SafeSelect

Agents can look. They cannot mutate.

Read-only PostgreSQL & MongoDB access for coding agents.

Debug with real database context, without exposing write tools.
Local, project-scoped policy—even when your existing credentials allow writes.

Get started → · Demo · Agents · Security · Docs · Website

What you can do

• Inspect PostgreSQL — discover tables, indexes and query plans; read bounded rows.
• Explore MongoDB — discover collections, infer sampled schemas and run bounded reads.
• Reuse your connections — import from DBeaver, Docker Compose or MongoDB Compass.
• Connect your coding agent — install a project/environment-pinned MCP entry.
• Keep control — local stdio, project policy, external secrets and audit metadata.

Roadmap

See it in action

Complete onboarding: from Homebrew to a protected agent

SafeSelect onboarding: Homebrew, DBeaver SSH import, Keychain and OpenCode

Install SafeSelect from Homebrew, import an SSH-backed DBeaver connection, keep the password in macOS Keychain, install the OpenCode integration, and see the agent read a paid order while its DELETE attempt is rejected. Focused agent and backend clips remain in the complete demo gallery.

Linux onboarding: Watch current-source Compass → password-authenticated bastion → check, using environment references for both passwords, with hidden input and no secrets stored in TOML.

Quick Start

Run setup from your repository root. You will need a PostgreSQL or MongoDB connection and a Java 17+ runtime for database commands.

1. Install

On macOS with Homebrew:

brew install antonillos/tap/safeselect

curl -fsSL https://raw.githubusercontent.com/antonillos/safeselect/main/packaging/install/install-release.sh | sh

asdf plugin add safeselect https://github.com/antonillos/asdf-safeselect.git
asdf install safeselect latest
SAFESELECT_VERSION="$(asdf latest safeselect | sed -n '$p')"
asdf set -u safeselect "${SAFESELECT_VERSION}"
asdf reshim safeselect "${SAFESELECT_VERSION}"

SafeSelect uses any available Java 17+ runtime rather than requiring a specific package-manager formula. If Java is missing or too old, install or select a Java 17+ runtime before running database commands. On macOS with Homebrew, you can install one with brew install openjdk@17.

2. Import one connection source

Choose the source you already use; you do not need to run all three:

Use your actual export or Compass directory. Follow the importer's next steps for driver and password setup before continuing. Keep secrets out of project files. For an SSH-backed walkthrough, see DBeaver → Codex.

3. Check and connect your agent

# Check configured environments; this can open SSH tunnels and contact databases.
safeselect check

# Install for OpenCode (replace with codex for OpenAI Codex).
safeselect agent install opencode

# Inspect the installed MCP entry and its configuration location.
safeselect agent status

Multiple environments? Use safeselect check --environment <name> to avoid checking unrelated databases, and add --environment <name> to agent install to select the intended target. Installation infers the name only when there is one environment.

Open or restart your agent and approve the MCP server if prompted. Installation uses user scope by default; see supported agents and the client setup guide for project scope and client-specific steps.

4. Try a first read

Ask your agent:

Use SafeSelect to identify the connected backend and discover its available tables or collections. Describe one, then stop before querying row or document contents. Follow next suggestions only within that discovery-only scope.

Success looks like: the agent calls database_info, uses the matching schema-discovery tools, and reports the structure it found. No write tools or database passwords are needed in the conversation.

Stuck? Run safeselect doctor --environment <name> for concise diagnostics (this can contact the database), then follow the reported next step. Do not relax policy to get past a rejection. See agent recovery.

{
  "mcpServers": {
    "safeselect-myapp-testing": {
      "command": "safeselect",
      "args": ["serve", "--project", "/path/to/myapp", "--environment", "testing"]
    }
  }
}

Documentation

Where It Helps

• Debug an application against realistic data without exposing mutation tools.
• Let an agent inspect schemas, indexes, query plans, and bounded rows during development.
• Explore MongoDB collections through bounded reads and sampled schema inference.
• Reuse existing DBeaver, Docker Compose, or MongoDB Compass connections.
• Give coding agents database context while keeping policy, limits, secrets, and audit under your control.

Why SafeSelect?

SafeSelect is intentionally narrower than general-purpose database MCP servers. It is not a tool builder, SQL workbench, or remote database gateway. It is a local safety boundary for agents that need database visibility, not database power.

What Makes It Different?

The combination matters: PostgreSQL and MongoDB inspection, a fixed database read surface, local stdio, project policy, connection import and reproducible security evidence. Read-only modes and layered controls also exist in other projects; they are not exclusive to SafeSelect.

