KanaAI

現在の製品・配布先は Kotori日本語入力 です。 紹介サイトはKotoriの現行情報を案内します。 以下のKanaAI実装・beta.2検証記録は旧プロジェクトの履歴です。

Mozcを基盤に、ローカルAIを育てるWindows日本語IME。

A local-first Windows Japanese IME built on Mozc, Rust, and bounded local AI.

Windows TSF · Mozc baseline · Local AI · No cloud dependency

Important

Status: public beta published - unsigned, with a bundled local AI that now actually starts.

Download: v0.1.0-beta.2 (KanaAI-0.1.0-Setup.exe / KanaAI-0.1.0-x64.msi, Windows x64, about 1.1 GB - the model is in the package).

Japanese introduction and installation guide: KanaAI website. This repository contains the public source and engineering record. The Release body records the evidence and limitations of each published binary. The product is a native Windows TSF text service, not the retired Workbench/CLI demo. Do not treat a local build, a DLL load, a source test, or a loopback model mock as proof that the end-user IME is ready.

What beta.2 established, and what it did not. The installer lifecycle is machine-verified by receipt for the published hash (11/11 phases), Japanese input is operator-confirmed with the text service observed loaded in eleven real processes, and the bundled AI is measured starting from the package's own opt-in on the product path - a 1.1 GB model loaded into a child process at a 1.6 GB working set, with byte-hash verification and a completed inference. What is not established is that any of that improves a conversion: the published beta.2 AI-on/AI-off measurement found no change in conversion results, and the last recorded model-quality evaluation in STATE.md was not shippable in any of its six roles. Read the Release body before installing.

Everything still runs offline. The runtime listens on loopback only, behind a per-process key, and performs no network access.

This beta is not a completed product and .goal-complete has not been created.

Windows release contract · Product requirements · Current engineering state · Open work and release checklist · Local AI runtime plan · Security · Contributing

What KanaAI is

KanaAI is being built as a practical Japanese input method for Windows:

1. Mozc owns everyday input. Romaji composition, segmentation, conversion, dictionary behavior, candidates, preedit, and commit remain owned by the pinned Mozc engine.
2. Rust owns bounded orchestration. A private broker tracks sessions and generations, validates results, and keeps optional AI work away from the synchronous key path.
3. Local AI improves candidates when it is safe. The intended first bundle uses a pinned Qwen GGUF and a CPU-only llama.cpp runtime. A model result can only reorder a bounded, unchanged candidate set; it cannot invent text or take ownership of commit.
4. Mozc remains the fallback. Missing model, timeout, malformed output, process failure, protected field, or low resources must leave normal input usable.

The current implementation target is a Windows x64 native TSF TIP. A browser Workbench, HTTP API, CLI, or source bridge is development tooling—not the product.

Product principles

Principle	What it means
Native first	The release target is a registered Windows TSF text service used in ordinary desktop applications.
Fast path stays fast	Key and preedit callbacks do not wait for a model, disk I/O, or network I/O.
AI is bounded	At most a small candidate window is reranked; IDs, text, and commit authority are validated.
Local by default	The intended release runs its model/runtime locally and does not require a cloud account or remote endpoint.
Explainable fallback	A missing or rejected AI result returns the Mozc baseline instead of silently rewriting text.
Privacy is a release gate	Protected/password contexts, bounded context, no user-data logging, and no hidden telemetry are required.

Current status

This table describes evidence, not marketing claims.

