dataporch

DataPorch

DataPorch is an open-source data access layer that lets AI agents discover and query relational databases through MCP without exposing database credentials directly to the agent.

PostgreSQL · SQLite · MySQL · MCP · Codex · Claude Code

---

Give agents data access, not database access

Giving an AI agent a database connection also gives its tool layer the credentials and privileges behind that connection.

Without DataPorch

AI Agent
    │
    │ database credentials
    ▼
Your Database

With DataPorch

AI Agent
    │
    │ MCP
    ▼
┌─────────────────────┐
│      DataPorch      │
│                     │
│  Discover schema    │
│  Keep secrets local │
│  Enforce read-only  │
│  Bound execution    │
└─────────────────────┘
    │
    ▼
Your Database

The agent only works with source IDs and database metadata. Credentials, DSNs, and secret references stay inside DataPorch.

---

Protecting your database from hallucinations

An agent investigating an issue decides to change production data.

❌ Direct database access

UPDATE invoices
SET status = 'paid'
WHERE customer_id = 42
RETURNING id, status;

The database executes the write if the supplied credentials allow it.

✅ Through DataPorch

> relational_database.query

UPDATE invoices
SET status = 'paid'
WHERE customer_id = 42
RETURNING id, status;

Result: rejected by the database read-only boundary

DataPorch executes queries through read-only database sessions. The agent cannot turn the query tool into a general-purpose write connection.

---

An agent asks for far more data than it actually needs.

❌ Direct database access

SELECT * FROM events;

The query can continue returning rows until the database, client, network, or agent gives up.

✅ Through DataPorch

> relational_database.query

SELECT * FROM events;

1,000 rows returned
truncated: true

DataPorch applies mandatory execution limits, including query time and encoded response size, with a bounded row limit enabled by default.

Guardrails live between the agent and the database instead of depending on the agent to remember them.

---

How it works

DataPorch gives agents a small set of typed MCP tools.

data_source.list
        ↓
relational_database.list_schemas
        ↓
relational_database.list_tables
        ↓
relational_database.list_columns
        ↓
relational_database.query

The agent starts with configured source IDs. It can then inspect schemas, tables, and columns before it sends a query.

Each database adapter implements the same discovery and query contract. DataPorch does not manage prompts, models, conversations, or agent reasoning. Your agent decides what to ask. DataPorch controls the data access path.

---

Quick start

DataPorch requires Go 1.25 or later when installed with Go. The preferred Homebrew installation is documented after the formula is accepted into homebrew/core; until then, install the published command with Go:

go install github.com/DataPorch/dataporch/cmd/dataporch@latest

The exact-version, reproducible installation becomes available when v0.1.0 is published:

go install github.com/DataPorch/dataporch/cmd/dataporch@v0.1.0

Contributors working from a checkout can use make install. Go installs the binary into GOBIN, or the default Go bin directory, which must be on PATH. Installation does not initialize DataPorch, create keys, register a service, or start a process.

Initialize the per-user local state and start the native background service:

dataporch secrets init
dataporch run
dataporch status

Bare dataporch, dataporch -h, and dataporch --help display the command overview. Use dataporch run -f for the foreground process in a terminal, container, CI job, or external supervisor. Bare invocation no longer starts a long-running process.

The background service uses launchd on macOS and systemd --user on Linux. dataporch status reports the PID, configured address, and the applicable log location. It returns exit code 3 when the service is stopped. DataPorch does not enable login startup; restart refreshes the service definition after an upgrade, while stop unregisters only the generated service definition.

