Skip to content

chaydavs/gyst

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

321 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Gyst

Give your AI agent your team's memory. In 90 seconds.

npm install -g gyst-mcp && gyst init

The problem

AI coding agents are fast — but stateless. Every session your agent starts fresh: no memory of why the auth service keeps timing out, no knowledge of the deploy-on-Fridays rule, no record of decisions your team made last month. Your agent figures things out and then forgets them. The team never gets smarter.


The solution

  • Self-populating context layer — mined automatically from your git history, code comments, markdown docs, and session transcripts
  • Every session starts ready — your team's rules, conventions, and past decisions are injected before the first prompt
  • Works with every major agent — Claude Code, Cursor, Codex CLI, Windsurf, Gemini CLI

90-second quickstart

npm install -g gyst-mcp && gyst init

Example output:

┌─────────────────────────────────────────────────────┐
│  gyst init                                          │
├─────────────────────────────────────────────────────┤
│  ✓ Database initialized      .gyst/wiki.db          │
│  ✓ MCP server registered     Claude Code            │
│  ✓ MCP server registered     Cursor                 │
│  ✓ Hooks installed           post-commit            │
│  ✓ Scanning codebase ...     42 files               │
│  ✓ Loaded conventions        8 entries              │
│  ✓ Loaded decisions          3 entries              │
│  ✓ Ghost knowledge generated 5 entries              │
├─────────────────────────────────────────────────────┤
│  Restart your AI tool. Context will inject on the   │
│  next session start.                                │
└─────────────────────────────────────────────────────┘

Restart your AI tool. That's it.


What your agent gets

Context type What it is Example
Ghost Knowledge High-confidence facts about your codebase the agent must know "Never deploy on Fridays — batch job runs at 22:00 UTC and will conflict"
Conventions Coding standards scoped to files and directories "All API responses use the ApiResponse<T> envelope shape"
Decisions Architecture choices and the reasons behind them "Switched from Prisma to raw SQL — N+1 queries were killing dashboard load time"
Error Patterns Known failure signatures with their fixes "SQLITE_BUSY on parallel writes — use WAL mode + retry with backoff"

Supported agents

Agent Auto-detected Context injection Hook coverage
Claude Code 12 hooks
Cursor 4 hooks
Codex CLI 4 hooks
Windsurf 4 hooks
Gemini CLI 4 hooks

How it works

Mining: gyst init scans your codebase and git history to build a context layer — zero config, zero manual writing. It extracts conventions from your source files, decisions from commit messages and ADRs, error patterns from TODO/FIXME comments, and structural knowledge from the files your team touches most. The post-commit hook keeps it current automatically after every commit.

Injection: At every session start, Gyst injects your team's ghost knowledge and top conventions into the agent's context window. Subagents get the same injection. When a session ends, anything the agent learned gets distilled back into the context layer, so the next session starts even stronger.


Learn more


Dashboard

gyst dashboard   # opens at localhost:3579

Browse all context entries, inspect the knowledge graph, review low-confidence entries, and track context economics (leverage ratio, zero-result rate, intent mix).


Team mode

Git-sync (zero infrastructure)

The context layer lives in your repo. git pull syncs it automatically via the post-merge hook.

Shared HTTP server

# Admin — set up once
gyst create team "Acme Engineering"
GYST_API_KEY=gyst_admin_... gyst team invite
gyst serve --http --port 3456

# Each developer — one command
gyst join gyst_invite_abc123... "Alice" --server http://your-host:3456

Every developer's AI tools automatically point at the shared context layer. Knowledge grows as the team grows.


License

MIT — see LICENSE.

About

Team knowledge layer for AI coding agents. Extends Karpathy's LLM Wiki pattern for teams. Served via MCP. Works with Claude Code, Cursor, Codex CLI, Gemini CLI, Cline, Windsurf, and self-hosted LLMs.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors