

Follow along on GitHub ↗