dbx-tools-graphiti

Native launcher for Graphiti with local Neo4j and LiteLLM processes configured for Databricks Model Serving. It runs directly on the host without Docker, Podman, or another container runtime.

Install from PyPI:

uv add dbx-tools-graphiti

Or install the current main branch:

uv add "dbx-tools-graphiti @ git+https://github.com/reggie-db/dbx-tools.git@main#subdirectory=packages/py/graphiti"

Key features

• launches upstream Graphiti's HTTP MCP server at http://127.0.0.1:8000/mcp/;
• runs Neo4j Community 5.26 as a native background process;
• starts dbx-tools-litellm with an optional CLI profile, the configured CLI default, or ambient Databricks App authentication;
• supervises Graphiti and managed LiteLLM with Honcho so they share one lifecycle, receive SIGTERM as process groups, and receive SIGKILL after Honcho's bounded shutdown grace if needed;
• journals successful graph mutations to Postgres and reconstructs an ephemeral graph backend during startup;
• defaults to databricks-gpt-5-nano and the 1024-dimensional databricks-gte-large-en embedding model;
• reuses executables from PATH and installs missing Java 21, uv, Neo4j, and Graphiti source through mise;
• pins Graphiti and Neo4j versions for repeatable local environments;
• caches downloads, Python dependencies, Neo4j data, credentials, and logs;
• needs no caller-owned config.yaml and does not vendor Graphiti code.

Quick start

Outside a Databricks App, configure a working Databricks CLI profile. A Databricks App uses its ambient service-principal authentication. The launcher installs mise when needed, handles Java, LiteLLM, Graphiti, and Neo4j, and installs uv only when it is not already available:

uv run dbx-graphiti start

The launcher uses DATABRICKS_CONFIG_PROFILE when set. Otherwise it runs databricks auth profiles --output json --skip-validate and uses the one entry marked "default": true. --profile <name> is an optional override, not a requirement.

The first run downloads about 120 MB of Neo4j plus the pinned Graphiti release, creates Graphiti's uv environment, generates a local Neo4j password, starts LiteLLM and Neo4j, and then runs Graphiti in the foreground. Later runs reuse the installed assets.

Honcho stops the sibling process when Graphiti or managed LiteLLM exits. On Ctrl-C or SIGTERM it forwards SIGTERM to each child process group, waits up to five seconds, then sends SIGKILL to any remaining group. The launcher stops Neo4j after Honcho finishes.

For background operation:

uv run dbx-graphiti up
uv run dbx-graphiti status
uv run dbx-graphiti down

Commands

• start starts Neo4j, then runs Graphiti and managed LiteLLM under Honcho in the foreground. Missing prerequisites are installed on demand. This is the default.
• up starts all three services in the background.
• down signals the Honcho supervisor, which stops Graphiti and managed LiteLLM before the launcher stops Neo4j.
• status prints process state, model selection, and the MCP URL as JSON.
• env prints resolved database, proxy, and model settings as JSON. Its output includes the Neo4j password and must be treated as secret.

Arguments after -- are forwarded to upstream Graphiti:

uv run dbx-graphiti start -- --port 9000 --group-id my-agent

Postgres persistence

DelegatingGraphDriver accepts any Graphiti GraphDriver and delegates its provider behavior, operations, sessions, transactions, search, and maintenance to that driver. Mutating Cypher statements are appended to a supplied ordered storage driver before the graph operation or transaction commits. During the first index setup, the wrapper clears the delegated graph and replays the stored mutations in order without journaling them again.

PostgresWriteStorage provides the durable implementation. It stores a namespaced append-only JSONB journal and accepts the async SQLAlchemy engine created by dbx-tools-postgres:

from databricks.sdk import WorkspaceClient
from dbx_tools.graphiti.persistence import (
    DelegatingGraphDriver,
    PostgresWriteStorage,
)
from dbx_tools.postgres import create_async_engine

engine = create_async_engine(WorkspaceClient(), pool_pre_ping=True)
storage = PostgresWriteStorage(engine, namespace="memory-service")
driver = DelegatingGraphDriver(graph_driver, storage)

The bundled MCP launcher enables this automatically when any of these settings is present:

• JOURNAL_DATABASE_URL: explicit PostgreSQL URL. The launcher uses asyncpg.
• PGHOST, LAKEBASE_ENDPOINT, or LAKEBASE_INSTANCE_NAME: resolve the connection and rotating credential through dbx-tools-postgres and WorkspaceClient.
• JOURNAL_NAMESPACE: isolates one journal within the table. The launcher derives a stable value from its data directory when omitted. A direct dbx_tools.graphiti.server invocation must set it explicitly.
• JOURNAL_TABLE: journal table name. Defaults to graphiti_write_journal.

