Operators

Running kglite’s standalone binaries in production or automation. The Python package + wheel are out of scope here — those are the Python guide. This track is for operators deploying the standalone binaries.

CLI

kglite-cli ships the standalone kglite command for one-shot Cypher queries, scoped writes, agent JSONL sessions, graph description output, text export, diffs, and the interactive shell.

Bolt server (Neo4j wire protocol)

kglite-bolt-server speaks the Bolt v5.x wire protocol. The official Python driver path is regression-tested; other Bolt clients may connect, subject to KGLite’s documented protocol and Cypher limits.

After Phase F (2026-05-24):

  • neo4j:// URIs work via single-server routing (--advertise-addr flag for reverse-proxy deployments).

  • TLS via --tls-cert / --tls-key. Drivers connect with bolt+s:// or neo4j+s://.

  • db.labels() / db.relationshipTypes() yield Neo4j-conventional column names (label / relationshipType).

  • OCC enforced on commit (concurrent writers get Neo.ClientError.Transaction.ConflictDetected).

cargo install --path crates/kglite-bolt-server
kglite-bolt-server --graph my-graph.kgl --bind 127.0.0.1 --port 7687

MCP server (Model Context Protocol)

The kglite-mcp-server binary exposes Cypher queries + schema introspection over MCP for use with Claude Code / other MCP clients. It’s a single, pure-Rust implementation — reachable two ways, both running the identical server:

# Standalone binary (no Python at all — this operators track's default):
cargo install kglite-mcp-server
kglite-mcp-server --graph my-graph.kgl --mcp-config my_graph_mcp.yaml

# Or, for Python users, the same server ships inside the wheel (0.10.26+):
pip install kglite
kglite-mcp-server --graph my-graph.kgl --mcp-config my_graph_mcp.yaml

No Python runtime is required for the cargo install binary. The pip install path bundles the Rust server inside the wheel (statically linked into the extension, sharing the one engine — no separate wheel, no duplicated engine) and exposes the same kglite-mcp-server command via a thin console-script shim. (History: through 0.10.24 the wheel shipped a Python server; 0.10.25 retired it for cargo-only to stop two implementations drifting; 0.10.26 brought the command back to pip as the bundled Rust server.)

Two workspace archetypes for code intelligence

Beyond serving a single pre-built --graph file, the two most-wanted deployments both build a code graph over source on the fly (builder injected by codingest-mcp). Both ship as copy-pasteable manifests in the repo’s examples/:

  • Clone-and-explore GitHub repos (workspace.kind: github) — open_source_workspace_mcp.yaml. The agent calls repo_management('org/repo') to clone and graph a repo on demand.

  • Review a local directory (workspace.kind: local + set_root_dir + watch) — local_code_review_mcp.yaml. Point it at a checked-out tree; swap roots at runtime with set_root_dir(path); watch-mode auto-rebuilds on file change.

Register with the absolute binary path, then verify

Point the client’s command at the absolute path to the binary you mean (which kglite-mcp-server inside your active env), not a bare kglite-mcp-server — a bare name resolves against $PATH and can silently launch an older PATH-shadowing install with a different tool set. Then confirm the deployment stands up:

kglite-mcp-server --selftest --mcp-config my_graph_mcp.yaml

--selftest re-spawns the server with the same flags, drives a real MCP handshake, and prints green/red per capability (tools registered, github tools present when a token is set, graph hydrates). It exits non-zero on any failure, so it also works as a CI / deployment smoke gate. Remember to restart the server or client after editing any manifest / config — those are read once at boot.

Setup, manifest, and tool customization details live in the MCP servers guide; skill authoring (teaching agents how/when to use each tool) is in MCP skills.