The 30-second mental model.
Read disk truth
Git diff, changed files, recent commits, notes, file tree, and dependency graph. No live AI session required.
Build hot context
Hivy trims noise, flags drift, and keeps the useful state instead of dumping the whole transcript.
Format for the next tool
The continuation packet becomes Claude, Codex, Gemini, Cursor, Devin, or another target's starting state.
7 core surfaces and 2 delivery methods.
The seven core surfaces match the changelog: dashboard, extension, MCP, CLI, desktop, autosave and local LLM failover. Pickup files and web handoff are two delivery methods, not additional core surfaces. The nine cards below cover both groups.
Review and generate
Scan a project, review the packet, copy or download handoffs, and use Expert view for graph and savings details.
Alt+G in VS Code/Cursor
One-hotkey recovery from inside the editor.
49 callable tools
Claude Code, Cursor, Devin, and Windsurf can call scan, recall, handoff, checkpoint, memory, council, and blueprint tools.
Terminal recovery
Run ShardStitch from the project folder through pip, npm, or npx.
Windows, macOS, and Linux
Standalone app with dashboard, tray, and global hotkey. macOS is available.
Recovery before collapse
Silent local checkpoints keep a recent packet available if the session ends badly.
Keep moving locally
Ollama, vLLM, or LM Studio can help when a connected model is rate-limited. BYOK cloud routing is optional.
Native handoff files
Writes continuation packets to tool-readable files like CLAUDE.md, AGENTS.md, GEMINI.md, and .cursorrules where supported.
Clipboard packets for browser chats
Creates a clean copy/paste packet for web chats such as ChatGPT, Claude.ai, Gemini web, Grok, DeepSeek, Perplexity, Z.ai, Replit, Lovable, and Kimi.
Hivy: the on-device brain.
Hivy is a cascade of tiny local models and heuristics that preprocesses the recovery packet before a bigger model sees it.
Spend only when needed
Easy work can stay local. Hard queries can route to a user-configured cloud key or owned GPU node.
Spot lockout language
Recognizes rate limits, usage limits, quota exceeded, 429s, out-of-credit errors, and similar tool failures.
Trim the noise
Keeps salient state and drops dead weight so the next session starts lean.
Drop the dead branch
Flags off-task drift and keeps the lesson without carrying the failed path forward.
Fail-closed secret gate
Replaces secrets, keys, and emails with typed placeholders before outbound text is included.
Rotate before decay
Estimates thread pressure and nudges a checkpoint before the session collapses.
Council: review before trust.
Council is an opt-in 12-agent review engine for important recovery packets. It retrieves focused project chunks through Shard RRF/BM25, checks the continuation packet through specialized lenses, and runs a refutation pass before the next AI treats the handoff as ground truth.
12 review lenses
Warden, Architect, Tracker, Scribe, Auditor, Foreman, Historian, Critic, Planner, Sentinel, Mediator, and Oracle each look for a different failure mode: risky files, stale assumptions, missing tests, bad sequencing, security concerns, and conflicting signals.
Refute plausible mistakes
Council Verify tries to disprove packet claims before they become the next session's ground truth. It pairs with Claimed-vs-Disk checks so uncertain claims become warnings instead of confident instructions.
Recovery and continuity.
Local survival
Conversation evidence survives crashes, rate limits, and misclick-deletes when capture is available.
Trust wedge
Compares what the AI said happened against what git diff and files actually show.
Verified vs inferred
Every recap line can be marked as disk-verified or inferred, so guesses do not masquerade as facts.
Impact radius
Flags dependent files, central god nodes, and architecture communities before the next agent edits blindly.
Scoped project memory
Stores decisions, entries, and summaries by project scope.
Local ROI proof
Records handoffs and estimates token/dollar savings versus re-explaining from scratch.
Tools, adapters, and MCP.
ShardStitch supports 31 AI handoff targets across native integrations and web-chat handoff with tool-shaped handoff adapters: Aider, Amazon Q, Amp, Antigravity, ChatGPT, Claude Code, Claude Desktop, Claude.ai, Cline, Codex, Crush, Cursor, DeepSeek, Factory Droid, Gemini (web), Gemini CLI, Grok, Kilo Code, Kimi, Kiro, OpenClaw, OpenCode, Perplexity, Qwen Code, Roo Code, Trae, Devin and Windsurf, Z.ai (GLM), Replit, Lovable.
| Surface | What it means |
|---|---|
| Disk scan | Works for all tools because the source of truth is the project folder, not the old AI. |
| Manual copy/paste | Fallback for any tool without pickup files or MCP. |
| File injection | Writes to pickup files like CLAUDE.md, AGENTS.md, GEMINI.md, and .cursorrules where supported. |
| MCP | 49 MCP tools for scan, handoff, route, session handoff, rotation check, claims, memory, blueprint, council, provider switching, and more. |
Privacy and local-only behavior.
- Your code stays on your machine. No telemetry and no cloud sync.
- License validation talks to Polar, the storefront that issued the key, not a ShardStitch code server.
- Optional BYOK cloud routing is opt-in and only for hard queries you choose to send.
- Low-end mode still works: no GPU, no cloud, no model required for disk-truth recovery and BM25 retrieval.
Honest notes.
- Numbers are V9 build numbers. Package surfaces can differ, so marketing should keep one canonical number per surface.
- RRF uses hybrid BM25 plus sentence-transformer embeddings; do not market a fine-tuned reranker unless the live build is using one.
- Recall is strongest for disk-log tools. Browser-chat capture is best-effort and should be verified per tool.
- Dollar-savings claims should be shown only when wired to the live local savings endpoint.
- Council Verify and frontier-local routing depend on the relevant flags/nodes being available.