See the dated comparison for DBHub, MongoDB MCP Server, Postgres MCP Pro and SchemaBrain—including when each is a better fit.

Agents can look, but they cannot mutate through SafeSelect's database tools. This boundary does not cover a shell, another MCP server or direct credentials also available to the agent. Use least-privilege database users and review the threat model and limits.

Backend Support

Architecture

The agent talks to SafeSelect through MCP stdio. SafeSelect enforces policy in Rust, stores secrets outside project files, and reaches databases through an embedded Java sidecar: JDBC for SQL backends and the MongoDB driver for MongoDB. The Rust to Java channel is JSON-lines over stdin/stdout: no sockets, no open ports.

Guided MCP Context

Clients that support MCP prompts can invoke read_only_database_debugging for a safe investigation checklist. Clients can also read safeselect://guide/read-only-database-debugging for the same static workflow and boundary notes. Neither capability exposes database data, credentials, or write access; use the database tools below for discovery and bounded reads.

Agent Workflow

Agents should use SafeSelect in this order:

1. database_info
2. list_tables then describe_table; inspect list_table_indexes or bounded statistics when useful for SQL
3. list_databases, list_collections, then discover_document_schema for NoSQL
4. select / explain, or the bounded MongoDB read tool that matches the task
5. check, connect, or reconnect when connectivity is stale

MCP check verifies the backend used by the current session (SELECT 1 for PostgreSQL, a database ping for MongoDB). An existing usable connection or tunnel does not require a separate successful bastion probe. Sidecar startup and response reads have deadlines; stalled or malformed responses invalidate the sidecar without closing the MCP session. For a stale connection, call reconnect once, then check, without reopening the client. A startup or configuration failure still requires fixing its cause before retrying; do not loop on reconnection.

Agents must discover relation or collection structure before querying unfamiliar data and use each discovery response's next_suggestion instead of guessing column or field names. SQL descriptions are catalog metadata; MongoDB schemas are inferred from a bounded, non-exhaustive sample.

MongoDB query documents must remain complete nested JSON values. Clients that flatten nested tool arguments can pass filter, projection, and sort as JSON-encoded object strings and pipeline as a JSON-encoded array string. redact_fields also accepts a JSON-encoded string array. Flattened keys are rejected so a lost filter or redaction can never become a less constrained fallback.

MongoDB server-side JavaScript is never available: $where, $function, and $accumulator are rejected recursively in filters, projections, sorts, and aggregation pipelines before the MongoDB driver receives them. When rejected, rebuild the request with declarative MQL operators; SafeSelect has no setting that enables JavaScript.

Query responses include row_count, byte_count, elapsed_ms, and a human-readable elapsed value so agents can reason about result size and latency.

Every MCP success and error includes one contextual next_suggestion. Agents should follow that single safe action, never blindly repeat an invalid request, and stop when the suggestion is terminal. For clients that only show an MCP error summary, SafeSelect also includes the trusted next suggestion in that summary without exposing database-derived detail.

Security Model

• Fail closed: security violations terminate the MCP process.
• Read only: SQL allows SELECT, EXPLAIN, and WITH; NoSQL backends allow discovery and read-only document reads.
• No server-side JavaScript: MongoDB $where, $function, and $accumulator are rejected in Rust and again in the Java sidecar.
• Scoped access: schemas, relations, databases, and collections can be allowed or denied.
• Hard limits: row count, result bytes, and timeouts are enforced; MongoDB read commands receive the same timeout as maxTimeMS.
• Secret isolation: passwords live in macOS Keychain or environment variables, never in project config.
• Driver verification: JDBC drivers are checked by SHA-256 before use.
• Audit trail: query text is hashed before being recorded; the current session exposes bounded audit metadata through audit_status and audit_recent.

Deliberate Limits

• SafeSelect does not expose database writes, migrations, administration, or arbitrary command execution.
• PostgreSQL and MongoDB are the supported backends today; broad connector count is not the goal.
• MCP transport is local stdio. SafeSelect is not a remote database gateway.
• MongoDB schema discovery is sampled and bounded, not an exhaustive schema guarantee.
• SafeSelect complements database-native least privilege; it does not replace it.

MCP Tools

get_maintenance_diagnostics supports PostgreSQL 15, 16, 17, and 18.

When no .safeselect/ directory exists, safeselect serve scans for PostgreSQL Compose services. If found, it enters setup mode automatically: it imports them, writes project configuration, and starts a setup-only MCP server. Otherwise it prints setup instructions and exits. An existing but empty or invalid configuration is rejected, not replaced by setup.