The default state is kept under ~/.dataporch:

~/.dataporch/admin.sock
~/.dataporch/master.key
~/.dataporch/secrets.store
~/.dataporch/connections.store
~/.dataporch/mcp-token.json
~/.dataporch/mcp.sock
~/.dataporch/mcp-control-token
~/.dataporch/logs/

For containers, system services, or an explicitly managed deployment, set the individual DATAPORCH_*_PATH overrides and use foreground mode. Existing state is never removed by stop, upgrades, rollback, or binary removal.

export DATAPORCH_ADMIN_SOCKET_PATH=/run/dataporch/admin.sock
export DATAPORCH_MASTER_KEY_PATH=/etc/dataporch/master.key
export DATAPORCH_SECRETS_STORE_PATH=/var/lib/dataporch/secrets.store
export DATAPORCH_CONNECTIONS_STORE_PATH=/var/lib/dataporch/connections.store
export DATAPORCH_MCP_TOKEN_STORE_PATH=/var/lib/dataporch/mcp-token.json
export DATAPORCH_MCP_SOCKET_PATH=/run/dataporch/mcp.sock
export DATAPORCH_MCP_CONTROL_TOKEN_PATH=/var/lib/dataporch/mcp-control-token
dataporch run -f

DataPorch listens on 127.0.0.1:8080 by default.

Upgrade or roll back an exact Go installation, then refresh the running service:

go install github.com/DataPorch/dataporch/cmd/dataporch@v0.1.0
dataporch restart

dataporch stop

There is no dataporch start command and no dataporch run --foreground alias. Use dataporch run for the native user service and dataporch run -f for foreground execution.

Connect a database

Import a connection through the local administration socket:

Database	Connection format
PostgreSQL	postgres://user:password@host[:port]/database
MySQL	mysql://user:password@host[:port]/database
SQLite	sqlite:///absolute/path/database.db

dataporch connections import --id finance --kind postgres

DataPorch reads the connection string from a hidden terminal prompt. The MCP interface does not receive the connection string. Change --kind and enter the matching connection format for MySQL or SQLite. Import validates and stores the normalized definition without opening or testing the database. The first schema, table, or column discovery call—or a query—opens the source and reports connectivity or permission failures.

Local agent access

dataporch secrets init
dataporch run

Install the Codex or Claude Code plugin after the runtime is running. The plugin launches dataporch mcp over stdio and authenticates through runtime-only local state.

Local plugin users do not run dataporch mcp-token create, copy or export a credential, or edit a shell profile. New terminals, client restarts, and runtime restarts require no repeated token setup.

Direct HTTP MCP clients

Direct HTTP clients can continue using a long-lived bearer token:

dataporch mcp-token create
export DATAPORCH_MCP_TOKEN='dp-...'

The token is for http://127.0.0.1:8080/mcp only. Rotate or revoke it with:

dataporch mcp-token list
dataporch mcp-token rotate
dataporch mcp-token revoke
dataporch mcp-token revoke --yes

Do not place the token in plugin manifests, repository files, shell history, or logs. Hosted OAuth is separate work.

---

Connect an agent

DataPorch includes plugins for Codex and Claude Code. Both plugins connect to the same local MCP service.

Codex

codex plugin marketplace add /absolute/path/to/dataporch
codex plugin add dataporch@dataporch

Claude Code

claude plugin marketplace add /absolute/path/to/dataporch
claude plugin install dataporch@dataporch

See plugins/dataporch/README.md for local stdio setup, updates, removal, direct HTTP usage, and troubleshooting.

---

Compatibility

Component	Support
PostgreSQL	Tested with PostgreSQL 14 and 18
SQLite	Supported
MySQL	MySQL 8.4 LTS
MCP	2026-07-28
Agent clients	Codex and Claude Code
Go	1.25+

---

Security

DataPorch keeps the database access boundary outside the agent.

• Local by default: The MCP service listens on 127.0.0.1:8080.
• Authenticated MCP: MCP requests require a Bearer token.
• Local credentials: Agents receive source IDs instead of credentials, DSNs, or secret references.
• Encrypted secrets: DataPorch encrypts stored connector credentials.
• Read-only queries: Database adapters execute queries through read-only paths.
• Bounded execution: DataPorch limits query time and encoded response size. It limits returned rows by default.

Database permissions still define what the configured database identity can read. Read-only SQL does not make arbitrary database functions free of external side effects.

Keep the default local network boundary unless you add secure transport and authorization for remote access.

---

Development

Run the local quality gate:

make check

Run database integration tests:

make test-integration-postgres
make test-integration-sqlite
make test-integration-mysql

PostgreSQL and MySQL integration tests require their corresponding test connection strings.

---

License

DataPorch is licensed under Apache-2.0.