MEMGRAPH-MCP
智能体系统
一个持久化的多智能体编排器,具有以下特性:
显式运行图和检查点/恢复功能
编排器控制的并行委派
有界研究集群执行
编码、审查、修复、CI 和审批循环
通过 MCP 资源和工具暴露的后端存储中的项目内存
用于内存检索的真实 SQLite 向量索引
支持 Tavily 的可插拔外部研究后端
范围
此实现针对 MCP 2025-11-25 规范基准,使用官方 Python MCP SDK 和用于内存表面的 FastMCP 服务器。对于本地开发,它通过 stdio 运行。对于远程部署,请参阅 docs/remote_auth.md。
Related MCP server: ellmos-homebase-mcp
布局
app/runtime: 运行状态、调度器、编排器循环、检查点app/planner: 规划和图修订辅助工具app/agents: 用于研究、代码、审查、修复、CI、综合、审批的节点执行器app/memory: 基于 SQLite 的内存、检索和工件索引app/mcp_server:FastMCP资源、工具、提示词和服务器入口点tests: 验收和单元测试覆盖
本地使用
uv sync --group dev
uv run pytest
uv run agent-system-mcp检索和研究后端
内存条目使用
sqlite-vec索引到本地 SQLite 向量表中。默认嵌入提供程序为
auto:它优先使用真实的sentence-transformers模型,仅在模型无法加载时回退到确定性哈希提供程序。当节点提供
inputs.corpus时,研究使用内存语料库后端。如果设置了
TAVILY_API_KEY,无语料库的研究节点可以使用 Tavily 后端进行外部网络研究。如果既没有语料库也没有 Tavily 密钥,研究将返回有界的空结果,而不是编造来源。
嵌入配置
AGENT_SYSTEM_EMBEDDING_PROVIDER=auto|sentence-transformers|hashAGENT_SYSTEM_EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2AGENT_SYSTEM_EMBEDDING_CACHE_DIR=/path/to/cacheAGENT_SYSTEM_EMBEDDING_LOCAL_ONLY=true|false
示例:
AGENT_SYSTEM_EMBEDDING_PROVIDER=sentence-transformers uv run agent-system create-run "improve scheduler"本地传输
开发使用 MCP stdio 传输。
远程部署
远程部署在 v1 版本中仅提供文档说明。服务器记录了兼容 OAuth 2.1 的同意路径,并保留本地 stdio 作为默认开发模式。
完整文档
state.md: 当前项目状态、创建以来的变更以及路线图
docs/getting_started.md: 实现本地运行和 MCP 服务器的最快路径
docs/operator_guide.md: 系统全貌、操作和最佳实践
docs/developer_custom_graphs.md: Python API 使用、自定义图设计和节点负载参考
docs/codex_mcp_usage.md: Codex 应如何使用此 MCP 进行大型应用规划、内存和检查点执行
examples/plan_large_app.py: 以编程方式生成大型应用蓝图和运行程序的示例
Available Tools
10 toolscheckpoint_runBIdempotent
Apply a small state patch and create an explicit checkpoint.
Inputs: run_id (required), state_patch_json (required). Returns: object
create_runA
Create a new orchestrator run from a high-level goal and optional JSON spec.
Inputs: goal (required), spec_json. Returns: object
get_file_contextBRead-onlyIdempotent
Read file context from the orchestrator workspace.
Inputs: paths (required), include_neighbors. Returns: object
plan_goalBRead-onlyIdempotent
Generate a large-app execution blueprint from a high-level goal and optional JSON spec.
Inputs: goal (required), spec_json. Returns: object
record_decisionCIdempotent
Record a high-value architectural or implementation decision.
Inputs: title (required), run_id (required), node_id (required), rationale (required), +2 more. Returns: object
record_failure_modeAIdempotent
Record a reusable failure pattern or bug signature.
Inputs: run_id (required), node_id (required), summary (required), signature (required), +1 more. Returns: object
resume_runAIdempotent
Resume an orchestrator run until completion, pause, or failure.
Inputs: run_id (required). Returns: object
search_project_memoryARead-onlyIdempotent
Search project memory by query string and optional type filters.
Inputs: query (required), kind, limit, status. Returns: result
write_project_memoryCIdempotent
Write a durable memory entry with provenance to the backing store.
Inputs: key (required), kind (required), run_id (required), node_id (required), +5 more. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 10 tools
Most tools target distinct actions or resources, but plan_goal and create_run both take a high-level goal and optional JSON spec, making their boundaries unclear. Similarly, search_project_memory and list_related_artifacts both retrieve stored information in overlapping ways. Descriptions help clarify intent, but some pairs could still lead to misselection.
All tools follow a consistent verb_noun snake_case pattern, such as resume_run, write_project_memory, record_decision, and create_run. The naming is uniform and predictable across the entire toolset.
Ten tools is a well-scoped count for an orchestrator and memory management server. Each tool covers a meaningful operation without the set feeling bloated or overly sparse.
The core lifecycle is covered: creating and resuming runs, checkpointing, planning, memory read/write, and recording decisions/failures. Minor gaps exist, such as no explicit run status/query tool or delete operation for memory entries, but agents can generally work within the provided surface.
Maintenance
Related MCP Connectors
Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Persistent, inspectable memory for AI agents with lineage, correction, and a hosted MCP endpoint.
The AI orchestration agent for modern software teams.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceShared memory and orchestration for coding agents, enabling persistent knowledge, multi-agent coordination, and a canonical workflow across MCP-compatible AI clients.83-

ellmos-homebase-mcpofficial
AlicenseCqualityBmaintenanceEnables local-first LLM orchestration with persistent memory, knowledge management, routing, swarm patterns, API probing, tests, automation planning, and plugin discovery via a stdio MCP server, using SQLite for offline storage.51173 npm2MIT- AlicenseNot gradedqualityDmaintenanceA production-grade multi-agent workflow orchestrator built on the Model Context Protocol, featuring planner/executor/critic agents, durable run state, and replayable traces.15MIT
- AlicenseNot gradedqualityBmaintenanceEnables autonomous multi-agent orchestration using directed acyclic graph workflows to interact with external developer infrastructure such as Git, Postgres, and Docker through safety-gated MCP servers.Apache 2.0