The split is what makes it usable: analysis, secret scanning and dependency auditing happen on the filesystem with no token at all, while anything that changes a repository — a pull request, an issue — returns a preview first and only executes when the same call is repeated with confirmation. Credentials go into the OS keyring through one setup call rather than sitting in a JSON config, which also means the same token works across every client on the machine.
A single server covering the jobs an agent needs around a codebase — local git and static analysis, GitHub reads and writes, vulnerability data, and process and log inspection — with credentials kept in the operating system keyring rather than in a config file.
- Local git history read directly from the repository — authors, frequency, recent changes and commit patterns, filtered by date range or author
- Python static analysis that flags unused imports, likely bugs, style violations and high-complexity functions without executing anything
- A secret and PII scan over local source: API keys, tokens, passwords and private keys, emails, SSNs and card numbers, plus missing .gitignore protections
- A dependency audit against the OSV vulnerability database covering Python, JavaScript, Go, Java and C++ manifests, reporting CVEs and outdated versions
- Code search by pattern across a local directory or a remote GitHub repository
- GitHub reads: a repository's metadata and structure, and any file at a given path and ref
- GitHub writes behind a confirmation step — pull requests opened and issues created, updated, closed or triaged, each returning a preview until you call it again with confirmed=true
- GitHub security data: code scanning alerts from CodeQL, Semgrep or another SAST tool, and Dependabot alerts with CVE id, severity, affected package and fix version
- Web work: any URL fetched with its title, meta description, Open Graph tags, canonical URL, robots directives, heading structure and link breakdown, and search through a SearXNG instance you run
- Local operations: whether a named service or port is up with its PID, and the last N lines of a log file with a filter applied
- A search over the server's own tools by keyword, tag or module, returning which connector each one needs
Python 3.12 or newer; pip install agentic-store-mcp, then run agentic-store-mcp over stdio. Nothing is needed for the local tools — git history, linting, secret scanning and dependency audit read the filesystem. GitHub tokens are optional for public repositories and required for private ones; opening pull requests and managing issues needs repo scope, and the security alert tools need security_events or repo. Tokens are stored once through the configure tool into the system keyring and persist across sessions and clients. Web search needs a SearXNG instance running locally via docker compose. A bundled local web UI, started with agentic-store-webapp at http://localhost:8765, holds the connector keys and decides which tools are exposed to the client. MIT.
One command — pip install agentic-store-mcp --upgrade