ShardStitch

3:47pm. Mid-refactor. "Rate limit reached."

Your code is half-written. Your plan is in your head. ShardStitch captures your entire session — git diff, changed files, dependency graph, intent — and stitches it into the next AI tool. Keep building like nothing happened. Works after you're locked out, when most tools can't.

Install & activate

pip install shardstitch              # install the launcher
shardstitch install <your-license-key>   # download + activate (key from your purchase email)
shardstitch                          # launch the local dashboard at localhost:8765

One-time purchase, no subscription. The download is verified by SHA-256 before it ever runs. Get a key at https://shardstitch.com.

What it does

When Claude, Cursor, Codex, or Gemini locks you out mid-session, ShardStitch:

1. Scans your project — git diff, changed files, recent commits, your notes
2. Builds a dependency graph — impact radius, god nodes — and warns the next AI what's dangerous to touch
3. Generates a continuation prompt formatted for the target tool
4. Injects it into the tool's pickup file (CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules)

Under 30 seconds, end to end. Reads disk, not the dead AI session — so it works after a rate limit, context overflow, or crash.

Meet Hivy — the on-device layer

Hivy is a hive of small local models that pre-process every session before any big AI sees it. It runs on your machine, so it's there exactly when the rate-limited model isn't.

• Trims the context — Hivy returns only the relevant slice of your project instead of dumping everything. 94% fewer tokens, measured on a real repo — not a borrowed study.
• Knows when to spend — the Oracle Router answers easy questions on a free local model and sends only the hard ones to a paid one. Base-plan tokens, pro-plan results.
• Catches the drift — Hivy filters off-task tangents so the handoff carries the work, not the noise.

3 retrieval tiers

Pick the depth the work needs — each ships with Hivy:

23 supported AI tools

Claude Code, Claude Desktop, Cursor, Codex CLI, Gemini CLI, Windsurf, Aider, Kiro, Amazon Q Developer, DeepSeek, OpenCode, Trae, Factory Droid, OpenClaw, Amp, Cline, Roo Code, Kilo Code, Crush, Kimi, Qwen Code, Antigravity, and Grok.

7 surfaces

• Web dashboard — chat (Normal) + full dashboard (Expert)
• VS Code / Cursor extension — Alt+G hotkey
• MCP server — 27 tools; Claude Code, Cursor, Windsurf call it directly
• CLI — shardstitch from any project folder
• Desktop app — standalone Windows + Linux exe (Mac coming soon)
• Autosave hooks — auto-capture on session end
• Local LLM failover — Ollama, vLLM, LM Studio (zero cloud)

Key features

• Per-agent formatting — one handoff shape per supported tool
• Oracle Router — routes cheap queries local, hard ones to your cloud key
• Persistent memory — project decisions survive across sessions
• Session Vault — always-on local recovery history
• Hash-chained audit trail — append-only SHA-256 timeline (tamper-detecting)
• Dependency graph — impact radius + god-node detection

Local by default

Your code stays on your machine. No telemetry, no cloud sync. License validation goes to Polar at activation and during a quiet periodic re-check — that's the only routine network call. A cloud LLM you configure is used only when you explicitly enable it.

Links

• Website: https://shardstitch.com
• GitHub: https://github.com/shardstitch/shardstitch
• X / Twitter: https://x.com/ShardStitch
• Support: support@shardstitch.com