sanjay3290/ai-skills: 24 Agent Skills That Install Into Six Different Coding Assistants

The gap ai-skills fills between MCP servers and hand-written tool descriptions

An agent that can query a database, post to Telegram or read a Google Doc needs three things: credentials handling, a description of when to invoke the capability, and the invocation code itself. The usual answer is an MCP server per service, which means a separate process, a separate config block, and a separate auth story for each one. The ai-skills repository takes the other route. Each capability ships as a skill folder under skills/, following the Agent Skills Standard that agentskills.io documents, and the client loads the folder directly from its own skills directory. The README lists 24 skills spanning read-only SQL for PostgreSQL, MySQL and Microsoft SQL Server, Google Workspace apps (Docs, Sheets, Slides, Drive, Calendar, Gmail, Chat), Atlassian Jira and Confluence, Azure DevOps, Outline, NotebookLM, ElevenLabs and Google Cloud TTS, image generation through Gemini, the deep-research agent, and delegation skills for Jules, Manus and Grok Build. The intended audience is an engineer who already has an agent open in a terminal and wants it to reach a service without standing up another server. The repository's own framing is that these are lightweight alternatives to full MCP servers, which is accurate for the Workspace set and the honest way to read the whole collection.

One folder per capability, loaded from the client's own skills path

The architecture is deliberately flat. There is no daemon, no proxy and no shared runtime: a skill is a directory the agent reads, and the agent decides when the described capability applies to the task at hand. The README's support table is the load-bearing part of the design, because it shows that the same folder is read from a different path depending on the client. Claude Code looks in ~/.claude/skills/ or .claude/skills/, Gemini CLI in ~/.gemini/skills/ or .gemini/skills/, Cursor in ~/.cursor/skills/ or .cursor/skills/, Codex CLI in ~/.codex/skills/ or .codex/skills/, Goose in ~/.config/goose/skills/ or .goose/skills/, and Google Antigravity in ~/.gemini/antigravity/skills/ or .agent/skills/. The project-level and user-level variants of each path matter: a skill in the project directory travels with the repository, while the home-directory version is available everywhere. Python is the primary language of the repository, which is consistent with skills that shell out to API clients rather than embedding logic in the host agent. The cross-platform claim rests entirely on this path convention plus the standard, not on any adapter layer the repository maintains.

Installing through the Claude Code plugin marketplace or npx skills

Claude Code users have the shortest path, and it needs no extra tooling. Two slash commands from the README: /plugin marketplace add sanjay3290/ai-skills followed by /plugin install ai-skills@ai-skills. That installs all 24 skills as a single plugin, and /plugin marketplace update ai-skills refreshes it later. Everyone else goes through the skills CLI, which the README describes as the package manager for the Agent Skills ecosystem. The listing command is npx skills add sanjay3290/ai-skills --list, which shows all 24 without installing. A single skill installs with npx skills add sanjay3290/ai-skills --skill postgres, and multiple skills chain the flag: --skill postgres --skill mysql --skill mssql. The --all flag takes everything. Agent targeting uses -a, so npx skills add sanjay3290/ai-skills --skill postgres -a claude-code writes into Claude Code's directory specifically, and -a claude-code -a gemini-cli -a cursor fans one skill out to three clients. The wildcard form, --all -a '*', installs every skill into every supported agent. By default the CLI writes to the current project; -g switches to a user-level install available across projects. The README's final code block is truncated mid-command, so the exact -g example is not fully visible in the supplied material.

Read-only is the security model for the SQL skills, and that is the whole point

The three database skills are described in the README with security qualifiers rather than feature lists. Postgres offers read-only queries with what the README calls defense-in-depth security. MySQL provides read-only queries with session-level write protection. MSSQL provides read-only queries with query validation security. Those are three different enforcement mechanisms, which suggests each skill was written against the constraints of its own driver rather than sharing one generic guard. The practical consequence is that if your task is to migrate a schema, insert rows or run a stored procedure, these skills will not do it and are not meant to. The failure mode is a silent one: an agent asked to fix data may attempt a write, get refused, and either report the refusal or route around it by suggesting the SQL for a human to run. That second behaviour is worth watching for, because the skill's refusal does not stop the model from printing a destructive statement into the transcript. The security posture here is a boundary, not a sandbox around the agent.

