ai-skill-ai MCP server
Commit Oct 31, 2026Provides a library of agent skills, role agents, and MCP servers to help AI assistants build software, conduct research, and perform quantitative analysis.
Also known as devpilotX/ai-skill-ai
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 31, 2026
- License
- MIT
What is ai-skill-ai MCP?
ai-skill-ai provides a library of agent skills, role agents, and MCP servers to help AI assistants build software, conduct research, and perform quantitative analysis.
Summary generated from the project's README
About ai-skill-ai MCP server
Developers and researchers can use this library to provide AI assistants with specialized capabilities for software engineering, quantitative analysis, and academic research. The system provides sixty-six distinct skills, thirteen role agents, and eight MCP servers that work together to ensure an assistant verifies its work instead of guessing. By integrating these components, an AI can move beyond generic responses to perform precise tasks like backend architecture planning, SEO audits, or mathematical verification.
To use the library, you can clone the repository and copy specific skill folders into your host environment. For MCP client integration, you install uv and add the server entries to your configuration. The system relies on Python 3.9 or later and includes forty-six scripts that act as tested checkers. It is designed to prevent common AI failures such as flattery or high-probability convergence by enforcing strict procedures and using exact math tools.
What you can do with it
- 1Software development lifecycle managementAn AI can manage a full build process through staged phases including research, planning, and implementation.
- 2Technical SEO auditingA developer can use the assistant to perform a polite crawl of a live website to generate a prioritized list of fixes.
- 3Improving writing qualityYou can task an AI with removing the stylistic markers of machine writing and verifying the result with a rule detector.
- 4Infrastructure and deployment planningAn assistant can estimate hosting costs and ensure rollback procedures are actually performed.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 16, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachabledevpilotx/ai-skill-ai is public and reachablegithub.com/devpilotx/ai-skill-ai ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-31github.com/devpilotx/ai-skill-ai ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubai-skill
Skills, agents and MCP servers that make an AI assistant check its work instead of guessing.
Quick start · Skills · Agents · MCP servers · Scripts · Verification
Sixty-six agent skills, thirteen role agents, eight MCP servers and a set of house rules for AI
assistants that build software, write, research and work with numbers. Every skill follows the open
Agent Skills specification, so it works in any host that reads
SKILL.md folders.
The skills are built around one rule: say the true thing, in the user's own situation, without the accent of machine writing. Two failures show up in almost every assistant. Flattery produces agreement instead of assessment, and convergence hands every user the same high-probability answer. Fixing only the first gives you a confident, rude, generic answer, which is not an improvement.
| Part | Where | What it is |
|---|---|---|
| 66 skills | skills/ |
One folder per job: a SKILL.md, reference files for depth, and a tested checker or calculator where the job needs one. |
| 13 agents | agents/ |
Role agents that each load a set of skills, from builder and architect to security, analyst and quant. |
| 8 MCP servers | mcp/ |
70 tools over the skills' checkers, free public APIs and exact maths, on a shared safety library. |
| 46 scripts | skills/*/scripts/ |
Standard-library Python 3.9 or later, each with --help and a --self-test that CI runs. |
| House rules | rules/ |
Always-on rules and a one-line-per-skill index, to keep in an assistant's context. |
flowchart LR
Q["Your request"] --> X["rules/skill-index.md<br/>picks the skill"]
H["rules/house-rules.md<br/>always loaded"] -.-> S
X --> S["skills/*/SKILL.md<br/>non-negotiables and procedure"]
G["agents/*.md<br/>a role with a set of skills"] --> S
S --> R["references/<br/>depth, read on demand"]
S --> C["scripts/<br/>tested checkers"]
M["mcp/servers/<br/>tools for any MCP client"] --> C
M --> P["Public APIs and exact maths"]
A host loads each skill's name and description at startup and opens the full skill only when a request matches it. The skill sets the rules and the steps, its references hold the depth, and its script measures the result. The MCP servers expose the same checkers, plus public data and exact maths, as tools.
git clone https://github.com/devpilotX/ai-skill-ai.git
- Skills. Copy the folders you want from
skills/into the folder your host reads skills from. Each one is self-contained, and the folder name is the skill name. Single skills, or all of them, are also on the latest release as zips, withSHA256SUMS.txtto verify the download. - House rules. Add
rules/house-rules.mdandrules/skill-index.mdto the instructions your host loads on every turn, such as anAGENTS.mdfile or the host's own rules folder. - Agents. Each agent is one Markdown file with its skills and access level in the frontmatter and its system prompt as the body.
agents/README.mdexplains the format, the permission policy the prompts expect, and how to map both onto a host. - MCP servers. Install uv, copy the entries you want from
mcp/mcp.example.jsoninto your MCP client's configuration, and replace/path/to/ai-skillwith the path of your clone.mcp/README.mdcovers the tools, the keys and the safety rules.
