Vector MCP server
Commit Nov 13, 2026Lets an AI client create, populate, list and delete vector collections and run semantic, lexical, and fused searches across them.
Also known as Knuckles-Team/vector-mcp
- Stars
- 16
- Tools
- 1
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 13, 2026
- License
- MIT
What is Vector MCP?
Vector lets an AI client create, populate, list and delete vector collections and run semantic, lexical, and fused searches across them. It exposes 1 tool, such as x-mcp-enabled.
Summary generated from the project's README
About Vector MCP server
Managing vector collections and performing advanced retrieval is what vector-mcp enables for developers using AI clients. It provides a governed way to create, populate, list, and delete collections while supporting semantic, lexical, and fused search methods. The server is designed to work locally through stdio for IDEs like Cursor or Claude Desktop, or remotely via streamable-HTTP for production environments.
Users can choose from several storage backends depending on their infrastructure needs. The system defaults to the native medulla engine for a zero-infrastructure setup, but also supports PostgreSQL with pgvector, Qdrant, and MongoDB Atlas. Deployment is handled through uvx or by installing specific extras for the chosen provider. To maintain security, the server uses secret references for credentials and allows operators to filter available tools through environment variables or CLI arguments to manage the LLM context window.
What you can do with it
- 1Hybrid information retrievalDevelopers can instruct an AI to perform complex searches combining different retrieval strategies.
- 2Context window optimizationOperators can restrict the specific tools exposed to an LLM to prevent excessive token usage.
- 3Multi-provider vector storageTeams can switch between different database backends like Qdrant or MongoDB depending on their environment.
Written from the project's README. Every use case is checked against a quote from it.
Vector MCP tools 1
Detected from the READMEx-mcp-enabled
Configuration
COLLECTION_MANAGEMENTTOOLDATABASE_TYPEDB_PASSWORD_REFDB_PORTDB_USERNAME_REFMCP_TOOL_MODEMONGODB_MAX_POOL_SIZEMONGODB_REQUEST_TIMEOUT_MSMONGODB_URI_REFPOSTGRES_MAX_POOL_SIZEPOSTGRES_REQUEST_TIMEOUTQDRANT_API_KEY_REFQDRANT_REQUEST_TIMEOUTVECTOR_DB_TYPE
Automated checks
Ran Nov 13, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableknuckles-team/vector-mcp is public and reachablegithub.com/knuckles-team/vector-mcp ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/knuckles-team/vector-mcp ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model1 tool description and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNo tool named like delete, drop or execPass
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubvector-mcp
Action-routed MCP and agent interfaces for governed vector collection management and retrieval. The native default is medulla. Secure opt-in providers cover PostgreSQL/pgvector, Qdrant, and MongoDB Atlas.
Governed capability
- MCP tools:
vector_create_collection,vector_add_documents,vector_delete_collection,vector_list_collections,vector_semantic_search,vector_lexical_searchandvector_search - Skill provider: the consolidated
vector-mcp-operationsworkflow - Ontology provider: the packaged vector retrieval ontology
- Source connector provider: a read-only vector collection inventory preset
- Runtime configuration: AgentConfig, environment variables, and secret references
- Privacy posture: no checked-in endpoints, credentials, personal identity, or host paths
Install
Use the smallest extra set required by the deployment:
uvx --from 'vector-mcp[mcp]' vector-mcp
The runtime requires scion-sdk>=0.2.0 and its self-contained full
medulla engine contract. A bare numeric-only or partial engine profile is not a
supported deployment.
For a selected storage provider:
uv add 'vector-mcp[postgres]'
uv add 'vector-mcp[qdrant]'
uv add 'vector-mcp[mongodb]'
The all extra enables every supported optional provider plus the agent, Langfuse, and
Logfire runtimes. Production images should install only the providers they operate.
MCP configuration
The package includes a neutral agent-start configuration containing only the command, condensed tool mode, and tool toggles. Runtime values are inherited from AgentConfig or injected by the operator. Detailed instructions on how to use the underlying API wrappers, extended schema bindings, and developer SDK references are maintained in docs/index.md.
MCP
This server use dynamic Action-Routed tools to optimize token overhead and maximize IDE compatibility.
Available MCP Tools
Auto-generated from the live MCP server — do not edit by hand.
