SyncytiumMD

One canonical .syncytium/ context transpiled to 10+ AI-tool instruction formats (CLAUDE.md, AGENTS.md, .cursor/rules, …) with an agent handoff baton, lease-based workspace lock, CI drift gate, and a 23-tool MCP server.

Documentation

🧬 SyncytiumMD

One context. Every AI coding tool. Coordinated agents. Zero drift.

AGENTS.md is the instruction standard. SyncytiumMD is the workflow infrastructure around it.

npm CI MCP Registry Node License PRs Welcome

English Β· TΓΌrkΓ§e


SyncytiumMD in action

One init scaffolds .syncytium/, sync generates 25 bridge files across every tool, handoff passes the baton, and diff --check fails CI the moment a generated file drifts from the source of truth.


The problem

You use Cursor at work, Claude Code in the terminal, Copilot in pull requests, and three CLI agents on CI. Each one wants its own instruction file. So you maintain four copies of the same rules, they drift, and no agent knows what the others did yesterday.

The idea

Keep the rules once, in .syncytium/, and transpile them into every tool's native format.

.syncytium/          canonical, human-authored, git-tracked
  rules/*.md         code style, security, testing, architecture, …
  memory/decisions.md  ADRs β€” why the project is the way it is
  HANDOFF.md         live agent baton: who has it, what's next
        β”‚
        β”‚  syncytium sync
        β–Ό
CLAUDE.md   AGENTS.md   GEMINI.md   AGENT.md   CONVENTIONS.md
.cursorrules   .clinerules   .roomodes   .windsurfrules   .traerules
.cursor/rules/*.mdc          .gemini/antigravity/rules/*.md
.github/copilot-instructions.md + .github/instructions/*.instructions.md

Every generated file carries a banner saying where it came from, and syncytium diff --check fails your CI if any of them drift. syncytium clean removes only banner-tagged files β€” your hand-written rules are never touched.

Why not just use AGENTS.md?

You should use AGENTS.md β€” it is the right standard for shared instructions, and SyncytiumMD generates it for you (alongside every other format). But a standard file cannot answer these questions:

  • Which agent has the workspace right now? Cursor and Claude Code editing the same file at the same time produce silent, expensive conflicts.
  • What did the last agent actually finish β€” and what did it leave for me?
  • Did someone hand-edit one generated copy and forget the others?

SyncytiumMD is the workflow layer that answers them: a handoff baton, a lease-based lock, and a CI drift gate β€” all local-first, all git-tracked, all exposed as MCP tools so the agents can operate them themselves.

CapabilityAGENTS.md aloneSyncytiumMD
Shared instructions across toolsβœ… (one file, per-tool agents re-read it)βœ… (25 files, every tool's native format)
Agent handoff with audit trailβŒβœ…
Lease-based workspace lockβŒβœ…
CI drift gate (real unified diffs, exit 1)βŒβœ…
MCP server so agents self-serveβŒβœ… (23 zod-validated tools)
Recoverable deletion of rules/ADRsβŒβœ… (trash + restore)
3D knowledge graph of the whole contextβŒβœ…
Reverse-import existing configsβŒβœ…

How is this different from Rulesync, Ruler or conforme?

Those tools (and peers like them) are excellent rules compilers: one source, many tool configs. SyncytiumMD generates configs too β€” but the bet is different: the hard multi-agent problem is not formatting instructions, it is coordinating the agents that read them. Nobody else in this space ships an audited handoff baton, an agent-renewable workspace lease, an MCP server that operates the canonical context itself, and a 3D view of the whole workspace context in one CLI. Rulesync has 66 tool targets and no coordination; SyncytiumMD has 10 adapters and the coordination the incumbents skip. If tool coverage is all you need, use them β€” and syncytium import will happily absorb what they generated.

Quick start

npx syncytium init          # scaffold .syncytium/ (stack auto-detected)
npx syncytium sync          # generate every bridge file
npx syncytium graph         # open the 3D knowledge graph in your browser

Requires Node.js 20+. No global install needed; npx works.

What you get

10 built-in adaptersCursor, Claude Code, GitHub Copilot, Cline/Roo, Antigravity, Windsurf, Trae, OpenCode, AGENTS.md, GEMINI.md
Stack-aware startersTypeScript, JavaScript, Python, Go, Rust, Java, Kotlin, PHP, Ruby, .NET, Swift, Elixir, generic
Multi-agent safetyA lease-based workspace lock with heartbeat, a handoff baton, and a 100-entry audit trail
Drift detectionReal unified diffs, orphan detection, and a --check flag that actually exits non-zero
Live 3D graphWebGL knowledge cosmos, vault explorer, Markdown document reader β€” bound to 127.0.0.1
MCP server23 tools with zod-validated inputs, honest safety annotations, built-in rate limits, and recoverable deletions
CI + git hookssyncytium ci and syncytium hook install both gate on zero drift
--json everywhereEvery command emits machine-readable output for agents and CI

Commands

Run syncytium --help for the full list.

