Find a skill, connect your agent and run it
Find a skill, connect Claude Code, Codex, Cursor or Gemini CLI, run it, and manage allowances, keys and devices.
This guide covers community and hosted skills. Community skills install locally from their listed sources without a SkillGild session or account. Hosted skills also need sign-in and an MCP connection; their monthly allowances are shown on each listing.
#Find a skill
Before choosing a platform, compare Skillry's downloadable skills with SkillGild's hosted tools to understand what you install and where the work runs.
#On the website
| Page | What it offers |
|---|---|
/skills |
The full catalog. Search, filter by category and access (free or paid), switch between grid and list, and sort by Recommended, Name, Price or Most free runs. Filters are kept in the URL, so you can share a filtered view. |
/categories/{slug} |
Every skill in one category. |
/collections, /collections/{slug} |
Curated groups of skills picked by the SkillGild team. |
/skills/{slug} |
One skill: description, access tier, free-run allowance, its public input schema, the install command for each agent, and run snippets for the CLI, REST, TypeScript and Python. |
Press Ctrl+K (⌘+K on macOS) or / anywhere on the site to open quick search.
#From the terminal
skillgild search "plot"academic-plotting free Academic Plotting — Publication-quality figures for ML papers: picks the chart type…#From your agent
Once the MCP server is connected, your agent can call skillgild_search_skills itself. The tool returns public metadata, access tier, version and runtime type; never the private implementation.
Search matches a case-insensitive substring of a skill's name, slug or description, and returns up to 100 results.
#Connect your agent
You set this up once per machine. After that, every skill is available to every connected agent.
#1. Install the CLI
The skillgild CLI is both the sign-in tool and the local MCP server your agent talks to. It is a single binary with no dependencies.
macOS and Linux:
curl -fsSL https://raw.githubusercontent.com/SkillGild/skillgild/main/scripts/install-cli.sh | shWindows (PowerShell):
irm https://raw.githubusercontent.com/SkillGild/skillgild/main/scripts/install-cli.ps1 | iexThe installers download the release for your platform from GitHub Releases, verify its SHA-256 against the release's checksums.txt, and put skillgild in ~/.local/bin (macOS and Linux) or %LOCALAPPDATA%\Programs\skillgild (Windows). Set SKILLGILD_VERSION to pin a release and SKILLGILD_INSTALL_DIR to choose the directory. You can also download an archive from the releases page by hand, or build from source with Go 1.26 or later:
cd skillgild/backend
go install ./cmd/skillgildCheck the installation with skillgild version. Running the installer again updates to the latest release.
#2. Sign in from the terminal
skillgild setupSkillGild detects your installed agents and asks you to choose if it finds several. It checks your existing sign-in, opens the browser when authorization is needed, and connects the selected agent. Approve the displayed device code in your browser, then restart or reload your agent. Your selection is remembered for future installs.
To choose explicitly:
skillgild setup --agent claude-codeUse codex, cursor, gemini-cli, antigravity or antigravity-cli for another
supported agent. Claude Code, Codex and Gemini CLI must be on PATH. If you already
installed an older SkillGild CLI, run skillgild update first.
You can also skip a separate setup step: skillgild install <slug> completes setup
for hosted skills automatically. Community skills do not require sign-in or MCP.
Run skillgild status to check your connected account and credential storage.
skillgild login remains available to sign in separately or replace a credential.
#3. Register the MCP server with your agent
Automatic setup already connects supported agents. The manual alternative below is optional. For a client without automatic setup, run skillgild login first. Find the absolute path of the binary with command -v skillgild, then register it. Apps launched from an IDE may not inherit your terminal's PATH, so use the absolute path.
| Agent | Command or config |
|---|---|
| Claude Code | claude mcp add --scope user skillgild -- /absolute/path/to/skillgild mcp |
| Claude Desktop | Download skillgild.mcpb and open it. Claude Desktop installs the server and asks for an API key, which it stores as a secret; you can skip steps 1 and 2 |
| Codex CLI | codex mcp add skillgild -- /absolute/path/to/skillgild mcp |
| Gemini CLI | gemini mcp add --scope user skillgild /absolute/path/to/skillgild mcp |
| Cursor | Add the JSON below to ~/.cursor/mcp.json, or to .cursor/mcp.json in a project |
| Any MCP client | Launch /absolute/path/to/skillgild mcp as a local stdio server |
{
"mcpServers": {
"skillgild": {
"type": "stdio",
"command": "/absolute/path/to/skillgild",
"args": ["mcp"]
}
}
}After setup, check that the skillgild server shows as connected in your agent's MCP settings. You can register the server before signing in: its search tool works without a credential, and the other tools answer with a request to run skillgild login. Once you have signed in, the running server picks the credential up on the next call; no restart is needed.
#4. Install a skill
skillgild install academic-plotting --agent claude-codeInstalled hosted wrapper at /Users/you/.claude/skills/academic-plotting/SKILL.md
Open your agent and ask it to use Academic Plotting.| Flag | Effect |
|---|---|
--agent |
auto (default for hosted skills), codex, claude-code, cursor, gemini-cli, antigravity or antigravity-cli. shared installs files without setup. Picks the skills folder; see Core concepts. |
--target /absolute/path |
Write under a custom skills root without changing login or MCP configuration |
--force |
Replace an existing wrapper |
--no-setup |
Install files only; handle sign-in and MCP registration separately |
For a hosted skill, installing writes a small SKILL.md wrapper rather than the private server implementation. Community listings instead install the full source folder (from the listed repository or its available archive), and run locally without MCP or hosted quota. The hosted wrapper tells your agent when to use the skill and how to start it:
---
name: academic-plotting
description: "Publication-quality figures for ML papers: picks the chart type for your results and writes the matplotlib or seaborn code, or plans an architecture diagram from your method section."