Condensed action-routed tools (MCP_TOOL_MODE=condensed)
| MCP Tool | Toggle Env Var | Description |
|---|---|---|
vector_create_collection, vector_add_documents, vector_delete_collection, vector_list_collections |
COLLECTION_MANAGEMENTTOOL |
Collection lifecycle. |
vector_semantic_search, vector_lexical_search, vector_search |
SEARCHTOOL |
Semantic, lexical and fused retrieval. |
2 action-routed tool(s) · 0 verbose 1:1 tool(s). Each is enabled unless its <DOMAIN>TOOL toggle is set false; MCP_TOOL_MODE selects the surface (intent default — the six verb-tools, granular set loaded on demand · condensed action-routed · verbose 1:1 · both). Auto-generated — do not edit.
Detailed tool schemas, parameter shapes, and validation constraints are preserved in the usage guide.
Dynamic Tool Selection & Visibility
This MCP server supports dynamic toolset selection and visibility filtering at runtime. This allows the operator to restrict the set of exposed tools to prevent blowing up the LLM's context window.
The operator can configure tool filtering via multiple input channels:
- CLI Arguments: Pass
--toolsor--toolsets(or their disabled counterparts--disabled-toolsand--disabled-toolsets) during startup. - Environment Variables: Define standard environment variables:
MCP_ENABLED_TOOLS/MCP_DISABLED_TOOLSMCP_ENABLED_TAGS/MCP_DISABLED_TAGS
- HTTP SSE Request Headers: Pass custom headers during transport initialization:
x-mcp-enabled-tools/x-mcp-disabled-toolsx-mcp-enabled-tags/x-mcp-disabled-tags
- HTTP SSE Request Query Parameters: Append query parameters directly to the operator's transport connection URL:
?tools=tool1,tool2?tags=tag1
When query strings or parameters are provided, an LLM-free Knowledge Graph resolution layer (using DynamicToolOrchestrator) matches query intents against known tool tags, names, or descriptions, with safe fallback and automated 24-hour background cache refreshing.
MCP Configuration Examples
Install the connector-focused
[mcp]extra. Examples usevector-mcp[mcp]to add FastMCP / FastAPI throughscion-sdk[mcp].
stdio Transport (local IDEs — Cursor, Claude Desktop, VS Code)
{
"mcpServers": {
"vector-mcp": {
"command": "uvx",
"args": [
"--from",
"vector-mcp[mcp]",
"vector-mcp"
],
"env": {
"MCP_TOOL_MODE": "intent",
"COLLECTION_MANAGEMENTTOOL": "True",
"DATABASE_TYPE": "medulla",
"DB_PASSWORD_REF": "secret://runtime/db-password",
"DB_PORT": "5432",
"DB_USERNAME_REF": "secret://runtime/db-username",
"MONGODB_MAX_POOL_SIZE": "20",
"MONGODB_REQUEST_TIMEOUT_MS": "30000",
"MONGODB_URI_REF": "secret://runtime/mongodb-uri",
"POSTGRES_MAX_POOL_SIZE": "20",
"POSTGRES_REQUEST_TIMEOUT": "30",
"QDRANT_API_KEY_REF": "secret://runtime/qdrant-api-key",
"QDRANT_REQUEST_TIMEOUT": "30",
"SEARCHTOOL": "True",
"VECTOR_DB_TYPE": "medulla"
}
}
}
}
Runtime references require an alias-aware launcher such as Arbora. Other launchers must omit those entries and inject the resolved values through their own runtime secret boundary.