CLI Essentials

Visual command gallery

The CLI is easier to scan by task than as one long list. The web command gallery uses the same synthetic-demo captures. query is kept separate because it is a direct SQL workflow; agents should normally discover schema first through MCP tools.

safeselect import-dbeaver ~/Downloads/connections.zip

safeselect import-compose --path .

safeselect import-compass --path ~/.config/MongoDB Compass

safeselect config show --project demo --environment postgres

safeselect driver list

safeselect agent detect

safeselect check

safeselect doctor

safeselect posture --strict

safeselect serve

safeselect connect

safeselect disconnect

safeselect reconnect

safeselect query --sql "SELECT order_id, status, subtotal FROM public.demo_orders WHERE status = 'paid' LIMIT 3"

list_tables({"schema":"public"})

describe_table({"schema":"public","table":"demo_orders"})

get_maintenance_diagnostics({"schema":"public"})

list_databases()

list_collections({"database":"safeselect_demo"})

discover_document_schema({"database":"safeselect_demo","collection":"orders","sample_size":5})

find_documents({"database":"safeselect_demo","collection":"orders","filter":{"status":"paid"},"limit":3})

Use safeselect --help or a command-specific --help for the full CLI.

Uninstall checks both release-installer and Cargo binary locations. MongoDB Compass imports support SSH-tunneled mongodb+srv:// connections by resolving the SRV target and rewriting the local endpoint with the required TLS and direct-connection options.

Configuration

Global state lives in ~/.config/safeselect/ by default. Project policy lives in .safeselect/ at the repository root:

<repo-root>/
└── .safeselect/
    ├── project.toml
    └── environments/
        └── <env>.toml

SafeSelect walks upward from the current directory to find .safeselect/. Use --project <path> when an agent or script should target a specific repository.

Convention before configuration

From inside a configured repository, commands infer the project from the nearest .safeselect/ directory and infer the environment when exactly one environments/*.toml file exists. For example, safeselect serve, safeselect check, safeselect query --sql "SELECT 1", and safeselect config show need no project or environment flags in a single-environment project. --project and --environment remain available for scripts, other working directories, and deliberate selection. If multiple environments exist, commands that act on one fail rather than guess and tell you to pass --environment <name>.

The defaults differ by operation:

• serve, query, connect, disconnect, config show, and password commands require one explicit or uniquely inferred environment. With no environments they fail; serve has the separate first-run behavior above.
• check, doctor, posture, reconnect, and config validate inspect or process all environments when the flag is omitted. Checks are not offline: they can resolve secrets, establish SSH tunnels, and contact databases. Use --environment <name> to avoid touching unrelated or production environments.
• query supports JDBC environments only and still needs SQL through --sql or stdin (interactive stdin waits for EOF). Selecting a MongoDB environment does not turn it into a MongoDB query command.
• CLI connect and disconnect operate on a new, temporary JDBC sidecar and shut it down before exiting. They do not control an already-running MCP session; use that session's MCP connection tools instead.
• Password commands modify local Keychain/configuration, not the database password itself. config set-ssh-password switches SSH authentication to password and clears the configured identity-file reference.

Password references on all platforms

Use OpenCode-style {env:NAME} references for database and SSH passwords on macOS, Linux and WSL. These commands save only the variable name; they do not read its value or store the password:

safeselect config set-password --environment staging --password '{env:DB_PASSWORD}'
safeselect config set-ssh-password --environment staging --password '{env:BASTION_PASSWORD}'

Enter the values without putting passwords in shell history (Bash):

read -rsp 'Database password: ' DB_PASSWORD; echo; export DB_PASSWORD
read -rsp 'Bastion password: ' BASTION_PASSWORD; echo; export BASTION_PASSWORD
safeselect check --environment staging

The variables must be exported in the process that launches SafeSelect or the MCP client. A missing or empty required variable fails closed. No Python helper, .env auto-loading, shell evaluation, {file:...} reading or general TOML interpolation is performed. Password-based SSH tunnels still require sshpass and do not fall back to SSH-agent/key authentication. The saved configuration uses the existing database.secret.source = "env" and variable fields, and SSH secret_variable; no new password field is needed.

On macOS, secure password prompts also accept {env:NAME}; ordinary passwords continue to use Keychain. Linux/WSL SSH import prompts accept either NAME or {env:NAME}. Use config set-password after import to select a short database variable name instead of the generated default. config_set_password through MCP accepts the same reference; prefer references rather than sending raw passwords through an agent.