Area	Current state
Mozc/TIP and Windows runtime build	Windows x64 TIP/server/renderer artifacts have been built and statically checked.
Installer	v0.1.0-beta.2: unsigned Windows x64 Setup/MSI, about 1.1 GB, with a bundled local AI model and runtime. Read the release limitations before installing.
Windows registration	The installed candidate has x64/x86 COM registration (mozc_tip64.dll / mozc_tip32.dll) and a Japanese input profile.
Real app input	Operator-confirmed, no machine-verified receipt. On 2026-09-27 the operator confirmed on the validation machine that the installed candidate works as an IME (kana input, conversion, commit). mozc_tip64.dll was independently observed loaded in several live processes of that session. The automatic W1 harness has not produced a passing receipt; the blockers were harness observation defects, documented in the Release body. Notepad/Edge/Office composition, candidate, cancel, focus-loss, and restart checks are not machine-verified.
AI bundle	beta.2 includes Qwen2.5-1.5B GGUF and the llama.cpp CPU runtime. Startup on the product path is measured; the published AI-on/AI-off comparison found no change in conversion results. No quality improvement is claimed.
W1/W2	W2 (installer lifecycle) is machine-verified by receipt for this exact hash. W1 (real-app Japanese input) is operator-confirmed only.
Public release	GitHub prerelease v0.1.0-beta.2, with fixed hashes and limitations in the Release body. Japanese introduction and installation guide.
Code signing	Not required for the beta by user decision. This beta is unsigned (NotSigned for both MSI and Setup), and the Release body states the resulting Windows warning.
Product completion	Not complete. .goal-complete has not been created.

The authoritative current record is STATE.md. Historical Linux/WSL milestones are retained there for provenance and are not current Windows release evidence.

First AI bundle candidate

The user approved this pairing for the first implementation pass:

Component	Pinned identity	License / digest
Model	Qwen/Qwen2.5-1.5B-Instruct-GGUF @ 91cad51170dc346986eccefdc2dd33a9da36ead9	Apache-2.0; qwen2.5-1.5b-instruct-q4_k_m.gguf; 1,117,320,736 bytes; SHA-256 6a1a2eb6d15622bf3c96857206351ba97e1af16c30d7a74ee38970e434e9407e
Runtime	ggml-org/llama.cpp release b11146 @ 7fe450e19305b828c199d602c23a8337aaa1f03b	MIT; llama-b11146-bin-win-cpu-x64.zip; 18,560,055 bytes; SHA-256 14cf1303ca9ac3abd94816850532f9f9a69ac66fbaca3776fc6f9061c2fac1d1

The complete package is expected to be roughly 1.1 GB. These are approved implementation inputs, not evidence of Japanese IME quality, latency, memory use, crash recovery, or completed legal review. The full license, notices, SBOM, model provenance, and runtime closure must be present in the eventual release. See docs/LOCAL_AI.md.

Architecture

              Windows applications
                      │
             Windows TSF TIP (x64)
                      │
       authenticated private named pipe
                      │
             Rust broker / session owner
          ┌───────────┴───────────┐
          │                       │
  bounded Mozc conversion   local AI queue
          │                       │
   pinned Mozc engine       llama-server.exe
                                  │
                    pinned Qwen GGUF + CPU runtime

Repository map

Path	Responsibility
platform/windows-tsf/	Native TSF integration, reviewed Mozc patches, registration, installer, and smoke contracts.
crates/kanai-broker/	Authenticated protocol, session/generation ownership, bounded enhancement queue, and local model adapter.
crates/kanai-mozc/	Supervised Mozc bridge and bounded session pool used by the broker.
crates/kanai-core/	Conversion contracts, ranking policy, cache/learning interfaces, and model metadata.
crates/kanai-api/ / crates/kanai-cli/	Development lab and health tooling; not the end-user IME.
evals/	Development fixtures and evaluation tooling; synthetic fixtures are not release quality evidence.
third_party/mozc/	Pinned upstream Mozc source; not KanaAI-owned code.

Try the source checkout

This is a developer checkout, not an end-user installation. Native Windows builds require the pinned Visual Studio/Bazel/Mozc toolchain described in the handoff guide.

git clone --recurse-submodules https://github.com/aruiki/kanai.git
cd kanai
npm ci

Run the source-only checks:

cargo fmt --all -- --check
cargo clippy --locked --workspace --all-targets -- -D warnings
cargo test --locked --workspace --all-targets
npm test
npm run build

For Windows-specific build and validation notes, see the Windows roadmap. The developer Workbench can then be started with:

npm run dev

The Workbench is loopback-only development software. It is not the Windows IME, and its API must not be exposed to a network.