Streamable-HTTP Transport (networked / production)
{
"mcpServers": {
"vector-mcp": {
"command": "uvx",
"args": [
"--from",
"vector-mcp[mcp]",
"vector-mcp",
"--transport",
"streamable-http",
"--port",
"8000"
],
"env": {
"TRANSPORT": "streamable-http",
"HOST": "127.0.0.1",
"PORT": "8000",
"MCP_TOOL_MODE": "intent",
"COLLECTION_MANAGEMENTTOOL": "True",
"DATABASE_TYPE": "medulla",
"DB_PASSWORD_REF": "secret://runtime/db-password",
"DB_PORT": "5432",
"DB_USERNAME_REF": "secret://runtime/db-username",
"MONGODB_MAX_POOL_SIZE": "20",
"MONGODB_REQUEST_TIMEOUT_MS": "30000",
"MONGODB_URI_REF": "secret://runtime/mongodb-uri",
"POSTGRES_MAX_POOL_SIZE": "20",
"POSTGRES_REQUEST_TIMEOUT": "30",
"QDRANT_API_KEY_REF": "secret://runtime/qdrant-api-key",
"QDRANT_REQUEST_TIMEOUT": "30",
"SEARCHTOOL": "True",
"VECTOR_DB_TYPE": "medulla"
}
}
}
}
Alternatively, connect to a pre-deployed Streamable-HTTP instance by url:
{
"mcpServers": {
"vector-mcp": {
"url": "http://localhost:8000/vector-mcp/mcp"
}
}
}
Run a reviewed container image as a least-privilege stdio child (no listener or published port):
docker run -i --rm \
--read-only \
--cap-drop=ALL \
--security-opt=no-new-privileges \
--pids-limit=256 \
--tmpfs /tmp:rw,noexec,nosuid,nodev,size=64m \
-e TRANSPORT=stdio \
-e MCP_TOOL_MODE=intent \
-e COLLECTION_MANAGEMENTTOOL=True \
-e DATABASE_TYPE=medulla \
-e DB_PASSWORD_REF \
-e DB_PORT=5432 \
-e DB_USERNAME_REF \
-e MONGODB_MAX_POOL_SIZE=20 \
-e MONGODB_REQUEST_TIMEOUT_MS=30000 \
-e MONGODB_URI_REF \
-e POSTGRES_MAX_POOL_SIZE=20 \
-e POSTGRES_REQUEST_TIMEOUT=30 \
-e QDRANT_API_KEY_REF \
-e QDRANT_REQUEST_TIMEOUT=30 \
-e SEARCHTOOL=True \
-e VECTOR_DB_TYPE=medulla \
registry.example.invalid/vector-mcp@sha256:<digest> vector-mcp
For containerized network HTTP, supply an authenticated TLS ingress (or
direct server TLS), exact MCP_ALLOWED_HOSTS, and an exact trusted-proxy
CIDR policy through the operator-owned deployment profile. The generator
does not emit an unauthenticated non-loopback listener.
Auto-generated from the code-read env surface (MCP_TOOL_MODE + package vars) — do not edit.
Additional Deployment Options
vector-mcp can also run as a local container (Docker / Podman / uv) or be
consumed from a remote deployment. The
Deployment guide has full, copy-paste
mcp_config.json for all four transports — stdio, streamable-http,
local container / uv, and remote URL:
- Local container / uv — launch the server from
mcp_config.jsonviauvx,docker run, orpodman run, or point at a local streamable-http container byurl. - Remote URL — connect to a server deployed behind Caddy at
https://vector-mcp.example.invalid/mcpusing the"url"key.
Environment Variables
Package environment variables
| Variable | Example | Description |
|---|---|---|
HOST |
127.0.0.1 |
|
PORT |
8000 |
|
TRANSPORT |
stdio |
options: stdio, streamable-http, sse |
ENABLE_OTEL |
— | |
EMBEDDING_TLS_PROFILE_REF |
secret://runtime/embedding-tls-profile |
Configure AgentConfig EMBEDDING_MODELS and its referenced runtime credentials. |
DOCUMENT_DIRECTORY |
— | Required only for filesystem ingestion. Supply the operator-owned root at runtime. |
DATABASE_TYPE |
medulla |
Backend used when db_type is unspecified. Default is the native medulla engine (local, zero-infra, durable). Options: medulla, postgres, mongodb, qdrant. DATABASE_TYPE is the canonical variable; VECTOR_DB_TYPE is accepted as an alias for backward compatibility. |
VECTOR_DB_TYPE |
medulla |
|
DB_HOST |
— | postgres/qdrant host |
DBNAME |
— | postgres/mongodb database name |
DB_PORT |
5432 |
|
DB_USERNAME_REF |
secret://runtime/db-username |
|
DB_PASSWORD_REF |
secret://runtime/db-password |
|
MONGODB_URI_REF |
secret://runtime/mongodb-uri |
|
QDRANT_API_KEY_REF |
secret://runtime/qdrant-api-key |
|
QDRANT_HTTP_ALLOWED_PRIVATE_HOSTS |
— | comma-separated SSRF allowlist for a private Qdrant host |
POSTGRES_REQUEST_TIMEOUT |
30 |
seconds, 1-300 |
POSTGRES_MAX_POOL_SIZE |
20 |
connections, 1-100 |
