Agent Coherence MCP Server ★ New (July 12)¶
Stop AI agents from silently overwriting each other's shared files — a single-host coherence guard for multi-agent systems. Uses TLA+-verified algorithms to prevent file conflicts when multiple agents access the same filesystem concurrently.
Source: mcp.so (submitted July 12, 2026) · Author: Vlad Parakhin
What It Is¶
Agent Coherence is an infrastructure MCP server that prevents multi-agent file corruption. When multiple AI agents share a filesystem (e.g., a project directory, a docs repo, a shared workspace), concurrent writes can silently overwrite each other. This server provides file-level locking, version tracking, and conflict detection — verified correct with TLA+ formal methods.
Business Relevance¶
- Platform operators running multi-agent systems need file integrity guarantees
- Dev teams with multiple coding agents sharing a repo need write conflict prevention
- Content operators with multiple agents updating docs need version consistency
- CorpusIQ complement for any team running 2+ Hermes agents on shared infrastructure
Tools Available¶
| Tool | Description |
|---|---|
acquire_lock |
Request exclusive write access to a file |
release_lock |
Release file lock after write completes |
check_conflicts |
Detect if any agent's changes conflict with current state |
list_locks |
Show all active file locks and their owners |
get_version |
Get current version/hash of a file |
Quick Start¶
# Install via npx
npx @vparakhin/agent-coherence-mcp
# Or add to MCP client config
{
"mcpServers": {
"agent-coherence": {
"command": "npx",
"args": ["@vparakhin/agent-coherence-mcp"],
"env": {
"WATCH_DIR": "/path/to/shared/workspace"
}
}
}
}
Business Use Cases¶
- Multi-agent docs editing: 3 agents updating the same knowledge base — Coherence ensures no silent overwrites
- Code generation pipelines: Agent A generates code while Agent B reviews — Coherence prevents review of stale files
- Shared configuration management: Multiple agents updating system configs — version tracking prevents drift
- Agent audit trail: Track which agent wrote what, when, with full version history
Limitations¶
- Single-host only (not distributed — one machine, multiple agents)
- File-level granularity (not line-level — whole file locks)
- Requires agents to use the lock/check/release protocol (cooperative, not enforced)
- New project; production hardening likely needed
- TLA+ verification covers the algorithm, not the implementation
See Also¶
- SPM Structured Project Memory — Project memory management
- MCP Long-Term Memory — Agent memory persistence