A host matches requests to skills on its own. In hosts that expose skills as slash commands, you can also call one by name:
/build-pilot build a booking system for a two person dental practice
/reality-check I want to start a timber company with 30,000
/incident-response checkout is returning 502s for a third of users
/email-deliverability our invoices land in spam at Gmail
/data-analysis why does the dashboard revenue differ from finance's number
/privacy-engineering what do we have to delete when a customer closes their account
/quant-research is this momentum backtest real or overfit
Every skill has an off switch. Saying "stop", "I've decided" or "just execute" ends it for the session. Skills that nag get uninstalled.
Every skill has an activation section with an off switch, numbered non-negotiables, a procedure, a self-audit and a statement of what it cannot do, and the validator fails the build if any of those is missing. Each one routes a request that belongs to a sibling skill to that skill. Open a group to see its skills.
| Skill | What it does |
|---|---|
| build-pilot | The whole build as staged phases: think, research, argue with itself, decide, plan, implement, verify. Gates where your input is needed, and no code before the deciding facts are verified. |
| website-build | Delivery playbook for a website or web app, from goals and page inventory to a launch checklist where every item names the skill or script that verifies it. |
| frontend-build | The six data states most interfaces omit, state in the narrowest scope that works, accessibility during the build, bundle size budgeted in kilobytes. |
| backend-build | Object level authorisation, a stated concurrency assumption per write path, timeouts everywhere, idempotency before the retry arrives. |
| mobile-build | Offline behaviour decided before the screens, process death survived, store rejection causes covered. |
| ml-build | Evaluation set and trivial baseline before the model. Treats a good score as leakage until proven otherwise. |
| web-3d | Real-time 3D with Three.js and React Three Fiber: a device budget first, compressed assets, rendering on demand, and real HTML kept for search. |
| data-layer | Schemas, indexes read from the query plan, and migrations treated as production operations with a lock estimate. |
| infra-deploy | Simplest hosting the requirements allow, rollback actually performed once, cost estimated from retrieved prices. |
| arch-decide | Classifies a decision by what reversal costs, spends effort in proportion, and writes the decision record. Refuses complexity your team size cannot operate. |
| stack-select | Language, framework, runtime and datastore per component, from versions and end-of-life dates checked live against endoflife.date. |
| migration-plan | Incremental migration with both systems running and every step reversible. No cutover weekend. |
| i18n-build | Many languages and regions, including the major Indian scripts, Arabic and Urdu right to left, ICU plurals, and lakh and crore number formats. |
| mcp-build | MCP servers built from the threat model out: one named job per tool, typed and bounded inputs, honest annotations, previews before side effects, and a linter for the tool list. |
| email-deliverability | SPF, DKIM, DMARC, MX and transport policy measured from live DNS before and after each change, against the current Gmail, Yahoo and Microsoft sender rules and RFC 9989. |
| Skill | What it does |
|---|---|
| ui-compose | Pages assembled from openly licensed components, icons, fonts and imagery, with every licence checked at its source and recorded. |
| visual-design | Type scale, spacing, colour and dark mode tokens, with anti-generic rules and a script that flags the templated look. |
| motion-design | Interface motion on Material 3 and Carbon tokens, a reduced-motion branch for everything that moves, and compositor-only properties. |
| brand-logo | Logos, wordmarks, favicons and app icons built in vector and tested at 16 px, with licence and trademark risk checked. |
| ux-writing | Buttons, labels, errors and empty states in sentence case with one term per concept, checked by a copy linter. |
| web-analytics | Tags, consent, events and conversions that load lawfully and measure the outcomes the site exists to produce. |
| Skill | What it does |
|---|---|
| seo-technical | Crawlable, renderable, canonical pages: rendering mode per framework, URL design, redirects and robots, measured on the built HTML. |
| seo-content | Keyword research, search intent, topical maps and briefs from free first-party sources, with paid tools optional. |
