Katana: The Bittensor Workspace Built by Ninja (SN 66)

Most coding agent subnets score miners against fixed benchmarks. Ninja (SN66) chose a different scoring method, one significantly harder to game.

The subnet's evaluation framework generates tasks from real GitHub commits, runs agents against those tasks, then scores results with a combination of changed-line similarity and LLM-based judging. The king of the hill keeps their spot only as long as no challenger beats them on the next real task. There's no static leaderboard to optimize against, no fixed test set to overfit.

Katana Is the Workspace That Bittensor Participation Has Always Needed

Participating in Bittensor today means assembling a stack by hand. You manage a wallet in one terminal, run mining workflows somewhere else, debug validators in another window, and research subnets from a browser tab you lose and reopen constantly.

Katana, the new workspace from the Ninja (SN66) team, was built to collapse that fragmentation into one environment. The xninja agent harness is installable directly from PyPI.

The product description is specific about what it integrates: Chutes (SN64), Targon (SN4), Lium (SN51), Ninja (SN66), Affine (SN120), Ditto (SN118), and Hippius (SN75) are all wired in natively. That list covers inference, trusted compute, GPU rental, coding agent evaluation, reinforcement learning, long-term agent memory, and decentralized storage. The subnets that a serious Bittensor participant touches most often are available without a separate setup process for each one.

What a Persistent Sandbox Actually Changes

Every workspace in Katana is a shareable, persistent cloud sandbox with a real filesystem, and sessions carry their context forward with a single prompt rather than resetting when you close the browser.

The more consequential feature is what happens when you attach your own hardware. External machines, including mining rigs and validator nodes, connect to the workspace and share their memory and context with the cloud sandbox, so an operation started on local hardware picks up in the cloud environment exactly where you left off. For miners managing multiple hotkeys and validator operators running across different machines, that's the specific friction Katana removes.

You manage environment variables and secrets globally, then scope them to individual workspaces or specific members within those workspaces. That level of operational control hasn't existed in a purpose-built Bittensor product before.

Multiple Harnesses, One Environment

Katana ships with native support for five agent harnesses: pi, Codex, Claude Code, Cursor, and xninja. Swapping the agent leaves the workspace intact, with no reconfiguration required.

xninja runs the Ninja agent locally and calls the validator-compatible contract the subnet expects, meaning it runs exactly the same evaluation path a validator uses to score miner submissions. Mining on SN66 through Katana means the development environment and the evaluation environment are identical. The gap between local testing and on-chain scoring disappears.

The team flagged Ditto (SN118) as a first-class integration going forward. Ditto provides a shared long-term memory and context layer for AI agents on the Bittensor network, and a coding agent retaining its working context across sessions through Ditto's persistent knowledge graphs changes how iterative mining work on SN66 operates.

The Workspace Is Also the Participation Tool

Katana wasn't scoped to mining alone. The product description names subnet research, wallet configuration, validator debugging, and code shipping as core use cases alongside mining workflows, treating every layer of Bittensor participation as work belonging in one environment.

The integration map tells you what the team actually believes about Bittensor's current infrastructure. Hippius (SN75) delivers decentralized file hosting and S3-compatible object storage on Bittensor, handling persistent storage for the workspace's real filesystem. Lium (SN51) supplies verified GPU compute. Chutes (SN64) covers serverless inference. Targon (SN4) adds trusted execution for privacy-sensitive workloads. Together, those integrations cover storage, compute, inference, and trusted execution in a single workspace, assembled for how Bittensor participants already work.

Katana Is Live. The Fragmented Setup Is Optional Now.

The SN66 team built the most defensible coding agent evaluation mechanism on Bittensor by choosing the mechanism hardest to game. Katana extends that logic to the workspace itself: one environment wired directly to the network's best commodities, replacing a five-tool fragmented stack.

Katana is live now, with additional integrations flagged for this week. If you're a miner, validator, or researcher on Bittensor, the question is whether you want to run a fragmented setup or an integrated one.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or trading advice. The information provided should not be interpreted as an endorsement of any digital asset, security, or investment strategy. Readers should conduct their own research and consult with a licensed financial professional before making any investment decisions. The publisher and its contributors are not responsible for any losses that may arise from reliance on the information presented.