Put an S3-compatible policy and audit boundary between agent credentials and your bucket.
Use your bucket
Bring your bucket.
Connect S3-compatible storage and apply a versioned WebAssembly pipeline to supported writes and processed reads. Each operation is bound to the policy revision that ran.
Use managed storage
Skip the bucket.
Managed object storage with the same workspace credentials, pipelines, and transaction history. Use MCP or S3; 1 GiB per workspace is included.
Policy enforcement
Supported writes run before storage; processed reads run before bytes reach the agent context. Both bind execution to the exact pipeline revision.
ReadAgents see the view you allow. The raw object stays in storage.
WriteApply protection before the object reaches storage.
Live pipeline
See the difference between the original record, the view an agent receives, and a protected value that remains usable for matching.
[
{
"email": "alice@example.com",
"card": "4111••••1111"
},
{
"email": "alice@example.com",
"card": "4111••••1111"
}
][
{
"email": "[redacted]",
"card": "[redacted]"
},
{
"email": "[redacted]",
"card": "[redacted]"
}
][
{
"email": "join_8f3d",
"card": "join_c4a9"
},
{
"email": "join_8f3d",
"card": "join_c4a9"
}
]Original record
Two records begin with the same private details.
Agent-safe view
Maskura redacts the fields that should not enter agent context.
Protected matching
Same source values become the same keys for matching.
Access controls
Apply policy to the representation an agent receives, not only to the object key it can request.
Each protected operation resolves one immutable pipeline revision. Component digests, configuration, limits, byte counts, fuel, and duration follow the execution evidence.
Deterministic AES-SIV protects selected values while preserving equality, so joins and deduplication can work without exposing the original field.
Hybrid X25519 + ML-KEM-768 encapsulation protects recoverable fields. Only the public key crosses the boundary; decryption remains in the client.
Compose built-in or custom Wasm components through a public WIT contract. Fresh instances run with bounded memory, fuel, deadlines, and no storage or credential access.
Use Maskura-managed storage or connect an S3-compatible provider. Policy stays at the access boundary instead of inside one storage vendor.
Point an existing S3 client at Maskura, or use the open-source maskura-mcp server. Both enter through the same data path.
Format coverage
Text records, schema-aware Avro OCF objects, columnar Parquet and Lance files, and Iceberg tables run through format-specific pipelines.
JSONL, JSON, CSV, TSV, plain text, and Avro OCF are selected by Content-Type and can be processed on write and read.
New! Lance and Parquet bring schema-aware columnar processing (each behind a default-off gate), and the Iceberg table layer applies one consistent policy across every file of a table. Other unsupported binary objects remain raw S3 passthrough.
Limits are enforced before processing and vary by path and format.
Read the limits and streaming detailsClients + tools
Use the same S3-compatible gateway from your CLI, application code, or local MCP server. Writes run the configured pipeline; processed reads opt in with x-maskura-process: read.
Maskura speaks the S3 API — point any S3 client at the gateway. PII is redacted before it lands in your bucket.
# Maskura credentials retain the s4_ and s4s_ prefixes for compatibility
export AWS_ACCESS_KEY_ID=s4_your_access_key
export AWS_SECRET_ACCESS_KEY=s4s_your_secret_key
export AWS_ENDPOINT_URL=https://api.maskura.dev
# Write an object — email + card are detected and redacted in transit
aws s3 cp customer.json s3://my-bucket/customer.json \
--endpoint-url $AWS_ENDPOINT_URL
# Or with curl (header auth)
curl -X PUT https://api.maskura.dev/my-bucket/contact.json \
-H "x-maskura-access-key: s4_your_access_key" \
-H "x-maskura-secret-key: s4s_your_secret_key" \
-H "Content-Type: application/json" \
--data-binary '{"email":"alice@example.com","card":"4111111111111111"}'
# Read it back — stored object is redacted
curl https://api.maskura.dev/my-bucket/contact.json \
-H "x-maskura-access-key: s4_your_access_key" \
-H "x-maskura-secret-key: s4s_your_secret_key"Use boto3 with a Maskura API key as SigV4 credentials and point it at the gateway endpoint.
pip install boto3
import boto3
s3 = boto3.client(
"s3",
endpoint_url="https://api.maskura.dev",
aws_access_key_id="s4_your_access_key",
aws_secret_access_key="s4s_your_secret_key",
region_name="us-east-1",
)
# Write through Maskura.
s3.put_object(
Bucket="my-bucket",
Key="customer.json",
ContentType="application/json",
Body=b'{"email":"alice@example.com","card":"4111111111111111"}',
)
# Read the stored representation.
stored = s3.get_object(Bucket="my-bucket", Key="customer.json")
print(stored["Body"].read().decode())Install the public stdio server, create an MCP token in the dashboard, and register the command with your MCP client. Full details are in docs/mcp.md.
cargo install --git https://github.com/231self/maskura maskura-mcp
// claude_desktop_config.json, .cursor/mcp.json, or equivalent
{
"mcpServers": {
"maskura": {
"command": "maskura-mcp",
"env": {
"MASKURA_GATEWAY_URL": "https://api.maskura.dev",
"MASKURA_MCP_TOKEN": "maskura_mcp_your_token"
}
}
}
}
# Available tools:
# maskura_put_object, maskura_get_object, maskura_list_objects, maskura_delete_object
# Legacy s4_* tools and s4-mcp remain aliases during the transition.Open source · Apache-2.0
The repository contains Maskura’s self-hosted S3 data plane, content-addressed Wasm runtime, official plugins, CLI, local MCP server, SDKs, and storage adapters. Hosted Maskura adds workspace identity, versioned policy management, audit history, billing, managed operations, and optional managed storage. See the exact boundary or inspect the source .