Skip to main content
Glama
auto-skill

auto-skill-connector

by auto-skill
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

Auto-Skill

Just-in-time skill routing for coding agents. Auto-Skill watches what a task needs, hands the agent a verified, relevant technique for it, and gets out of the way. No skill installation, no directory browsing, no maintaining a local skills folder.

Works with Claude Code, Codex CLI, Cursor, and GitHub Copilot.

Does it actually help?

We measured it rather than assumed it. Routing a skill into context can just as easily override a model's own reasoning as inform it — and our first pass at this got that wrong, measurably. After fixing that (the wrapper now says "if you don't already know a correct way to do this, use this technique — it should let you do this better than you could on your own," not "apply this now") and building a judge-graded benchmark to check the result:

condition

baseline pass rate

with Auto-Skill

lift

well-matched task + genuinely relevant skill

50%

90%

+40 points

general, unfiltered task mix

18–59%

24–65%

+6 to +12 points

Both numbers are real, from the same investigation, on a local weak model (llama3.2:3b) — not cherry-picked, not from a paid frontier model that already knows most things unaided. The first row is what Auto-Skill's mechanism does when the pieces line up (a real capability gap, a routed skill that's actually right for it); the second is closer to what a representative task mix looks like. Read both, not just the good one: backend/bench/WEAK_MODEL_BENCHMARK_REPORT_20260803.md has the full methodology, every regression found along the way, and how each was fixed.

Related MCP server: @skill-hub/mcp-server

Quickstart

Requires Python 3.10+ and git.

git clone https://github.com/auto-skill/auto-skill-connector
cd auto-skill-connector
python -m venv .venv
.venv\Scripts\python -m pip install -e ".[dev]"
auto-skill login
auto-skill route "build a React landing page with Tailwind"
backend: self-hosted-route
route: skill
tier: full
skill plan (coding):
  policy: ponytail
  primary: frontend-design

next action: apply the returned skill plan in-turn; no skill installation is required

Installing the connector does not install skills — catalog skills stay in Auto-Skill's hosted database and are retrieved per task. Free accounts get 100 routes/month; the CLI session is a 30-day sliding token (refreshes on use).

How it works, briefly

  1. A task gets classified (family + action), not just keyword-matched.

  2. Auto-Skill composes an ordered plan: an always-on team/task-family policy, plus one primary specialist for this specific task.

  3. Only high-confidence, content-hash-verified, risk-0 content is delivered as active guidance (full tier). Ambiguous matches come back as hints, not instructions. Content tied to one specific repo or to an agent's own sandbox filesystem is excluded — it wouldn't generalize to your machine.

  4. Your own project/team instructions always win over a routed skill if they conflict.

Full pipeline details, response tiers, and current launch scope: docs/ROUTING.md.

Using it

Three ways to get routing, in increasing order of automation:

Mode

What it does

Setup

CLI, on demand

You call auto-skill route "<task>" explicitly

Nothing extra

MCP, proactive

A connected agent calls route_task itself near the start of substantial work

Add the MCP server — see docs/MCP.md

Auto Mode, opt-in hook

Every eligible prompt is preflighted locally and routed automatically

auto-skill enable-hook (Claude Code) or auto-skill enable-hook --target codex

Client

Delivery path

Skill install required?

Claude Code

MCP or Auto Mode hook

No

Codex CLI

MCP or Auto Mode hook

No

Cursor

MCP only (no inject hook yet — vendor limit)

No

GitHub Copilot

MCP only (no inject hook yet — vendor limit)

No

Auto Mode never means silent, persistent installation, and it's off until you opt in — read SECURITY.md first, then docs/AUTO_MODE.md for exactly what data moves where and how to turn it off.

Commands

auto-skill search "<task>"
auto-skill route "<task>"
auto-skill route "<task>" --json --show-content
auto-skill preview "<task-or-url>"
auto-skill doctor
auto-skill login
auto-skill enable-hook [--target codex]
auto-skill disable-hook [--target codex]

Full command reference, including route-prompt, feedback, metrics, and the optional developer install/validate export for skill authors: docs/AUTHORING.md.

Repo layout

  • Client (repo root): auto_skill_cli.py, mcp_server.py, hooks/ — the package end users install. Talks to the hosted backend over HTTP.

  • Backend (backend/): the FastAPI/SQLite service behind skills.autoskill.dev — ingestion, embeddings, routing, content storage, deploy scripts. Own run story and CI; not part of the pip package. See backend/README.md and backend/RUNBOOK.md.

The backend subtree tracks the standalone backend repository at https://github.com/auto-skill/auto-skill.git. After syncing it, update .autoskill-backend-subtree.json and run python scripts/check_backend_subtree.py --check-remote.