MONGODB_REQUEST_TIMEOUT_MS |
30000 |
milliseconds, 1-300000 |
MONGODB_MAX_POOL_SIZE |
20 |
connections, 1-100 |
QDRANT_REQUEST_TIMEOUT |
30 |
seconds, 1-300 |
COLLECTION_MANAGEMENTTOOL |
True |
|
SEARCHTOOL |
True |
|
VECTOR_MCP_MCP_IMAGE |
— | e.g. registry.example.invalid/vector-mcp@sha256: |
VECTOR_MCP_AGENT_IMAGE |
— | e.g. registry.example.invalid/vector-mcp@sha256: |
Inherited agent-utilities variables (apply to every connector)
| Variable | Example | Description |
|---|---|---|
MCP_TOOL_MODE |
intent |
Tool surface: intent | condensed | verbose | both |
MCP_ENABLED_TOOLS |
— | Comma-separated tool allow-list |
MCP_DISABLED_TOOLS |
— | Comma-separated tool deny-list |
MCP_ENABLED_TAGS |
— | Comma-separated tag allow-list |
MCP_DISABLED_TAGS |
— | Comma-separated tag deny-list |
EUNOMIA_TYPE |
none |
Authorization mode: none | embedded | remote |
EUNOMIA_POLICY_FILE |
mcp_policies.json |
Embedded Eunomia policy file |
EUNOMIA_REMOTE_URL |
— | Remote Eunomia authorization server URL |
OTEL_EXPORTER_OTLP_ENDPOINT |
— | OTLP collector endpoint |
MCP_CLIENT_AUTH |
— | Outbound MCP child auth: oidc-client-credentials | basic | none |
OIDC_CLIENT_ID |
— | OIDC client id (service-account auth) |
OIDC_CLIENT_SECRET_REF |
secret://identity/oidc-client-secret |
Runtime secret reference for the OIDC service account |
MCP_BASIC_AUTH_USERNAME |
— | HTTP Basic username (MCP_CLIENT_AUTH=basic) |
MCP_BASIC_AUTH_PASSWORD_REF |
secret://identity/mcp-basic-password |
Runtime secret reference for HTTP Basic auth (MCP_CLIENT_AUTH=basic) |
DEBUG |
False |
Verbose logging |
PYTHONUNBUFFERED |
1 |
Unbuffered stdout (recommended in containers) |
MCP_URL |
http://localhost:8000/mcp |
URL of the MCP server the agent connects to |
PROVIDER |
— | Operator-configured LLM provider for the agent |
MODEL_ID |
— | Operator-configured model id for the agent |
ENABLE_WEB_UI |
True |
Serve the AG-UI web interface |
25 package + 20 inherited variable(s). Auto-generated from .env.example + the shared agent-utilities set — do not edit.
Every variable the server reads, grouped by purpose. See .env.example for the
canonical, copy-paste list — including the DATABASE_TYPE / GRAPH_SERVICE_SOCKET /
GRAPH_SERVICE_AUTH_SECRET connection settings for the native medulla backend. Backend
endpoints, database locations, and credentials for opt-in providers (Postgres/Qdrant/Mongo/
Chroma/Couchbase) are never README-documented literal values or MCP tool arguments — they resolve
through AgentConfig and secret:///env:///vault:// references at runtime.
MCP server / transport
| Variable | Description | Default |
|---|---|---|
TRANSPORT |
stdio, streamable-http, or sse |
stdio |
HOST |
Bind host (HTTP transports) | 0.0.0.0 |
PORT |
Bind port (HTTP transports) | 8000 |
MCP_TOOL_MODE |
Tool surface: condensed, verbose, or both |
condensed |
MCP_ENABLED_TOOLS / MCP_DISABLED_TOOLS |
Comma-separated tool allow/deny list | — |
MCP_ENABLED_TAGS / MCP_DISABLED_TAGS |
Comma-separated tag allow/deny list | — |
PYTHONUNBUFFERED |
Unbuffered stdout (recommended in containers) | 1 |
Tool toggles
Each action-routed tool can be disabled individually via its toggle env var (set to false).
The full list is in the Available MCP Tools table above.
| Variable | Description | Default |
|---|---|---|
COLLECTION_MANAGEMENTTOOL |
Enable the collection-management tool | True |
SEARCHTOOL |
Enable the search tool | True |
Telemetry & governance
| Variable | Description | Default |
|---|---|---|
ENABLE_OTEL |
Enable OpenTelemetry export | True |
OTEL_EXPORTER_OTLP_ENDPOINT |
OTLP collector endpoint | — |
OTEL_EXPORTER_OTLP_PUBLIC_KEY / OTEL_EXPORTER_OTLP_SECRET_KEY |
OTLP auth keys | — |
OTEL_EXPORTER_OTLP_PROTOCOL |
OTLP protocol (e.g. http/protobuf) |
— |
EUNOMIA_TYPE |
Authorization mode: none, embedded, remote |
none |
EUNOMIA_POLICY_FILE |
Embedded policy file | mcp_policies.json |
EUNOMIA_REMOTE_URL |
Remote Eunomia server URL | — |
See .env.example for a copy-paste starting point.
