Live now: しりとり部, a real-time party game on Cloudflare WorkersLIVE

Tools for people who run
many AI agents at once.

gooji builds the missing layer between developers and their AI coding agents — Claude Code hooks, MCP servers and agent-ready services. Our flagship designs are stress-tested in structured debates between Claude and a second model before any code is written.

See the products How we build

One-person studio · Saitama, Japan · Founded June 2026

miruka echo’s finished screen design (Japanese UI, sample data). The data layer is built; these screens are being implemented now.

676commits since July
across 5 codebases
8,942recorded passing
automated tests & checks
44rounds of Claude-vs-Codex
design review
~2 daysfrom first commit to a
live multiplayer game

Counts from our own repositories, July 11 – Oct 8, 2026. Test suites differ in size and type.

PRODUCTS

Six products. One builder.

Each one is labelled honestly: live, in development, or concept.

miruka echo

IN DEVELOPMENT

One inbox for every AI session that needs you.

When you run several Claude Code and Codex sessions in parallel, you come back and can’t remember what each one was waiting for. miruka echo makes every session file a structured status card at the end of each turn — who is waiting on you, the AI’s exact question, and what it did — so you decide from one screen instead of touring every terminal.

Built on Claude Code: session hooks record each turn, an MCP tool (send_card) lets the agent file its card, and a Stop-hook check sends the turn back once if no card was filed. A local daemon stores everything in SQLite; nothing leaves your machine.

Who it’s for: early simulated testing suggests the value appears at three or more parallel sessions — light users don’t need it, so heavy agent users are who we’re building for.

Honest status: the data layer is built and in a gated acceptance test (25 of 33 checks passed; 4 failures being fixed). The screen design is finished and being implemented. Value has not yet been tested with users.

24rounds of design debate between Claude Fable 5.1 and gpt-6-astra
2,593line implementation spec, stress-tested in 4 adversarial review passes
9 / 10internal trial cases where the founder could decide from the AI-written card alone
105automated tests passing; backend built in 3 days
miruka echo session detail design: what you need to do, what the AI is doing, the question with options and their impact, and a reply drafted into the agent's input
Session detail design: what you need to do, what the AI is doing, each option’s impact — and your answer drafted straight into the agent’s input. Sample data.

しりとり部 (Shiritori Club)

LIVE

Party word games for friends, in the browser.

Rooms of 2–10 friends play auction, 大喜利 (improv comedy prompts) and NG-word games in real time while talking on Discord. Each room is one Cloudflare Durable Object with WebSockets, SQLite storage and alarm-driven deadlines, so a dropped player can rejoin within 60 seconds.

Public demo; capacity beyond small groups has not been load-tested. Built by an AI agent team under the same review process as our other products.

Play the live demo →
~2 daysfrom first commit to public deployment
3games, with 30 NG-word condition types
80passing tests (65 unit + 15 browser)
0console errors across 64 API operations in public QA

Automation Odyssey (working title) · by Imomushi Works, gooji’s game label

IN DEVELOPMENT

Write code. Watch the world start moving.

A programming-automation game for Steam (Godot 4.7 / C#). Players script robots in a Python-like language to bring a shut-down logistics base back online. The signature mechanic: you publish your own @public APIs on one robot and call them from another — API design is the puzzle. The language interpreter and a deterministic simulation run as a headless-tested core, separate from the engine.

Built with Claude Code: a planner → implementer → parallel-reviewer sub-agent workflow, with 129 commits co-authored by Claude. The game itself needs no AI at runtime — a deliberate choice for a one-time-purchase title.

Chapter 1 runs end to end on the developer’s machine. No public build or Steam page yet; outside playtesting hasn’t started.

8missions playable in Chapter 1
2,147automated checks passing (interpreter, simulation, in-engine self-tests)
191commits since July 2026
~45klines of C#, including tests

video-context

IN DEVELOPMENT

Turn what was said and shown in a meeting into context your agent can cite.

Decisions happen in meetings — in what people say and in the screens they share — and then they’re lost to the agents doing the work. video-context turns a recording into searchable, timestamped context that Claude Code can query through a JSON CLI: on-device speech recognition, screen captures and vector search, with nothing sent to the cloud. Agents can propose transcript corrections; every change keeps its provenance and the original is never overwritten.

Made for Claude Code: designed as a tool agents call — no human UI — so Claude can cite exactly what was said or shown, and when.

Today’s prototype ingests public YouTube videos as a test source; meeting recordings with screen sharing are the target and not yet supported. In private use by the founder; the correction feature is not yet deployed.

6,025passing tests (4,315 unit + 1,710 database)
364commits in 15 days
9.51%character error rate of the on-device speech model we measured
13 / 16segments acceptable in the agent-correction pilot

flow-deck

MVP · LOCAL

Let Claude draw your thinking.

A visual flow builder for untangling processes as auto-laid-out graphs. Its MCP server exposes create_flow: ask Claude Code to map a process, and the diagram appears live in the editor.

600passing tests
16 daysto a working MVP with Claude Code integration

BAND M (working name)

CONCEPT

Book the show your band actually wants to play.

Japanese indie bands often absorb fixed ticket quotas (ノルマ) on bills they didn’t choose. BAND M lets bands assemble compatible co-bills and split a venue’s cost transparently within a shared budget. The concept was designed through a 20-round Claude-vs-Codex review covering market research, payments regulation and unit economics.

Concept stage: no code yet, and demand has not been validated. Break-even is modelled, not proven.

See the concept pages (Japanese) →
20debate rounds; both models agreed on the proposal at round 16
36screen mockups for bands, venues and fans
3landing-page drafts, one per audience
7 / mobookings to break even — a hypothesis to test first
HOW WE BUILD

Claude argues with a second model before we write code.

A neutral facilitator runs Claude and a model from another vendor, both read-only, against the same sources. Each must cite evidence and is told not to simply agree. The debate ends only when both sides have looked for holes and both declare the design complete.

1
Adversarial design debate24 rounds for miruka echo, 20 for BAND M — every claim sourced.
2
Spec audits before codingThe implementer writes down its understanding; a stronger model audits it. One audit caught 11 misunderstandings and 4 missing features.
3
Independent review after codingJudges who didn’t write the code. After “82 tests passing”, review still found a path that silently hid records.
debate loop (simplified)
# one round = one turn from each model
for round in 1..N:
  claude -p --model claude-fable-5-1 \
    "critique v{n}; cite sources"
  codex exec -m gpt-6-astra -s read-only \
    "find what Claude missed"
  if both_declare_complete: break

# miruka echo: "complete" was declared at v3, v4, v5 —
# each time the other model found a real gap.
WHERE WE ARE

What’s true today.

DONE

  • しりとり部 live on Cloudflare Workers
  • flow-deck MVP with a Claude Code MCP server
  • miruka echo data layer (hooks · MCP · SQLite)
  • video-context ingestion and search (video prototype)
  • Automation Odyssey Chapter 1 (8 missions)

NEXT

  • Pass miruka echo Gate 1, then build its screens
  • Founder’s 14-day real-use evaluation of miruka echo
  • Claude-powered prompts and commentary for 大喜利 rounds
  • Validate BAND M demand with real bands and venues

NOT YET

  • No external users or revenue yet
  • No outside funding (bootstrapped)
  • No load testing of the live game
  • BAND M is a concept, not code

Running many agents? Let’s talk.

We’re looking for developers who run several Claude Code sessions a day to try miruka echo first.

sakurai@gooji.dev