| seo-indexing | Diagnoses Search Console indexing reports from live status codes, robots directives and access logs, server errors first. |
| seo-audit | A polite, robots-aware crawl of a live site, turned into a scored and prioritised fix list. |
| seo-authority | Links and mentions earned within Google's link spam policy, and a backlink profile read for risk. |
| structured-data | JSON-LD that matches the visible page and earns the rich results that still exist, checked on the built output. |
| local-seo | Google Business Profile, the map pack, NAP consistency, location pages, citations and reviews. |
| ai-search-visibility | Findable and citable by AI answer engines: crawler user agents checked against each vendor's own documentation, server-rendered HTML, referral measurement. |
| Skill | What it does |
|---|---|
| human-prose | Removes the stylistic residue of machine writing, then verifies with a tested 33 rule detector rather than a judgement call. |
| doc-forge | Documents chosen by what the reader does next, plus a zero dependency Markdown to print ready HTML and PDF converter. |
| prompt-forge | Diagnoses which of six defects a prompt has and rewrites it. Refuses to invent a score out of ten. |
| Skill | What it does |
|---|---|
| quality-gates | The same ordered verification loop on every change, and CI with actions pinned to commit SHAs and least-privilege permissions. |
| code-craft | Code that reads as written by an experienced engineer in the project's language, with a detector for machine-generated habits. |
| code-review | Defects that cost something. Ignores anything a linter owns. Reports a clean change as clean. |
| test-strategy | Effort allocated by what a failure costs. Says plainly that coverage percentage measures execution, not correctness. |
| refactor-safely | Never mixes a refactor with a behaviour change. Safety net first, small reversible steps, suite green between each. |
| debug-method | Reproduce, then one hypothesis at a time with a prediction that can be wrong. A symptom that stopped is not a fix. |
| performance-tuning | Profile first, fix the dominant cost, report before and after from the same method. |
| security-hardening | Threat model first, findings ranked by reachability, advisories retrieved rather than recalled. |
| privacy-engineering | A field-by-field personal data inventory, consent kept as evidence, deletion that reaches backups and processors, a breach clock written in advance, and a scanner for personal data in code and logs. |
| observability-setup | Works back from the questions you need answered at 3am. Every alert has an owner and an action. |
| incident-response | One commander, severity set early, harm stopped before the hunt for a cause, updates on a schedule, legal notification clocks, and a blameless postmortem checked by a linter. |
| release-manage | Separates deploying code from releasing behaviour, so a bad change is a toggle rather than an incident. |
| ship-audit | Twelve gate production readiness audit with a fixed severity to verdict rule. Every finding names a file, a line, the trigger, and what is lost. |
| devex-tooling | One command from clean checkout to running, with pinned versions and a feedback loop short enough that nobody skips it. |
| git-workflow | Short-lived branches, atomic commits, safe force pushes, recovery with reflog and revert, history rewrites with git-filter-repo, and an audit for oversized blobs and credential files. |
| numbers-check | Recomputes with a script, carries units, cross-checks a second way, names the assumption that decides the result. |
| accessibility | WCAG 2.2 AA, tested the way auditors test: a static markup check, then the named manual tests no script can do. |
| Skill | What it does |
|---|---|
| deep-research | Graded sources, reported contradictions, and an explicit list of what could not be established. |
| reality-check | Names the answer every other user got, bans it, then judges your idea against your own non-transferable advantages. BUILD, PIVOT or KILL with dated kill criteria. |
| oss-sourcing | Finds a free API, dataset or library and vets its licence, terms, rate limits, maintenance and supply-chain risk before you depend on it. |
| data-analysis | The metric defined before it is computed, every input profiled, totals tied to a trusted figure, Simpson's paradox and fan-out joins checked, and charts that do not mislead. |
The quantitative skills are written to the standard of institutional research, trading and risk desks: every figure computed by a tested script, every model assumption named, and every backtest treated as a false discovery until the evidence says otherwise. None of it is investment advice.