Windows product path

The intended user experience is a one-click Setup.exe flow:

1. Download the release asset from the GitHub Release page and verify its SHA-256 against the value published in the Release body. The beta is unsigned, so Windows may show a SmartScreen or publisher warning — do not disable SmartScreen, Smart App Control, antivirus, or enterprise policy to get past it.
2. Start Setup.exe and follow Windows UAC prompts. It writes an embedded MSI to a temporary folder and starts Windows Installer; it does not build, download, or require manual file placement.
3. Select KanaAI from the Windows Japanese input-method list, restarting the target application if it does not appear.
4. Use Mozc composition and conversion immediately. This beta contains no local AI model or runtime; no AI feature is active, and none is claimed.
5. Remove KanaAI from Settings → Apps → Installed apps when no longer needed.

Download: v0.1.0-beta.1 prerelease — KanaAI-0.1.0-Setup.exe and KanaAI-0.1.0-x64.msi, with SHA-256 values in the Release body. Do not use a local .local build as a release URL.

The installer lifecycle (install, registration, reinstall, upgrade, downgrade refusal, uninstall, absence) is machine-verified by receipt for that exact hash. Real-application Japanese input is operator-confirmed only, with no machine-verified receipt; privacy, performance, secure-field, AI fallback, and independent-verifier gates are not complete. Those gaps are listed in the Release body and are a documented deviation from the release contract.

Privacy and failure behavior

The release target is local-only. The intended production path does not send raw keys, full documents, clipboard contents, learning corpora, or model prompts to a remote provider. AI requests are bounded to the current reading, approved short context, and a small candidate set. Password/protected fields must skip AI, history, and learning.

These are release requirements, not claims that every current source slice is already complete. The current browser lab uses transparent localStorage; it is not the future encrypted native profile store. See SECURITY.md and docs/LOCAL_AI.md.

Release gates

Before a GitHub prerelease can be published, the same immutable source and artifacts must have receipts for:

• native Windows x64 TIP registration and ordinary application input;
• kana, conversion, candidate, commit, cancel, focus, and restart behavior;
• install, repair/upgrade, uninstall/reinstall, and rollback;
• bundled model/runtime size, SHA-256, SBOM, licenses, notices, and provenance;
• AI OFF/ON, model kill, timeout, malformed output, and Mozc fallback;
• secure/password fields, no user-data logs, and no unexpected network egress;
• CPU/memory/latency measurements and held-out Japanese quality evaluation;
• an independent verifier report tied to the source commit and artifact hashes.

Published beta deviation (2026-09-27). v0.1.0-beta.1 was published with the installer-lifecycle receipts only. Real-application input is operator-confirmed without a machine-verified receipt, and the secure-field, privacy, performance, AI-fallback, and independent-verifier items above are not met. The Release body lists each gap explicitly; nothing above is waived by the publication, and the remaining items stay open.

A prerelease is not the same as a completed product. Passing a beta gate does not create .goal-complete.

Roadmap

The first beta is deliberately narrow: Mozc baseline, x64, unsigned, no AI bundle, with the limitations stated in the Release body. The AI work below is still required for product completion; it is simply not what the first beta ships.

• Pin and stage the reviewed Mozc/TSF source path.
• Build Windows x64 TIP/server and unsigned installer candidates.
• Add bounded Rust broker/session/generation contracts.
• Complete W2 installer lifecycle evidence (receipt: 11/11 phases, one machine).
• Complete W1 native input evidence as a machine-verified receipt (currently operator-confirmed only; the harness cannot yet observe preedit or enumerate target modules).
• Freeze a clean source commit and publish the unsigned, Mozc-only GitHub prerelease with real hashes and real verified/unverified results.
• Bundle the approved Qwen GGUF and llama.cpp runtime.
• Add automatic runtime supervision and offline installer staging.
• Run real AI quality, privacy, resource, and failure-injection tests.
• Independently verify full product completion.

License and third-party notices

KanaAI-owned source is offered under either MIT or Apache-2.0, at the user's option. Mozc and every bundled runtime/model retain their own notices and license terms. The eventual release must include the exact third-party notices and SBOM; loading an asset does not relicense it.

KanaAI is not affiliated with or supported by Google. Mozc is an independent open-source project.

Contributing and support

Read AGENTS.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md before contributing. Please redact API keys, model files, user dictionaries, private text, and personal paths from issues and support reports.