---
# Academic Plotting
Free skill. This skill runs in this coding agent with your own model; SkillGild runs its private
server tools. This local file is only an invocation guide.
When the user's task matches this skill:
1. Call the SkillGild MCP tool `skillgild_start_session` with skill ID `<skill uuid>`. …
2. Follow those instructions with your normal file and shell tools. …
3. Call `skillgild_end_session` when the task is done.
…#5. Use it
Ask your agent for the task in plain language. When the task matches the skill's description, the agent starts a session, follows the instructions SkillGild sends for that session, calls the skill's server tools when the instructions ask for them, and ends the session. It applies changes with its own file tools; the wrapper tells it to show a diff and ask before destructive changes.
Your agent decides from the skill's description whether a task matches. If you want a particular skill, name it: "Use the SkillGild academic-plotting skill to …".
You can also drive a session from the terminal:
skillgild session academic-plotting
skillgild call <session-id> validate_palette --input '{"colors":{"ink":"#000000","paper":"#FFFFFF"},"pairs":[{"foreground":"ink","background":"paper","minimum_ratio":4.5}]}'
skillgild end <session-id>#Configure a skill's input
Each server tool declares a public input schema. You can see the tools and their fields on the skill page, and a session lists them when it starts. Your agent sends a JSON object that matches the tool's schema:
- include every key listed in
required; - use the declared type for each property.
Tool input is limited to 16 MiB per call, with smaller limits possible in each tool's schema. Legacy prompt runs have a separate 64 KiB default input limit.
#Access, pricing and quotas
| Skill type | What you get |
|---|---|
| Free hosted | The monthly allowance shown on each skill listing. Resets at 00:00 UTC on the 1st. |
| Community | Local execution under the listed source license; no SkillGild hosted run allowance. |
| Paid | Needs an active entitlement, held either by you or by an organization you belong to. The entitlement may set a monthly run limit. |
| Pro | Included with an active SkillGild Pro plan, with a monthly run limit per skill shown at checkout. Pro skills are not sold one by one. |
Every run response includes your current usage, so agents and scripts can tell how much allowance is left:
"usage": { "used": 3, "limit": 25, "reset_at": "2026-11-01T00:00:00Z", "access_tier": "free" }Platform-wide safety limits also apply:
| Limit | Default |
|---|---|
| Runs per minute, per user per skill version | 20 |
| Runs per minute, per API key | 10 |
| Runs per minute, per IP address | 60 |
| Runs in progress at once, per user | 5 |
| Run timeout | 90 seconds |
| Legacy prompt input / output | 64 KiB / 1 MiB by default |
| Hybrid server tool input | 16 MiB per call, subject to field-level schema limits |
#Manage your access
Everything lives in Account:
- Agent connections → devices lists each connected device with its last use and credential expiry. Revoking a device disables its key immediately.
skillgild logoutdoes the same from the terminal and removes the local credential. - Agent connections → API keys are for scripts and SDKs. You can have 5 active keys in total, including CLI device keys. A key's secret is shown once; keys expire after 90 days by default.
- Organizations & webhooks → webhooks send signed
execution.completedandexecution.failedevents to your HTTPS endpoint. See Webhooks. - Organizations & webhooks → organizations let a team share paid-skill access. Create a workspace, invite members with one-time tokens, and transfer ownership.
#Refer people and earn commission
Every account is an affiliate account. Account → Referrals and the affiliate page show your 8-character referral code and a sign-up link that carries it. Anyone who signs up with the link, or types your code on their account page within 30 days of joining, is credited to you.
- Commission. Each verified Pro payment by someone you referred credits you the program rate for that person's payment number: by default 25% of each of their first 10 payments, then 10% of every payment after that, in the currency they paid in. The live rates are shown on the affiliate page.
- Hold and reversal. A commission is pending for 30 days after the payment, then becomes available. A refund or lost dispute reverses it; if it was already paid out, the amount is deducted from your next payout.
- Payouts. Add a payout method (PayPal, bank transfer or UPI) and request a payout once your available balance reaches the equivalent of $20. The SkillGild team sends it within a few business days and notifies you in your account inbox. USD equivalents use SkillGild's own price list to compare currencies; payouts are made in the currency each commission was earned in.
- What you see. Referred members appear as first name and last initial, with the date they joined, how they were credited and how many of their payments earned you commission. Email addresses are never shown.
#Report a problem
Signed-in users can report a skill from its page: pick a reason and add optional details of up to 2,000 characters. Reports are reviewed by the SkillGild team, and you can follow them under Account → Activity & support. You can appeal a closed decision within 30 days, with an explanation of 40 to 4,000 characters. For anything else, open a support case from the same section.
#Privacy at a glance
- Run inputs and outputs are not stored in execution records.
- Abuse review stores only HMAC fingerprints of inputs and coarse signals, never the request text.
- If you send an
Idempotency-Key, the response is kept encrypted for 24 hours so a retry can replay it. - Catalog search analytics count searches in aggregate, with no query text, account or IP address.