Provider and ontology integration
The package contributes its skills, prompts, ontology, and source connector through Python entry points. The collection-inventory connector is intentionally read-only and registers collection metadata, not document or embedding payloads.
Generated connector signatures must be recreated only after the installed MCP schema is observed and a release signing key is provided at runtime. A signature from an older tool schema or ontology must never be copied forward.
Development checks
Low-cost checks that do not start providers:
python -m compileall -q vector_mcp
Provider tests use mocked SDK boundaries and make no network calls. Live qualification is a separate deployment gate and must use operator-provided AgentConfig and secrets.
Documentation
The slim :mcp streamable-http container (docker/mcp.compose.yml) publishes
0.0.0.0:8000 with a /health check; see Deployment for the full
compose service definition.
License
See LICENSE.
Deploy with agent-utilities-deployment
Provision this package with the consolidated agent-utilities-deployment
workflow. It selects an installed-package, editable-source, or immutable-container
path; records only runtime secret and TLS-profile references in AgentConfig; and
runs doctor, registration, policy, observability, and rollback gates. Ask your agent
to "deploy vector-mcp with agent-utilities-deployment".
| Install mode | Command |
|---|---|
| Installed package | uv tool install "vector-mcp[mcp]", then run vector-mcp |
| Editable source | uv pip install -e ".[mcp]", then run vector-mcp |
| Immutable container | deploy registry.example.invalid/vector-mcp@sha256:<digest> through the operator-selected orchestrator |
The repository embeds no deployment profile, credential value, certificate path, or
environment-specific endpoint. Supply those at runtime through AgentConfig and the
configured secret provider.
Installation
Pick the extra that matches what the operator want to run:
| Extra | Installs | Use when |
|---|---|---|
vector-mcp[mcp] |
Slim MCP server only (scion-sdk[mcp] — FastMCP/FastAPI) |
The operator only run the MCP server (smallest install / image) |
vector-mcp[all] |
Everything (mcp + all vector backends) |
Development / both surfaces |
# MCP server only (recommended for tool hosting — slim deps)
uv pip install "vector-mcp[mcp]"
# Full agent runtime (Pydantic AI + medulla engine)
uv pip install "vector-mcp[mcp]"
# Everything (development)
uv pip install "vector-mcp[all]" # or: python -m pip install "vector-mcp[all]"
Container image
The shared connector pipeline renders a slim :mcp image (vector-mcp[mcp], entrypoint
vector-mcp) from pyproject.toml at release time:
pipelines-connector render --root .
docker build -t knucklessg1/vector-mcp:mcp docker/
The render command creates the release-owned docker/mcp.compose.yml; it runs the image with
the configured host publication.
Knowledge-graph database (medulla)
Graph storage, vector search and queries are served by medulla, run as its own database container. Deployment recipes (single-node + Raft HA), connection config, and the database architecture are in the medulla deployment guide.
Repository Owners
Contribute
Contributions are welcome! Please ensure code quality by running local checks before submitting pull requests:
- Format code using
ruff format . - Lint code using
ruff check . - Validate type-safety with
mypy . - Run test suites using
pytest
Vector MCP questions
What tools does Vector MCP provide?
It exposes 1 tool, including x-mcp-enabled.
Which vector databases does vector-mcp support?
It defaults to the native medulla engine and offers opt-in providers for PostgreSQL/pgvector, Qdrant, and MongoDB Atlas.
What MCP tools does vector-mcp expose?
It exposes collection create, add documents, delete, and list tools, plus semantic, lexical, and fused search tools.
How do I install vector-mcp for use with an MCP client?
Run it with uvx using the mcp extra, for example uvx --from 'vector-mcp[mcp]' vector-mcp.
Which transport modes can vector-mcp run over?
It supports stdio, streamable-http, and sse transports.
Can I limit which tools an AI client sees?
Yes, operators can filter tools and toolsets via CLI arguments, environment variables, HTTP headers, or query parameters.