Context

CommandWhat it does
init [name]Scaffold .syncytium/. --template <stack>, --force
syncRegenerate all bridge files. -t/--target, --no-prune, -f/--force
watchAuto-sync on every change. -a/--agent also holds and renews the lock
diffShow drift with real patches. --check exits 1 in CI, --full for complete hunks
doctor11 health checks, each with an actionable fix: hint. --strict
lintValidate rules, frontmatter and ADR schema. --fix heals kebab-case and missing titles
validateValidate syncytium.config.json and every rule against the schema. --fix
cleanDelete only banner-tagged files. --dry-run
exportOne portable Markdown bundle you can paste anywhere. --html, --max-rule-chars
importReverse-migrate existing CLAUDE.md, .cursorrules, … into .syncytium/. --dry-run
adaptersList every registered adapter and its targets
statusOne-screen project + handoff + lock summary

Rules & decisions

syncytium rules [query]                       # search the catalog
syncytium rule add --title "..." --body "..." # add a canonical rule
syncytium rule show <id>                      # inspect one
syncytium rule update <id> --body-file x.md   # edit in place
syncytium rule remove <id>                    # delete + prune generated files

syncytium adr list                            # architectural decisions
syncytium adr add --title "..." --context "..." \
                  --decision "..." --consequences "..."
syncytium adr show ADR-006
syncytium adr remove ADR-006

Multi-agent coordination

syncytium lock acquire --agent "Cursor" --goal "Refactor auth"
syncytium lock heartbeat --agent "Cursor"    # extend the lease
syncytium lock status
syncytium lock release --agent "Cursor"

syncytium handoff -i                          # interactive wizard
syncytium handoff --from "Cursor" --to "Claude" \
  -s ready_for_review -g "Ship v1" -d "…" -t "…"
syncytium log -n 20                           # baton audit trail

Guards

syncytium hook install      # block commits on drift
syncytium hook install --auto-sync   # or just re-sync on every commit
syncytium ci                # .github/workflows/syncytium.yml drift gate
syncytium ci --force        # regenerate an existing workflow

Knowledge graph

syncytium graph                        # 3D WebGL cosmos + vault explorer
syncytium graph --category ide         # isolate IDE adapters
syncytium graph --compact              # hide file and tag nodes
syncytium graph --no-open --port 4000

SyncytiumMD 3D knowledge graph

The real syncytium graph cosmos for a generated project: rule nodes colored by category, adapters, generated files, handoff and ADR nodes β€” click any node to open the underlying Markdown.

The server binds to 127.0.0.1 and rejects non-loopback Host headers (DNS-rebinding guard). It refuses to serve .env*, *.pem, *.key, .git/** and other in-root secrets.

Clicking a tool in the 3D galaxy opens the actual Markdown files it generates: directory targets such as .cursor/rules/ become directory nodes that list their .mdc rules, each file opens in the document panel and is highlighted in 3D, and adapters are marked USED or EMPTY depending on whether their targets exist on disk.

MCP

{
  "mcpServers": {
    "syncytium": {
      "command": "npx",
      "args": ["--ignore-scripts", "-y", "-p", "syncytium-md@0.3.2", "syncytium-mcp"],
      "cwd": "${workspaceFolder}",
      "env": {
        "SYNCYTIUM_TOKEN": "${SYNCYTIUM_TOKEN}"
      }
    }
  }
}

Hardening notes: pin the exact package version instead of @latest, pass --ignore-scripts when launching through npx, keep only the 3-5 MCP connectors a workspace actually needs, and pass any short-lived, read-only credential through the env block rather than command arguments.

Or syncytium mcp if you already have the CLI installed. The server exposes 23 tools, including syncytium_get_context, syncytium_handoff, syncytium_create_rule, syncytium_update_rule, syncytium_remove_rule, syncytium_record_decision, syncytium_sync, syncytium_diff, syncytium_doctor, syncytium_validate, syncytium_validate_fix, syncytium_lock_acquire/release/status, syncytium_get_graph, syncytium_list_trash, syncytium_restore and more. Every input is zod-validated and closed with additionalProperties: false; every property is documented; every stable read tool declares an outputSchema and returns matching structuredContent.

Built-in MCP safety. Every tool advertises MCP readOnlyHint, destructiveHint, idempotentHint and openWorldHint annotations, and the server enforces a per-tool sliding-window rate limit (for example 5 destructive calls per minute). Destructive tools default to dryRun: true; deleting needs dryRun: false and confirm: true. Rules and ADRs are archived under .syncytium/memory/trash/ before removal, so syncytium_restore can undo the operation. syncytium_doctor is strictly read-only. These limits are enforced server-side, so they hold even when a client ignores the annotations.

MCP Registry. The repository root carries a server.json manifest (official registry.modelcontextprotocol.io format) describing the npm package, the stdio transport and the mcp launch argument. The discoverability workflow regenerates llms.txt and llms-full.txt, verifies the manifest against the package (including mcpName), and publishes it with GitHub OIDC - no stored secret and no manual step. A unit test pins the manifest version to package.json so the registry never advertises stale metadata.

.syncytiumignore

Excludes generated targets from both sync and diff. Full gitignore syntax, including **, ?, directory-only patterns and ! negation:

# Don't manage these tools
.cursorrules
.github/instructions/

# Except this one
!.github/instructions/keep.md

Custom adapters

Declare any tool in syncytium.config.json β€” no code required:

{
  "customAdapters": [
    {
      "id": "zed",
      "name": "Zed",
      "targetFile": ".rules/zed.md",
      "includeHandoff": true,
      "includeArchitecture": true
    }
  ],
  "enabledAdapters": ["cursor", "claude", "zed"]
}

Library usage

import { SyncytiumEngine } from 'syncytium-md';

const engine = new SyncytiumEngine(process.cwd());
await engine.sync();
const { hasDrift, summary } = await engine.diff();
const graph = await engine.getKnowledgeGraph({ category: 'ide' });

Development

npm install
npm run verify      # typecheck -> build -> test (Node 20+)

Demo assets are reproducible: node scripts/record-demo.mjs re-records the real CLI session and the live 3D graph into docs/assets/ (needs ffmpeg + a local Chromium; see the header comment in the script).

See CONTRIBUTING.md and .syncytium/architecture.md.

License

MIT Β© MiraΓ§ Birben