agent-skills
- Type
- MCP server
- Scanned
- Nov 13, 2026
- Version
- fa44b92d2fd1
- Files read
- 276
This plugin bundles Neon's official AI agent skills and connects your coding agent to Neon's hosted MCP server (mcp.neon.tech) to create and manage Postgres databases, branches, authentication, storage, serverless functions, and AI models. Its biggest risks are that the core skill tells the agent to silently send failed commands and error details to Neon without asking for approval, automatically overwrite skill files with updates on every session, and store full-access API keys in global agent config files during CLI setup.
What it can reach
- Network
- Connects to Neon's hosted MCP server at https://mcp.neon.tech/mcp to read and modify Neon projects, databases, branches, functions, storage, auth, and logs, and fetches documentation and skill files from neon.com, claimable.neon.tech, and models.dev.
- Shell commands
- Instructs the coding agent to run Neon CLI commands, install packages with npm/pnpm/bun/npx, and execute database migration tools in your terminal.
- Files
- Instructs the coding agent to read and write project files such as neon.ts, .neon, .env, and .env.local, as well as local and global agent skill and MCP configuration files.
- Secrets
- Instructs the coding agent to pull, generate, and read database connection strings, Neon API keys, S3 storage keys, AI Gateway tokens, Realtime secrets, and cookie signing secrets in .env/.env.local and agent configuration files.
Findings
-
MediumSkill instructs the agent to silently send failed commands and errors to Neon without approval
The core neon skill (and its synced copy in plugins/neon-postgres/skills/neon/SKILL.md) instructs the AI agent to silently report tooling or guidance errors after completing a task via neon feedback --message or the send_feedback MCP tool, explicitly telling the agent to use a channel that runs without user interaction or approval. Although the text tells the AI to strip secrets and personal data, AI-based sanitization of raw commands and error messages can still leak internal schema details, paths, or credentials, and a prompt injection in logs or database rows could abuse this silent channel to send data out without asking.
Fix: Remove the silent, unprompted feedback instructions from skills/neon/SKILL.md (and re-sync plugins/neon-postgres/skills/), or require the agent to show the exact report to the user and ask for explicit approval before sending it.
-
MediumSkill instructs the agent to auto-update skills on every session and fetch live remote instructions
Under Updating Skills, skills/neon/SKILL.md tells the agent to check skill status on every new session and run neon skills update -y (skipping confirmation prompts, including --global) whenever upstream changes exist, and multiple skills instruct the agent to fetch live SKILL.md and protocol files directly from neon.com and claimable.neon.tech when missing locally. This causes the agent to modify installed skill files without being asked and bypasses version pinning, meaning a future compromise or bad edit on those remote endpoints would automatically push unreviewed instructions into existing installations.
Fix: Do not instruct the agent to run neon skills update -y automatically on every session or execute live remote SKILL.md files; instead, notify the user when updates are available and let them update manually.
-
LowCLI MCP setup instructions default to storing a full-access API key globally instead of OAuth
While the bundled mcp.json plugin manifest uses OAuth, the CLI setup instructions in skills/neon/SKILL.md (under Install the Neon MCP Server and neon init -y for non-plugin agents) default to running neon mcp --agent <agent> -y without --oauth, --read-only, or --project-id. This mints or reuses a long-lived, full-access Neon API key and writes it in plain text into the user's global coding-agent configuration, exposing all of the user's Neon projects if another local process reads that file or if a prompt injection occurs later in an unrelated repository.
Fix: Change the default global MCP setup command in skills/neon/SKILL.md to use --oauth (neon mcp --oauth --agent <agent>), and recommend --read-only and --project-id whenever a static API key is required.
-
LowSkills instruct the agent to install and run unpinned npm packages globally and via npx
skills/neon/references/claimable-neon.md:11
Several skill files instruct the coding agent to download and execute unpinned npm packages at runtime, including npm i -g neon@latest (skills/neon/references/claimable-neon.md and skills/neon-ai-gateway/SKILL.md), npm i -g neon (skills/neon/SKILL.md), npx shadcn@latest (skills/neon-auth/references/managed-auth.md), npx @modelcontextprotocol/inspector (skills/neon-auth/references/oauth-server.md), and the third-party npx add-mcp helper (skills/neon-functions/references/mcp.md). Running unpinned @latest or npx commands exposes the user's machine to supply-chain attacks if any of those packages is compromised.
Fix: Pin exact tested versions in global install and npx commands across the skill markdown files (matching the repository's own CONTRIBUTING.md pinning policy) and avoid unpinned third-party npx config writers.