When persistence is configured, Postgres initialization or replay failure stops server startup rather than running without durability. The journal is restart recovery for one live graph instance. It does not replicate new writes into other concurrently running Graphiti instances. A process crash after the graph write-ahead append but before the graph commit can leave an unacknowledged mutation in the journal; restart recovery applies journal entries at least once. Graphiti's UUID-backed mutation queries are compatible with this replay model, but a custom delegate or write predicate must supply replay-safe mutations. If the local Neo4j credential no longer matches its ephemeral data directory, the launcher resets that directory only when a Postgres journal is configured, then Graphiti rebuilds it from the journal. Without durable storage, an authentication mismatch fails startup rather than deleting local graph data.

Provisioning and caching

The package deliberately keeps orchestration separate from Graphiti itself:

1. dbx_tools.core.bin checks PATH before asking mise for a tool.
2. When mise is missing on macOS or Linux, the official checksum-verifying installer runs under a cross-process lock.
3. Missing tools are installed globally with mise use -g --yes, then resolved with mise which or mise where.
4. Java 21, uv 0.11, and Neo4j Community 5.26.12 use their mise registry backends.
5. Graphiti 0.29.3 uses mise's HTTP backend against the pinned release source archive because the GitHub release has no platform binary asset.
6. uv sync --project <checkout>/mcp_server creates the upstream environment.
7. A generated Neo4j password is stored with mode 0600.
8. The packaged LiteLLM proxy starts against the selected Databricks profile, and Graphiti receives its OpenAI-compatible URL and model settings through environment variables and CLI flags.

The launcher is supported on macOS and Linux. The cache root is:

• macOS: ~/Library/Application Support/dbx-tools/graphiti
• Linux: ${XDG_DATA_HOME:-~/.local/share}/dbx-tools/graphiti

Set DBX_GRAPHITI_HOME to override it. The directory contains links to the mise-managed tools plus launcher state, logs, and Neo4j data. Removing it permanently removes the local graph data; mise manages its own download cache and installation directories separately.

Configuration

Callers do not supply a Graphiti config.yaml. The server creates an empty temporary YAML file for the lifetime of the upstream process because upstream requires the argument. Model and server settings resolve from CLI option, environment variable, then package default:

• --profile / DATABRICKS_CONFIG_PROFILE: an optional Databricks profile override for managed LiteLLM. When both are absent, the launcher uses the Databricks CLI profile marked as default.
• --model / MODEL_NAME: defaults to dbx/databricks-gpt-5-nano.
• --embedder-model / EMBEDDER_MODEL: defaults to dbx/databricks-gte-large-en.
• --embedder-dimensions / EMBEDDER_DIMENSIONS: defaults to 1024.
• --litellm-host / LITELLM_HOST: defaults to 127.0.0.1.
• --litellm-port / LITELLM_PORT: defaults to 4000.
• --litellm-url / LITELLM_URL: selects an external OpenAI-compatible LiteLLM endpoint.
• --manage-litellm, --no-manage-litellm / MANAGE_LITELLM: explicitly controls whether the launcher owns the proxy.
• LLM_STRUCTURED_OUTPUT_MODE: defaults to json_object.
• GRAPHITI_GROUP_ID: defaults upstream to main.
• GRAPHITI_HOST and GRAPHITI_PORT: environment-only listener settings. Without a port, the launcher uses DATABRICKS_APP_PORT when present and 8000 otherwise. Without a host, it binds 0.0.0.0 in a Databricks App and 127.0.0.1 elsewhere. The AppKit plugin selects a loopback endpoint for both.
• NEO4J_URI and NEO4J_DATABASE: default to bolt://127.0.0.1:7687 and neo4j.

The launcher sets Graphiti's OpenAI provider and embedding dimensions directly. No OpenAI key is required for its managed local proxy.

To use a separately managed LiteLLM instance:

uv run dbx-graphiti start \
  --litellm-url https://models.example/v1 \
  --no-manage-litellm

Setting LITELLM_URL also selects external mode automatically. A direct OPENAI_API_URL selects external OpenAI-compatible mode and requires OPENAI_API_KEY. --manage-litellm overrides either environment choice when the launcher should still own the local proxy.

Explicit NEO4J_* values override generated defaults, which lets the Graphiti process use an existing Neo4j server. The launcher still manages its local Neo4j process; use upstream Graphiti directly if lifecycle ownership belongs to an external database administrator.

Graphiti owns MCP tools, graph behavior, LLM calls, embeddings, and migrations. This package owns repeatable installation, Databricks defaults, and process lifecycle. To run it beside an AppKit server through one Databricks App port, use @dbx-tools/appkit-graphiti. See the upstream MCP server documentation for its complete API.

Modules

• cli: Cyclopts commands and CLI-over-environment option binding;
• settings: model, embedding, profile, and LiteLLM resolution;
• runtime: on-demand provisioning and Honcho lifecycle;
• server: upstream MCP entry point, temporary config, and persistence wiring;
• proxy: loopback Caddy process used by the AppKit plugin;
• persistence: delegating graph driver and Postgres write-ahead journal;
• supervisor: detached up entry point.