Git Insights

dev-mcp-servers

Ten small MCP servers for everyday development and security checks. Each one is a separate npm package you add to Claude Code, Claude Desktop, Cursor or any other MCP client with one line, and each does one job well: look up vulnerabilities, grade security headers, decode a JWT, test a regex, explain a cron expression, validate JSON against a schema, lint an OpenAPI document or a Dockerfile, read statistics out of a git repository, or build an llms.txt.

All servers are TypeScript on the official @modelcontextprotocol/sdk, speak stdio only, validate every input with zod, bound their inputs and run time, send no telemetry, and ship with node:test suites that cover every tool.

Servers

Every package README lists its tools with inputs and outputs, the exact claude mcp add command, a .mcp.json snippet, and what the server touches on disk and on the network.

Quick start

For anyone who uses an MCP client (Claude Code, Claude Desktop, Cursor) and wants quick, local answers to everyday security and developer questions. Every server runs from a public container image, so all you need is Docker:

claude mcp add jwt-tools -- docker run --rm -i ghcr.io/basitalisandhu/mcp-jwt-tools:0.2.2
claude mcp list

Then ask your client to decode a token. A real decode_jwt result for an unsigned, expired token (trimmed):

{
  "verified": false,
  "header": { "alg": "none" },
  "payload": { "sub": "user-1", "exp": 1700000000 },
  "findings": [
    { "code": "alg-none", "severity": "critical", "message": "alg is none: the token is unsigned. Any verifier that accepts none accepts forged tokens." },
    { "code": "expired", "severity": "high", "message": "Expired 1.0 h ago (exp 2023-11-14T22:13:20.000Z)." },
    { "code": "aud-missing", "severity": "medium", "message": "No aud claim: the token can be replayed against any service that shares the key." }
  ]
}

Add -s user to install for every project rather than the current one. For clients that read .mcp.json (Claude Code, Claude Desktop, Cursor), add an entry per server:

{
  "mcpServers": {
    "jwt-tools": { "command": "docker", "args": ["run", "--rm", "-i", "ghcr.io/basitalisandhu/mcp-jwt-tools:0.2.2"] },
    "osv-advisories": { "command": "docker", "args": ["run", "--rm", "-i", "ghcr.io/basitalisandhu/mcp-osv-advisories:0.2.2"] }
  }
}

Pin the version, as above. An unpinned image or package runs whatever was published last, which turns an update into code that runs inside your editor without review.

Without Docker, run from a checkout (Node 20 or newer):

git clone https://github.com/basitalisandhu/dev-mcp-servers
cd dev-mcp-servers
npm install
npm run build
claude mcp add jwt-tools -- node "$PWD/packages/jwt-tools/dist/index.js"

The packages are not on npmjs.com yet, so npx -y @basitalisandhu/mcp-<name> does not work; use a container, a checkout, or the GitHub Packages npm registry (see Install).

Install

Every release is published in two places by publish-github-packages.yml: an npm package per server on GitHub Packages, and a container image per server on the GitHub Container Registry. The images can be pulled without signing in; GitHub Packages npm needs a token even for public packages. Publishing to npmjs.com is set up but off (see Releasing).

npm from GitHub Packages

Point the @basitalisandhu scope at GitHub Packages in ~/.npmrc:

@basitalisandhu:registry=https://npm.pkg.github.com
//npm.pkg.github.com/:_authToken=${GITHUB_TOKEN}

GitHub's npm registry asks for a token even to install public packages. That is a GitHub limitation, not a setting of this repository: use a personal access token (classic) with the read:packages scope, exported as GITHUB_TOKEN. With that in place, install globally as in the table, or let the client start the server through npx:

claude mcp add jwt-tools -- npx -y @basitalisandhu/[email protected]

Each package installs a mcp-<server> command (for example mcp-jwt-tools) that speaks MCP on stdio.

Container images

Images are built for linux/amd64 and linux/arm64, run as the non-root node user, and speak stdio, so -i is required and no port is published. Each image is tagged with the version and latest; pin the version.

