rustfs-mcp

A per-add-on RustFS (S3) object-storage admin MCP for the kubero chat. When a user selects a RustFS add-on in the chat's @-context selector, the kubero server (the feature-MCP broker) surfaces this server's tools namespaced s3__*, gated by the add-on selection. The 5th feature MCP after postgres/redis/mysql/kafka.

RustFS is an S3-compatible object store (operator CR kind Tenant, group rustfs.com); in-cluster its S3 API is http://<instance>-io:9000, plain HTTP with path-style addressing.

Architecture

ONE long-lived ClusterIP Deployment (rustfs-mcp:3010 in ns kubero-mcp) serves every RustFS add-on in the workload cluster. The target endpoint + credentials

• access mode arrive per request as HTTP headers set by the kubero broker:

--connection-from-request reads these via the FastMCP lowlevel request_ctx contextvar; a missing/invalid access mode ⇒ restricted (fail-safe). The kubero server reaches the service via apiserver port-forward — no public ingress.

Tools

Reads (always available): list_buckets, list_objects, stat_object, bucket_usage, get_object_text, get_bucket_policy.

Writes (refused under restricted mode): create_bucket, delete_bucket, put_object_text, delete_object, set_bucket_policy.

Run

pip install -r requirements.txt

# Production (in-cluster): per-request connection from headers.
python main.py --transport streamable-http --connection-from-request \
  --streamable-http-host 0.0.0.0 --streamable-http-port 3010

# Local (stdio): binds one endpoint from S3_ENDPOINT_URL / AWS_* env (.env.template).
python main.py

CI / deploy

Mirrors the other stackblaze feature MCPs: auto-tag-on-main → docker-release (amd64, self-hosted kubero runner, ghcr.io/stackblaze/rustfs-mcp) → deploy.yaml rolls the ClusterIP Deployment to the workload cluster (gated on KUBECONFIG_B64_SHARED). Repo secrets: CHAIN_PAT, KUBECONFIG_B64_SHARED. Seed a v0.1.0-sb.0 tag once so auto-tag can increment.