Security

Skills are instructions that can shape agent behavior. Content-hash matching detects drift; it does not establish trust. Read SECURITY.md before enabling Auto Mode or installing an unfamiliar skill.

More

Available Tools

2 tools
install_skillA

Download a skill's SKILL.md (by url, as returned from recommend_skill) and install it into ~/.claude/skills//SKILL.md so Claude Code can invoke it as a normal /skill from now on, in any project.

NameRequiredDescriptionDefault
urlYes
nameNo

Output Schema

NameRequiredDescription
resultYes

TDQS

A4.3/5.0
Behavior3/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

No annotations are provided, so the description carries the full burden. It describes the file write operation and path, but does not disclose potential side effects like overwriting existing skills, error handling, or network requirements.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Single sentence packed with essential information, no unnecessary words.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given no annotations and no schema descriptions, the description covers the core workflow well. It could mention the return value (output schema exists) but the essential usage is clear.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters5/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

With 0% schema description coverage, the description adds critical meaning: 'url' is the source from recommend_skill, and 'name' determines the folder name in ~/.claude/skills/. Both parameters are effectively explained.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the tool downloads a SKILL.md from a URL and installs it to a specific path, distinguishing it from the sibling tool recommend_skill which returns URLs.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

It explains the tool is for installing skills obtained from recommend_skill, but does not explicitly state when not to use or provide alternatives beyond the sibling reference.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

recommend_skillA

Search the auto-skill database (~200k scraped Claude skills, MCP servers, and plugins) for the one that best matches a task, and return its full SKILL.md content so it can be read and followed immediately.

Call this whenever the user's request might already be covered by an existing packaged skill/MCP server, before building something from scratch. Pass a short, keyword-rich description of the task.

NameRequiredDescriptionDefault
taskYes

TDQS

A4.5/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

No annotations provided, so the description carries the full burden. It discloses the database size (~200k) and composition (scraped skills, MCP servers, plugins), and states it returns full SKILL.md content. It implies a read-only operation, though it could be more explicit about no side effects.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is concise with two paragraphs: one explaining what the tool does and why, and one providing usage advice. Every sentence adds value, and the structure is front-loaded.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Despite no output schema and a single required parameter, the description covers the tool's purpose, input format, and usage context thoroughly. It includes database size and composition, making it complete for a search tool.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The single parameter 'task' has 0% schema description coverage. The description compensates by specifying 'a short, keyword-rich description of the task,' adding meaningful guidance beyond the schema's 'Task' label.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the tool's purpose: it searches a database of ~200k skills and returns the full SKILL.md content. It uses specific verbs and resources, and distinguishes from the sibling tool 'install_skill' by focusing on recommendation before installation.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description explicitly advises calling the tool when the user's request might be covered by an existing skill, before building from scratch. It also recommends passing a keyword-rich description. While it doesn't list explicit when-not cases, the guidance is clear.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 2 tool updatesv0.1.0
    • First observedinstall_skill
    • First observedrecommend_skill
Deploy Server
Shareable, User-Owned Context BackendPeople, applications, and agents read, write, and coordinate together in one context lake. Reach it over MCP, a CLI, a Python SDK, or REST.fulcradynamics.comAd

TDQS

A4.3/5.0

Scored across 2 tools

Disambiguation5/5

The two tools have clearly distinct purposes: one searches and returns skill content, the other installs a skill from a URL. No overlap in functionality.

Naming Consistency5/5

Both tools follow a consistent verb_noun pattern: 'install_skill' and 'recommend_skill', making them predictable and easy to distinguish.

Tool Count3/5

With only two tools, the server feels minimal for a skill management system. While it covers the core workflow of finding and installing skills, it lacks tools for listing, updating, or removing skills, which limits its utility.

Completeness3/5

The server covers the basic find-and-install flow but is missing lifecycle operations like listing installed skills, updating, or removing them. Users may need additional tools to manage their skill set effectively.

Maintenance

ActivityInactive
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • claude-oracle-mcp

    Vvkmnn
    A
    license
    A
    quality
    D
    maintenance
    An MCP server for discovering Claude Code skills, plugins, and MCP servers by searching 15,000+ resources from 17 registries, GitHub, and the web with zero setup.
    3
    8 npm
    3
    MIT
  • @skill-hub/mcp-server by skillhub-club

    @skill-hub/mcp-serverofficial

    skillhub-club
    A
    license
    A
    quality
    C
    maintenance
    Enables AI assistants to discover, search, and install Claude Code Skills from SkillHub, with tools for semantic search, browsing, recommendations, and installation.
    5
    8 npm
    1
    MIT

  • skills-mcp

    devfellowship
    F
    license
    Not graded
    quality
    D
    maintenance
    Enables Claude/MCP agents to search, list, get details, and generate install commands for skills, MCP servers, and connections from the DFL Forge skills registry.
    -