With Claude Code:

claude mcp add jwt-tools -- docker run --rm -i ghcr.io/basitalisandhu/mcp-jwt-tools:0.2.2

With Cursor (.cursor/mcp.json) or any client that reads .mcp.json:

{
  "mcpServers": {
    "jwt-tools": {
      "command": "docker",
      "args": ["run", "--rm", "-i", "ghcr.io/basitalisandhu/mcp-jwt-tools:0.2.2"]
    }
  }
}

A container only sees the files you mount. For servers that read local files (dockerfile-lint, openapi-lint, osv-advisories lockfile scans, llms-txt), mount the directory read-only and pass paths inside it, for example docker run --rm -i -v "$PWD:/work:ro" ghcr.io/basitalisandhu/mcp-openapi-lint:0.2.2 and then /work/openapi.yaml. For git-insights, mount the repository at /repo (the image marks only /repo as a safe git directory): docker run --rm -i -v "$PWD:/repo:ro" ghcr.io/basitalisandhu/mcp-git-insights:0.2.2.

Every image is signed with cosign (keyless) and carries a build provenance attestation; an SPDX SBOM per image is attached to the GitHub release. To check an image before running it:

cosign verify ghcr.io/basitalisandhu/mcp-jwt-tools:0.2.2 \
  --certificate-identity-regexp '^https://github.com/basitalisandhu/dev-mcp-servers/\.github/workflows/publish-github-packages\.yml@refs/tags/v' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com
gh attestation verify oci://ghcr.io/basitalisandhu/mcp-jwt-tools:0.2.2 --owner basitalisandhu

To build an image locally from a checkout: docker build --build-arg SERVER=jwt-tools -t mcp-jwt-tools .

MCP registry

Each server is listed in the official MCP registry as io.github.basitalisandhu/mcp-<name>, for example io.github.basitalisandhu/mcp-jwt-tools. The listing points at the signed GHCR image for that version (ghcr.io/basitalisandhu/mcp-<name>:<version>, stdio transport), so a client or catalogue that reads the registry installs a server by running that image with docker run --rm -i, exactly as under Container images above. Look a server up directly:

curl -s 'https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.basitalisandhu/mcp-'
curl -s 'https://registry.modelcontextprotocol.io/v0.1/servers/io.github.basitalisandhu%2Fmcp-jwt-tools/versions/latest'

The registry lists the container images rather than the npm packages because it verifies npm ownership only on registry.npmjs.org, and these packages are on GitHub Packages. It verifies the images by pulling each one anonymously and checking its io.modelcontextprotocol.server.name label, which the Dockerfile and the publish workflow set to the server's registry name.

Security posture

• Stdio only. No server opens a port. Each one is a child process of your MCP client and exits with it.
• Validated, bounded inputs. Every tool input has a zod schema with length and range limits; files are read with size caps; regex runs, HTTP requests and git commands have timeouts; results are capped.
• Network only where the job is network, and only to the documented host. osv-advisories talks to api.osv.dev and refuses any other origin. security-headers fetches the URL you give it, refuses private, loopback, link-local and cloud-metadata addresses (including names that resolve to them), limits redirects and never reads a body. The other eight servers make no network requests at all.
• Read-only. No server writes files, runs a shell, or modifies a repository. git-insights spawns git without a shell from fixed argument lists with validated paths and revisions and GIT_OPTIONAL_LOCKS=0; the output of llms-txt is returned as text for you to save.
• No telemetry. Nothing phones home. There are no analytics, update checks or crash reporters.
• Honest tool descriptions. Tool descriptions say what the tool does and returns. They contain no instructions aimed at the model, and the outputs of decode_jwt and the lints say when something is unverified or heuristic.
• Small dependency trees. Runtime dependencies are @modelcontextprotocol/sdk and zod, plus ajv and ajv-formats for json-schema-tools and yaml for openapi-lint. npm ci from the committed lockfile reproduces the exact tree.
• Provenance. Container images get a build provenance attestation and a keyless cosign signature from the workflow run that built them, plus an SPDX SBOM on the release. When the npmjs.com release is enabled it publishes with npm publish --provenance, so each npmjs version links to its commit and workflow run.

Before adding any MCP server, including these, read its source: it runs with your user's permissions and its tool results go into the model's context. See SECURITY.md for how to report a problem.