The Google Workspace skills are Workspace-only, and that gate is not negotiable

Seven skills cover Google's productivity apps: google-chat, google-docs, google-sheets, google-slides, google-drive, google-calendar and gmail. The README states plainly that a Google Workspace account is required and that personal Gmail accounts are not supported, and it attributes this to the skills using the same OAuth infrastructure as the official Google Workspace MCP. Each skill carries standalone OAuth authentication, and tokens are stored cross-platform through keyring. If you are on a personal Gmail address, the entire Workspace block is unavailable to you regardless of how the installation goes, and no amount of CLI flags changes that. For a team already on Workspace, these are the skills most likely to replace an MCP server, since the README positions them explicitly as lightweight alternatives to the full Google Workspace MCP server. The cost of that lightness is that each skill authenticates on its own, which means provisioning credentials per capability rather than once for the whole suite.

Telegram, TTS and the delegation skills cover the long tail

The remaining skills are less about data access and more about routing work. The Telegram skill sends messages, files and alerts through a bot, supports ask-and-wait approvals with inline buttons, handles multi-bot and multi-target routing, and integrates with Claude Code hooks. That hook integration is the notable detail: it makes the skill usable as a notification channel for a running agent rather than only as an outbound send. The elevenlabs and google-tts skills both turn documents into narration or two-host podcasts, with ElevenLabs reading PDF, DOCX, MD and TXT files and Google Cloud TTS offering Neural2, WaveNet and Studio voices across more than 40 languages. The delegation skills push work outward: jules hands async bug fixes, docs, tests and features to Google's Jules agent, manus takes deep research and market analysis, and grok-build dispatches well-specified tasks to xAI's Grok Build CLI headlessly, with the assistant planning, specifying, dispatching and reviewing each diff. The azure-devops skill is the largest single item, listed at 99 tools across 13 domains covering work items, repos, PRs, pipelines, wikis, test plans, security alerts, variable groups, environments and approvals, branch policies and attachments. The atlassian skill supports OAuth 2.1 through an MCP server with an API token fallback, which is the one place in the repository where an MCP server is still in the picture.

Where a dedicated MCP server is still the better answer

The README's own comparison is the useful one: the Workspace skills are framed as lightweight alternatives to the full Google Workspace MCP server maintained under gemini-cli-extensions. The difference in approach is real. An MCP server runs as a process the client connects to, holds one authenticated session, and exposes a tool surface that stays consistent across every client that speaks MCP. A skill folder is read by the agent from a client-specific path and depends on that client implementing the Agent Skills Standard correctly. So a single MCP server gives you one auth flow and one upgrade point for seven Google services, while these skills give you seven independent installs and seven independent auth setups, in exchange for no server process and no MCP configuration. If you are running several agents and want one place to rotate credentials, the MCP route wins. If you want the capability to appear only in the project where it belongs, and you are willing to manage tokens per skill, the skill route is simpler to reason about. Neither is universally correct, and the repository does not pretend otherwise.

Maintenance, the Apache-2.0 licence, and what the release history shows

The repository is not archived, the default branch is main, and the most recent release listed is v1.2.0 dated 2026-07-09, tagged as the 24 cross-platform agent skills release, with the last push timestamp matching the release within a second. That is a single data point, not a maintenance trend, and the supplied material does not include a release cadence or a changelog beyond that entry. The licence is Apache-2.0, which permits commercial use and modification and includes an explicit patent grant; it also requires that you preserve notices and state significant changes if you redistribute. That is a summary of the licence text, not legal advice, and if you are vendoring these skills into a product you should read the LICENSE file in the repository yourself. The upgrade cost depends on install path. Claude Code users get a single command, /plugin marketplace update ai-skills, that refreshes the whole plugin. Anyone using the skills CLI reinstalls per skill or per agent, so a setup with several skills across several clients is several commands. Because skills are directories rather than a running service, an update can change what the agent is told to do without any process restart, which is convenient and also means a skill change is worth reading before you pull it.

Editorial conclusion

Adopt this if you already work inside Claude Code, Cursor, Codex CLI, Gemini CLI, Antigravity or Goose and want packaged skills for read-only SQL, Google Workspace, Telegram alerts or TTS narration without writing the tool descriptions yourself. Do not adopt it if you need write access to production databases, because the postgres, mysql and mssql skills are documented as read-only by design, or if you are on a personal Gmail account, which the README states is not supported for the Workspace skills. Before installing, check two things: whether your client reads skills from the directory listed in the support table, and whether the OAuth credentials each Workspace skill needs are something you can provision. If you only want one capability, install that single skill with npx skills add sanjay3290/ai-skills --skill postgres rather than pulling all 24 into your agent context.

Official sources

1. License: Apache-2.0
2. Project website
3. README
4. Releases
5. sanjay3290/ai-skills on GitHub