| Skill | What it does |
|---|---|
| business-model | Contribution margin, the working capital cycle, break-even, payback, LTV from a retention curve. Models cash rather than only profit, with a decimal calculator. |
| finance-books | Bookkeeping, reconciliation, period close and statement checking, with a reconciliation script. Stops at the line where a licensed accountant is required. |
| career-strategy | Career decisions priced in years and money, anchored in your financial position and rare skill combinations. |
| job-hunt | Diagnoses which stage of the funnel is broken before rewriting anything. |
| quant-reasoning | Probability set up formally and checked twice, expected value with variance, Bayesian updating, Kelly and ruin, market making games, forecast scoring. |
| quant-research | Falsifiable hypothesis, point-in-time data, a counted trial budget, purged validation, deflated Sharpe ratio, costs and capacity, dated kill criteria. |
| portfolio-risk | Shrunk and denoised covariance, risk budgeting, fractional Kelly, VaR and expected shortfall with backtests, stress and liquidity tests, limits with owners. |
| derivatives-pricing | Forward and parity first, arbitrage-free surfaces, validated numerics, Greeks with units, P&L explained by realised against implied volatility. |
| execution-microstructure | Implementation shortfall from the decision price, square root impact, Almgren-Chriss schedules, markouts, market structure and market making. |
| stat-arb | Pairs and residual reversion with Engle-Granger critical values, search correction, rules from the fitted half-life, and a break rule. |
| trading-systems | Pre-trade checks with no bypass, tested kill switches, order state machines, backtest and live parity, reconciliation, atomic deploys. |
| Skill | What it does |
|---|---|
| skill-forge | Writes and audits skills, rules files and agent definitions against the Agent Skills specification and this collection's rubric, then runs the validator. |
| session-continuity | Keeps a long run alive across context compaction, crashes and restarts with one state file, a checkpoint after every slice and a stuck protocol. |
build-pilot is the entry point for anything substantial, and website-build takes over when the
product is a website. They call the others in order: deep-research to verify the facts,
reality-check to test the premise, arch-decide and stack-select to record the choices, the build
skills to implement, test-strategy, code-review and quality-gates while working, then ship-audit
before release. In production, observability-setup notices, incident-response runs the response,
debug-method finds the cause and release-manage rolls back. For trading work, quant-research
tests the idea, execution-microstructure prices the trading, portfolio-risk sizes it, and
trading-systems carries it into production. Each skill also works alone.
| Agent | Skills it loads | Access |
|---|---|---|
| builder | website-build, build-pilot, ui-compose, code-craft, quality-gates | read-write |
| designer | ui-compose, visual-design, motion-design, web-3d, brand-logo, ux-writing, accessibility | read-write |
| architect | arch-decide, stack-select, migration-plan, data-layer, backend-build, infra-deploy | read-write |
| operator | infra-deploy, observability-setup, release-manage, incident-response, performance-tuning, devex-tooling, quality-gates | read-write |
| debugger | debug-method, performance-tuning, test-strategy, observability-setup, git-workflow | read-write |
| seo | the eight SEO and growth skills, plus web-analytics | read-write |
| analyst | data-analysis, numbers-check, data-layer, web-analytics, ml-build | read-write |
| reviewer | code-review, code-craft, security-hardening, accessibility, test-strategy | read-only |
| security | security-hardening, privacy-engineering, code-review, oss-sourcing, git-workflow, quality-gates | read-only |
| researcher | deep-research, reality-check, oss-sourcing, stack-select, arch-decide | read-write |
| strategist | reality-check, business-model, deep-research, numbers-check | read-only |
| writer | human-prose, ux-writing, doc-forge, seo-content, prompt-forge | read-write |
| quant | the seven quantitative skills and numbers-check | read-write |
The read-only agents run tests and scanners but never edit, so their output is findings. The quant agent analyses and models; it never places, changes or cancels a live order.
| Server | Tools | What it gives an agent | Keys |
|---|---|---|---|
| quality | 8 | The prose, code, copy, accessibility, UI and collection checkers, and the project's own quality gates. | None |
| seo | 17 | Site audits, page lint, index checks, log and JSON-LD checks, PageSpeed, CrUX, Search Console, Bing data, IndexNow. | Optional |
| research | 10 | OSV, deps.dev, end-of-life dates, package registries, arXiv, Crossref, OpenAlex, Wikipedia, repository vetting. | Optional |
| design | 11 | Icons, fonts, openly licensed media, CC0 3D assets and photos with their licences and attribution, plus the design checkers. | Photos only |
| media | 5 | Local upscaling, background removal, image conversion and SVG optimisation, and approval-gated image generation. | Generation only |
| market | 5 | Read-only FX reference rates, FRED series, equity quotes and crypto prices. No order or account functions. | Optional |
| netcheck | 4 | DNS over HTTPS, email authentication (SPF, DKIM, DMARC, MTA-STS), TLS certificates and HTTP security headers. | None |
| math | 10 | Exact algebra and calculus, arbitrary-precision decimals, matrices, statistics with a t interval, and unit conversion. | None |
No server exposes an arbitrary shell, code evaluation or filesystem tool. Every tool declares whether
it reads, writes or reaches the network. Paths stay inside the workspace, outbound connections are
refused when they resolve to a private address, secrets come only from environment variables, and
anything that spends money or sends data does nothing without an explicit confirm.
Forty-six checkers and calculators ship inside the skills. All of them are standard-library Python 3.9 or later with no third-party packages, and eac
ai-skill-ai MCP questions
What is included in ai-skill-ai?
The package contains 66 agent skills, 13 role agents, 8 MCP servers, and 46 scripts.