Open Source & Tooling

Agent Readiness MCP

A local stdio MCP server that exposes repository readiness and agent-context audits to an editor or other MCP host. Published on npm and in the official MCP Registry.

Why it matters

• Exposes audit_repository, audit_agent_context, and get_audit_report, with structured findings, engine commits, and paginated reports.
• The scanner writes nothing, runs no repository scripts, and makes no network requests. Workers have read access confined to the selected directory and their own script.
• Credential masking is pattern-based and applies to scanner output only; paths remain visible. The host client may send results and files it reads to its model provider.
• Verified with the MCP SDK, Inspector, and signed-in headless Claude Code. MCP in Cursor and Windows remain untested.
• The October 14 publishing record also lists mcpserver.dev and MCPRush; MCPRush auto-indexed the package and its ownership claim awaits review. Glama remains submitted, awaiting review.

Install

npx -y @alipajand/agent-readiness-mcp@0.1.2 --root /path/to/projects

What it demonstrates

I made the same pinned audit engines available through MCP, with bounded workers and structured reports instead of model-generated scores.

Related writing

• How to prepare integrations for the Anthropic and Cursor directories
• Agent instruction files are code. Lint them like code.
• Building CI tools for pull requests you don't trust

Agent Readiness skill and plugins

A portable agent-readiness-audit skill and plugins for Claude Code and Cursor, using the same bundled audit engines as the MCP server.

Why it matters

• The skill and Claude Code plugin install directly from the public repository. The plugin adds no hooks, MCP servers, or executables to PATH.
• Invoke /agent-readiness:audit in Claude Code or the agent-readiness-audit command in Cursor. Named invocation is the tested path; plain-language discovery is not verified as reliable.
• Signed-in headless Claude Code and the local Cursor plugin were tested, including denied execution. If the scanner cannot run, the loaded command must report that no audit ran and give no scores.
• As recorded October 14, the official plugin directory submission is scanning and the Cursor Marketplace submission awaits review. Local Cursor installation is tested; marketplace approval is not claimed.

Install

npx skills add alipajand/agent-readiness-action --skill agent-readiness-audit

claude plugin marketplace add alipajand/agent-readiness-action
claude plugin install agent-readiness@agent-readiness

What it demonstrates

I packaged deterministic audits for existing editor workflows while keeping scanner behavior separate from the host client's data handling.

Related writing

• How to prepare integrations for the Anthropic and Cursor directories
• Agent instruction files are code. Lint them like code.
• Building CI tools for pull requests you don't trust

Agent Readiness Doctor

My open-source contribution to the GitBot Library, available in its in-app Marketplace.

Why it matters

• Checks repository instructions against the code.
• Proposes fixes for review before editing.
• Uses agent-context-doctor and agent-readiness-kit when available.

Install

GitBot → Marketplace → Agent Readiness Doctor → Install

What it demonstrates

I contributed a repeatable workflow for maintaining the instructions developers rely on between refactors.

Related writing

• Keeping repository instructions useful with Agent Readiness Doctor

agent-context-doctor

Audits the files that steer AI coding agents: AGENTS.md, CLAUDE.md, Cursor and Copilot rules, and committed Claude Code and MCP settings. Reports each problem with a severity and scores the repository from 0 to 100.

Why it matters

• Agents follow their instruction files literally. A stray “skip tests if they’re slow,” a TODO placeholder, or two files that contradict each other quietly steer the agent toward bad changes.
• Committed agent settings run on every contributor’s machine. A bypassPermissions default, a hook that pipes a download into a shell, or an unpinned MCP server is reported as high severity.
• Also catches stale commands, broken file references, pasted secrets (redacted in the report), and hidden Unicode that can smuggle instructions past reviewers.
• Outputs terminal, JSON, Markdown, SARIF for code scanning, or pull-request annotations, with baselines and score or severity gates for CI.

Install

npm install -g github:alipajand/agent-context-doctor#v1.0.0

What it demonstrates

I treat agent instructions and agent settings as code: reviewed, linted, and gated in CI like anything else that changes behavior.

Related writing

• Agent instruction files are code. Lint them like code.
• Building CI tools for pull requests you don't trust

agent-pr-reviewer-lite

A deterministic risk pre-screen for pull requests. It reads the git diff and flags the files that deserve human review before merge: authentication, billing, migrations, CI pipelines, committed secrets, and agent permissions.

Why it matters

• Agent-written pull requests touch many files. An offline pre-screen that runs in about 100 ms tells reviewers, or an LLM reviewer, where attention is actually needed.
• It reads added lines too: skipped or focused tests, new lint suppressions, new dependencies, and agent commands that run shell without asking.
• Each finding names its CODEOWNERS, read from the base branch, and the config can be read from a trusted ref, so a pull request cannot change the rules that review it.
• Outputs text, Markdown, JSON, SARIF, JUnit, or pull-request annotations, and exits non-zero at a configurable risk level.

Install

pnpm add -D github:alipajand/agent-pr-reviewer-lite#v1.0.0

What it demonstrates

It is deliberately not an AI reviewer: fixed, readable rules, conservative by design, and identical output for identical input.

Related writing

• Which agent-written changes does a human need to read?
• Building CI tools for pull requests you don't trust

agent-readiness-kit

Scores how ready a repository is for AI coding agents, from 0 to 100 across 13 categories: instruction files, architecture docs, scripts, tests, safety boundaries, and more.

Why it matters

• Teams often adopt AI-assisted development before their conventions, documentation, and validation paths can carry it. The score makes those gaps concrete.
• ark init, ark generate, and ark fix scaffold starter files for Cursor, Claude Code, Codex, Copilot, and CI, including least-privilege Claude Code settings that pre-approve nothing.
• Writes JSON, JUnit, SARIF, Markdown, or HTML reports, keeps a score history, and gates CI with a minimum score.

Install

npm install -g github:alipajand/agent-readiness-kit#v1.0.0

What it demonstrates

I treat developer experience as a product problem: the score explains itself, and the fix is one command away.

Related writing

• Building CI tools for pull requests you don't trust

agent-readiness-action

A GitHub Action that runs agent-readiness-kit and agent-context-doctor in CI. It scores the repository, checks its instruction files and agent settings, and reports the result on pull requests.

Why it matters

• Fails the build below a minimum score or when a pull request lowers the score from the base branch, writes a Markdown report, and can post a summary comment.
• Published in the GitHub Actions Marketplace as v1.2.0, with the v1 tag pointing to it. Both engines are bundled at pinned commits; no registry downloads at run time or LLM scoring.
• Built for pull requests you don’t trust: file names cannot inject workflow commands, reports are never written through symlinks, and only the action’s own comment is ever updated.

Install

uses: alipajand/agent-readiness-action@v1

What it demonstrates

I design CI tooling on the assumption that the repository it runs on may be hostile, because on a pull request it can be.

Related writing

• Building CI tools for pull requests you don't trust
• Agent instruction files are code. Lint them like code.