Agenthold
Give Claude Code or Cursor a shared SQLite-backed ledger so parallel agents can claim keys, release locks, and read consistent versioned state instead of overwriting each other.
Overview
agenthold is a MCP server for the Build phase that provides shared versioned SQLite state and claim/release coordination for multi-agent workflows.
What is this MCP server?
- Standard tool set: claim, release, status, and wait for cooperative multi-agent locking
- Advanced mode adds raw get/set/delete, watch, list, history, clear, and export over the same store
- Persists state in a local SQLite file (default ./agenthold.db) via stdio MCP
- Install via uvx from PyPI with --tools standard|advanced switch
- Versioned shared state designed for workflows where agents hand off work sequentially or in parallel
- Server version 0.5.1 on PyPI
- Two tool profiles: standard (claim/release/status/wait) and advanced (get/set/delete/watch/list/history/clear/export)
- Default database path: ./agenthold.db
Community signal: 1 GitHub stars.
What problem does it solve?
When several agents touch the same repo or task list, they overwrite each other’s assumptions because there is no single versioned place to coordinate locks and status.
Who is it for?
Indie builders running parallel or chained coding agents on one machine who need simple locking and history without standing up Redis or a custom API.
Skip if: Teams that need cloud HA, fine-grained ACLs, or real-time collaboration UI—this is local SQLite coordination, not production workflow SaaS.
What do I get? / Deliverables
After you register agenthold, agents can claim shared keys, wait on peers, and read consistent state from one local database across sessions.
- Registered MCP server with standard or advanced tool surface
- Persistent shared state DB agents can read and coordinate through
- Optional history/export of state changes in advanced mode
Recommended MCP Servers
Journey fit
Multi-agent coordination is a build-time concern when you wire several agents into one repo or pipeline. Agent-tooling is the canonical shelf for MCP servers that manage agent runtime state rather than product features or deploy ops.
How it compares
MCP coordination store for agents, not a general database MCP or a single-agent memory skill.
Common Questions / FAQ
Who is Agenthold for?
Solo and indie builders using Claude Code, Cursor, or similar agents who run multiple automations against one codebase and need shared, versioned state with optional claim/release semantics.
When should I use Agenthold?
Use it during build when you orchestrate subagents or parallel tasks that must not double-book the same issue, file, or pipeline step.
How do I add Agenthold to my agent?
Add a stdio MCP entry that runs uvx Agenthold (or the PyPI package), set --tools to standard or advanced, optionally set --db to your SQLite path, then reload the agent’s MCP config.