References must occupy the entire input and use a valid shell variable name. Imported credentials are opaque passwords, not expressions. If a literal macOS password happens to look like a reference, use the secure literal prompt:

safeselect config set-password --environment staging --literal-password
# or: safeselect config set-ssh-password --environment staging --literal-password

Avoid supplying real literal passwords with --password: command arguments can appear in shell history and process listings. Existing Keychain and environment references remain valid. Custom variable names are preserved when renaming an environment; changing a source does not delete the old Keychain entry.

SSH passwords on Linux and WSL

SSH password imports use macOS Keychain only on macOS. On Linux/WSL, the Compass and DBeaver import prompts ask for an environment variable name, not the password. The generated default uses SAFESELECT_SSH_PASSWORD_<PROJECT_SHA256>_<HEX>, where <PROJECT_SHA256> namespaces the canonical project directory and <HEX> encodes the complete SSH account's UTF-8 bytes (project, environment and /ssh) without lossy normalization. Distinct account names therefore have distinct defaults; existing and explicitly selected references stay unchanged. SafeSelect saves that reference as secret_variable in the SSH configuration (including shared bastions), never the password itself:

[ssh]
enabled = true
auth_type = "PASSWORD"
secret_variable = "SAFESELECT_SSH_PASSWORD_DEMO"

Export the password in the shell that launches SafeSelect or the MCP client. For Bash in WSL, this avoids putting the password in shell history:

read -rsp 'SSH password: ' SAFESELECT_SSH_PASSWORD_DEMO; echo
export SAFESELECT_SSH_PASSWORD_DEMO

Password-based tunnels also require sshpass. Missing or empty secrets fail closed; SafeSelect does not fall back to a different secret source. Do not set both secret_account (macOS Keychain) and secret_variable. For an existing Linux/WSL environment, run safeselect config set-ssh-password --environment <env> without --password to configure an environment reference, then export the variable it prints. The command cannot export into its parent shell. Restart a running MCP client after changing its environment.

Newly imported database passwords use SAFESELECT_PASSWORD_<PROJECT_SHA256>_<ENV_HEX> on Linux/WSL. The project component is the SHA-256 of the canonical project directory (without exposing its path); the environment component encodes the environment name's UTF-8 bytes. This produces shell-valid, distinct references even for names like qa.eu and qa-eu. The printed setup guidance uses the exact generated reference; existing saved references are not renamed, including when moving a checkout. Imports defer automatic verification while a required database or SSH variable is unset or empty. Export that variable separately before checking the imported connection; credentials from the export are removed from the saved MongoDB URI rather than written to TOML.

Generated MCP entries deliberately pin both project and environment. Keep those arguments in client configuration and unattended scripts: inference is a CLI convenience, not a persistent default environment.

Supported Agents

SafeSelect never silently falls back to a broader scope. In particular, --local for Windsurf fails with a clear correction because Windsurf does not document a project-scoped MCP configuration.

Build From Source

# Installs makevn through Homebrew or asdf only when it is missing.
./install.sh --install-makevn
"$HOME/.local/bin/safeselect" --version

Requirements: Rust 1.85+, Java 17+, and makevn 0.1.14+. The bootstrap requires Homebrew or asdf; otherwise install makevn first. sshpass is optional for password-based SSH tunnels. Add ~/.local/bin to your PATH before invoking safeselect without its full path.

The installer runs makevn doctor --compact, then makevn init --force to refresh generated initialization while preserving local configuration, before makevn test package. A makevn failure stops the build without replacing the installed binary.

Documentation

• Installation guide
• AI agent integration
• On-demand Codex code review
• Security model
• Security Proof
• Security test suite
• Security policy
• Distribution
• Changelog

Release notes are generated from CHANGELOG.md.

Ecosystem

License

MIT - see LICENSE.

When using an Azure Bastion tunnel, open it before running SafeSelect and set its SSH host and port to the local listener (for example 127.0.0.1:2222). The SSH connectivity diagnostic reports the configured endpoint and the actual TCP connection or name resolution error; TCP success does not validate SSH credentials. Run the Azure CLI tunnel in the same WSL distribution as SafeSelect, or ensure that its Windows listener is accessible from WSL. On SSH connectivity failure, SafeSelect also prints a safely quoted TCP probe command for the configured endpoint, using Bash and timeout on Linux/WSL; neither nc nor telnet is required. Run it from the same terminal environment as SafeSelect. A successful TCP probe does not verify SSH or database authentication.