memgraph-ingester-mcp 0.2.1

MCP tools for Memgraph graphs created by memgraph-ingester

memgraph-ingester-mcp

memgraph-ingester-mcp exposes safe, high-level MCP tools for Memgraph graphs produced by memgraph-ingester. It keeps agents out of long Cypher instruction blocks for normal work while still offering a project-scoped read-only Cypher escape hatch for unusual lookups.

The package is intended to be published to PyPI as memgraph-ingester-mcp under the ousatov-ua account.

Tools

Code graph tools:

• server_status: graph inventory, memory counts, and vector index status.
• code_orientation: compact package/type/call overview; pass sections to keep it focused.
• code_quality_stats: graph-wide code quality and quantity metrics; defaults to non-test code and a 5-row first page.
• code_hot_paths: compact hot-path candidates from size and call graph metrics; defaults to 5 owner/name rows per selected section.
• code_search: CodeChunk vector search for broad discovery; defaults to 5 non-test, deduped, range-first hits and omits text.
• code_lookup_type: exact class/interface/annotation lookup; test code and member expansion are opt-in.
• code_lookup_methods: exact method lookup with source ranges; defaults to 10 compact owner/name/path rows.
• code_callers, code_callees: compact, paginated, non-test, range-first call graph lookups; default first page is 10 rows.
• code_method_context: bundled exact method lookup plus caller and callee context; use instead of separate lookup/callers/callees calls while tracing one target.
• code_hierarchy: class ancestry, children, interfaces, and interface implementors.

Memory tools:

• memory_orientation: rules plus open findings, tasks, questions, and risks; compact=true omits body fields.
• memory_schema: allowed memory types, upsert fields, controlled values, and CodeRef targets.
• memory_search: MemoryChunk vector search with index-only hit metadata.
• memory_get: canonical memory node plus resolved CodeRefs.
• memory_upsert: create/update Decision, ADR, Rule, Context, Finding, Task, Risk, Question, or Idea.
• memory_update_status: lifecycle status update.
• delete_memory: delete one memory node plus its derived chunk and orphan CodeRefs.
• memory_link_code_ref: link memory to a resolved CodeRef target.
• memory_refresh_chunk, memory_refresh_embeddings: maintain derived memory RAG data.

Fallback:

• raw_read_cypher: read-only, project-scoped Cypher for edge cases.

Configuration

Use namespaced environment variables so this server does not collide with generic shell settings:

Variable	Default	Purpose
MEMGRAPH_INGESTER_MCP_BOLT_URI	bolt://127.0.0.1:7687	Memgraph Bolt URI
MEMGRAPH_INGESTER_MCP_USERNAME	unset	Optional username
MEMGRAPH_INGESTER_MCP_PASSWORD	unset	Optional password
MEMGRAPH_INGESTER_MCP_DATABASE	unset	Optional database name
MEMGRAPH_INGESTER_MCP_PROJECT	unset	Default project name
MEMGRAPH_INGESTER_MCP_QUERY_TIMEOUT_SECONDS	30	Query timeout
MEMGRAPH_INGESTER_MCP_READ_ONLY	false	Disable write tools when true
MEMGRAPH_INGESTER_MCP_EMBEDDING_MODEL	default	Metadata stamped on refreshed chunks
MEMGRAPH_INGESTER_MCP_EMBEDDING_DIMENSIONS	384	Expected memory embedding dimension

Run Locally

uv sync --dev
uv run memgraph-ingester-mcp

Run tests:

uv run ruff check .
uv run pytest

Build the package:

uv build --no-sources

Client Setup

Replace my-project with the project name used when running memgraph-ingester.

Codex

Add this to ~/.codex/config.toml:

[mcp_servers.memgraphIngester]
command = "uvx"
args = ["memgraph-ingester-mcp"]
startup_timeout_ms = 20_000

[mcp_servers.memgraphIngester.env]
MEMGRAPH_INGESTER_MCP_BOLT_URI = "bolt://127.0.0.1:7687"
MEMGRAPH_INGESTER_MCP_PROJECT = "my-project"

Claude Desktop

Add this to claude_desktop_config.json:

{
  "mcpServers": {
    "memgraphIngester": {
      "type": "stdio",
      "command": "uvx",
      "args": ["memgraph-ingester-mcp"],
      "env": {
        "MEMGRAPH_INGESTER_MCP_BOLT_URI": "bolt://127.0.0.1:7687",
        "MEMGRAPH_INGESTER_MCP_PROJECT": "my-project"
      }
    }
  }
}

Claude Code can also add the same stdio server directly:

claude mcp add-json memgraphIngester \
  '{"type":"stdio","command":"uvx","args":["memgraph-ingester-mcp"],"env":{"MEMGRAPH_INGESTER_MCP_BOLT_URI":"bolt://127.0.0.1:7687","MEMGRAPH_INGESTER_MCP_PROJECT":"my-project"}}'

GitHub Copilot in VS Code

Create .vscode/mcp.json in the workspace or add the server to your user MCP configuration:

{
  "servers": {
    "memgraphIngester": {
      "type": "stdio",
      "command": "uvx",
      "args": ["memgraph-ingester-mcp"],
      "env": {
        "MEMGRAPH_INGESTER_MCP_BOLT_URI": "bolt://127.0.0.1:7687",
        "MEMGRAPH_INGESTER_MCP_PROJECT": "my-project"
      }
    }
  }
}

Gemini CLI

Add this to Gemini CLI settings.json:

{
  "mcpServers": {
    "memgraphIngester": {
      "command": "uvx",
      "args": ["memgraph-ingester-mcp"],
      "env": {
        "MEMGRAPH_INGESTER_MCP_BOLT_URI": "bolt://127.0.0.1:7687",
        "MEMGRAPH_INGESTER_MCP_PROJECT": "my-project"
      },
      "trust": false
    }
  }
}

Agent Guidance

Prefer the dedicated tools over raw_read_cypher. Use RAG search tools only for broad discovery, then follow up with exact lookup tools before making claims or edits. The default first pages are small on purpose: use meta.hasMore and meta.nextSkip to paginate instead of raising limits preemptively. For performance work, start with code_hot_paths; add code_quality_stats only when you need graph-wide baselines. Compact method/search/call-graph rows are range-first and omit full signatures; request non-compact output only when signatures or modifiers are needed. Use code_method_context when tracing one method so one MCP call replaces lookup plus callers plus callees. Most code tools exclude src/test/ by default; pass include_tests=true only for test coverage or test-specific questions. Use memory_orientation(compact=true) for status checks and memory write tools for task lifecycle changes and CodeRef links so derived MemoryChunks stay refreshable.