Repository layout

packages/<name>/
  src/index.ts     server entry: createServer(), tool registrations, stdio main()
  src/*.ts         the logic, importable and tested on its own
  test/*.test.mjs  node:test suites over the built dist/ using the SDK's in-memory transport
  package.json     @basitalisandhu/mcp-<name>, bin, mcpName
  server.json      MCP registry metadata (checked against the schema in CI)
  README.md        tools, install, what it touches
scripts/check-server-json.mjs   offline validation of every server.json against the MCP registry rules
.github/workflows/ci.yml        build and test every workspace on Node 20 and 22
.github/workflows/publish-github-packages.yml   npm packages and GHCR images on a version tag
.github/workflows/release.yml   npmjs.com publish with provenance (off until NPMJS_PUBLISH is set)
Dockerfile                      one image per server: --build-arg SERVER=<name>

Build and test everything:

npm install
npm run build
npm test
npm run check:server-json

Tests run offline: network-facing servers are tested against a fake fetch, and git-insights builds a temporary repository.

Adding a server

1. Copy the closest existing package to packages/<name> and rename it: name is @basitalisandhu/mcp-<name>, bin is mcp-<name>, mcpName is io.github.basitalisandhu/mcp-<name>, and server.json repeats that name and the version, with one OCI package ghcr.io/basitalisandhu/mcp-<name>:<version> and the packages/<name> subfolder.
2. Keep the pattern: createServer() returns an McpServer; main() connects a StdioServerTransport only when the file is run directly; every tool has a zod inputSchema with limits, annotations, and a description that states what it does, what it returns and what it refuses.
3. Keep logic in separate modules so it can be tested without MCP, and write test/*.test.mjs that connect over InMemoryTransport and call every tool at least once, including an invalid input.
4. No network unless the tool's purpose is network, and then: one documented host, an allowlist check before the request, AbortSignal.timeout, a response size cap, and a fake fetch in tests.
5. Add a row to the table above, a section in CHANGELOG.md, and run npm run check:server-json. The CI matrix picks up new workspaces automatically.

Releasing

Set every package's version (they move together), add the release to CHANGELOG.md, and push an annotated tag:

git tag -a v0.2.2 -m "0.2.2" && git push origin v0.2.2

publish-github-packages.yml builds, tests, checks that every version matches the tag, then:

• publishes each workspace to GitHub Packages npm with the workflow's GITHUB_TOKEN (a version that already exists is skipped, so a re-run is safe);
• builds and pushes ghcr.io/basitalisandhu/mcp-<server>:<version> and :latest for every server, generates an SPDX SBOM, records a build provenance attestation and signs the image digest with cosign;
• publishes every server.json to the MCP registry, after the images are pushed: it installs a pinned, checksum-verified mcp-publisher, logs in with mcp-publisher login github-oidc (the job's id-token: write permission; no secret), sets the version from the tag and skips versions the registry already has, so a re-run is safe;
• creates the GitHub release for the tag with generated notes and the SBOMs attached.

No secret is needed. Pull requests that change the Dockerfile, the workflow or a package.json run the same build as a dry run.

release.yml publishes the same packages to npmjs.com and is off until the repository variable NPMJS_PUBLISH is true. Before turning it on, set up one of:

• Trusted publishing (recommended). On npmjs.com, for each package, add this repository and the workflow file name release.yml as a trusted publisher. No secret is needed; the workflow's id-token: write permission lets npm verify the GitHub OIDC token. The workflow upgrades npm first because trusted publishing needs npm 11.5.1 or newer.
• An automation token. Create a granular access token with publish rights and store it as the repository secret NPM_TOKEN; the workflow passes it as NODE_AUTH_TOKEN.

It passes --registry https://registry.npmjs.org, which overrides the GitHub Packages registry in each publishConfig. Provenance requires the workflow to run on GitHub-hosted runners from the public repository.

Once npmjs.com publishing is on, an npm entry (registryType: npm, registryBaseUrl: https://registry.npmjs.org) can be added back to each server.json; each package.json already carries the matching mcpName that the registry checks.

Related

Project hub: https://basitalisandhu.github.io/ . More tools by the same author: https://github.com/basitalisandhu

• ai-agent-incidents: open dataset of AI agent and LLM security incidents, with a searchable site.
• awesome-agent-security: curated list of AI agent security tools, papers and datasets.
• agent-security-skills: Claude Code plugin and skill pack for securing LLM agents, including an MCP server over the AI agent incident dataset.

Licence

MIT. See LICENSE.