Credentials isolated
Credentials never enter the agent context.
Give agents data access,not database access.
DataPorch is an open-source data-access layer for AI agents. It connects agents to enterprise data through MCP and APIs, with isolated credentials, validated operations, and bounded execution.
Built for the stack you already use
Protect your database from hallucinations
Let agents ask for data. Keep control of the database.
With DataPorch: write operations stop at the access boundary and permitted reads return bounded results.
Writes stopped
Unsafe operations are blocked before execution.
Queries bounded
Time, rows, and response size stay within limits.
Autonomy preserved
Safe queries run without manual approval.
Credentials exposed
Credentials enter the agent context.
Privileges inherited
Agents inherit database permissions.
Queries unbounded
Queries run without standard limits.
Manual review required
Human review will be the guardrail.
One path from every AI tool to your data.
Connect any MCP-compatible agent. DataPorch controls how every request reaches your databases.
Coding agents, agent frameworks, and any MCP-compatible client connect through DataPorch to PostgreSQL, MySQL, and SQLite.
AI Tools
Coding agents
Claude Code
Codex
Cursor
VS Code
Agent frameworks
Mastra
LangChain
LlamaIndex
CrewAI
Agno
Any MCP client
DataPorch
Controlled data access
Data Sources
Available today
PostgreSQL
MySQL
SQLite
More adapters to come
From install to first answer.
Install DataPorch, connect a source, and ask your first question.
Read the quickstart →DataPorch installs as one local binary. Initialize encrypted state, then start the per-user service.
$ go install github.com/adamraziv/dataporch/cmd/dataporch@latest
✓ DataPorch installed
$ dataporch secrets init
✓ Local keys and encrypted stores initialized
$ dataporch run
✓ Healthy at 127.0.0.1:8080
Import a source over the local admin socket. Your agent receives a source ID, never the connection string.
$ dataporch connections import --id finance --kind postgres
connection string: ••••••••••••••
✓ Source tested and stored
DataPorch lets any agent explore your connected databases safely, within the scope you allow. Simply ask your agent anything about your data.
Which customers have overdue invoices?
Thinking...
data_source.listresolved finance
relational_database.list_tablesinspected public
relational_database.queryreturned 12 rows
12 customers have overdue invoices. Acme Labs has the largest outstanding balance at $18,420, followed by Northstar at $12,700.
›
Help us build DataPorch around real workflows.
We're looking for teams willing to run DataPorch against real agent workflows, share what works and what doesn't, and help shape what we build next.
Popular questions.
What is DataPorch?
DataPorch is an open-source data-access layer for AI agents. It connects agents to enterprise data through MCP and APIs, with isolated credentials, validated operations, and bounded execution. It sits between an agent and a relational database, exposing typed MCP tools for discovery and read-only queries without handing the agent database credentials.
What does the agent see, and what stays inside DataPorch?
The agent sees configured source IDs, schemas, tables, columns, and query results. Connection strings, DSNs, and secret references stay in DataPorch's local state, and imported connector secrets are encrypted at rest.
How does DataPorch bound query time and result size?
DataPorch applies mandatory execution limits, including query time and encoded response size, and limits returned rows by default. Large or truncated results are reported rather than passed through without a bound.
Where does DataPorch run, and how does my agent connect?
The runtime runs locally by default and listens on 127.0.0.1:8080. Codex and Claude Code plugins connect over local stdio and runtime-only local state; direct HTTP MCP clients can use the authenticated local endpoint.
Why shouldn't I give an AI agent direct database access?
Direct access puts database credentials and their privileges in the agent's tool layer. DataPorch keeps credentials and secret references inside the local runtime while the agent works with source IDs and database metadata.
Can an agent change data through DataPorch?
Relational queries use read-only database paths, so write statements are rejected by the read-only boundary. Database permissions still define what the configured identity can read, and read-only SQL does not make arbitrary database functions free of side effects.
Which databases and agent clients are supported today?
DataPorch v0.1 supports PostgreSQL, SQLite, and MySQL. It connects to Codex and Claude Code through local plugins, and it also exposes a local HTTP MCP endpoint for other MCP clients.
How do I install DataPorch and run my first query?
Install the CLI, initialize local state, start the runtime, import a source, and connect your agent. The quickstart walks through the exact commands for a first bounded query.