Agasthy
Reads local Git repository metadata such as branches, remotes, uncommitted work, and duplicate clones, and respects Git ignore/tracked-file rules for safe repo-aware analysis and cleanup.
Discovers Homebrew-installed packages and includes them in the machine knowledge graph for shared-dependency queries and impact analysis, such as what breaks if a package is removed.
Discovers npm packages and project dependencies, enabling shared-dependency queries and impact analysis before removing or upgrading packages.
Optionally connects to a local Ollama model so users can ask natural-language questions about their machine; the AI can only propose actions and cannot run commands.
Discovers pnpm-managed packages and project dependencies, enabling shared-dependency queries and impact analysis before removing or upgrading packages.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
https://github.com/user-attachments/assets/2718c8ae-1576-4c4f-a494-646731a019fa
Why you'll want it
A developer machine holds hundreds of moving parts: cloned repositories, several copies of the same runtime, packages from six package managers, AI tools with their own configs, and now coding agents starting servers in every terminal. Nothing shows you how they connect, what they cost, or what is safe to remove. AGASTHY reads the machine itself, links everything in one graph, and lets you act on it safely.
π¦ Everything, discovered | Repositories (branch, uncommitted work, remotes, duplicate clones), applications, CLI tools and runtimes (which copy is active, which are shadowed), and packages from Homebrew, npm, pnpm, uv, pip and cargo. Every fact carries its evidence. |
πΈοΈ A knowledge graph of your machine | Projects, runtimes, dependencies, packages, tools and AI configuration, linked. See which projects share a dependency and what relies on what. |
π₯ Impact analysis | "What breaks if I remove this?" for any package, runtime or tool, simulated before you touch anything. Try dependency upgrades in an isolated Update Lab. |
π€ Coding agents, live | Every Claude Code, Codex, Gemini CLI, Cursor, Aider, Hermes or OpenClaw session with its memory, ports, MCP and dev servers. When an agent exits and leaves a server running, AGASTHY says which session started it and stops it in one reviewed step. |
πͺ Token spend | Tokens, requests and estimated cost per day, session, project and model, from the agents' own local logs. Counters only, never your prompts. Export to CSV. |
π§ AI ecosystem | Installed AI tools, providers, local models, MCP servers in every config, and instruction files ( |
πΎ Safe disk cleanup | Rebuildable |
π‘οΈ Security & snapshots | Outdated packages, tools of unknown origin, opt-in vulnerability advisories. Snapshot your setup and get the commands to rebuild it on a new machine. |
π¬ Ask your machine | "What is using my RAM?" Β· "Free 10 GB safely" Β· "Which repos haven't I touched in 90 days?" Β· "What breaks if I remove openssl@3?" Works with no AI model; optionally connect Ollama, LM Studio or a cloud model. |
π§ Insights + menu bar | A ranked list of what needs attention right now (low disk, swap pressure, token spikes, stale data), one click away in the menu bar. |
Related MCP server: agenticscope
Tour
Overview. What needs attention now, each with one action. | Discovery. Every repo, app, tool and package, with evidence. |
Machine Graph. Projects, runtimes and dependencies, connected. | Impact Analysis. What breaks before you remove anything. |
AI Ecosystem. Tools, MCP servers and instruction files, with conflicts flagged. | Agents. Every coding agent, and what it left running. |
Token Usage. Tokens and cost by day, project and model. | Processes. Every process with its port and project. |
Storage & Cleanup. Rebuildable folders into a restorable quarantine. | Security. Advisories, outdated and unknown-origin software. |
AI Operator. Ask your machine; changes are only proposed. | Safe by design. Every change is a dry run you confirm. |
Screenshots and the video use a synthetic demo machine, not real data. pnpm screenshots and pnpm demo:video regenerate them.
Built to be trusted
Local-first. One SQLite file on your machine. No account, no telemetry, no analytics.
Nothing changes without you. Every action is a dry run first, shows exact targets, needs your confirmation, is verified and logged.
Hard limits. Never
sudo. Never deletes Git repositories or tracked files. Never touches Docker volumes. Never stops system processes. The AI can only propose; it cannot run commands.Evidence, not guesses. Every fact has a source and a confidence level. What it cannot measure shows as unavailable, never as 0.
Light. About 0.4% CPU and under 100 MB in normal use. A 13 MB app.
Works with your coding agents
Your agent can ask AGASTHY too. The AGASTHY skill gives OpenClaw, Hermes, Claude Code and other agents that read Agent Skills real answers about your machine: which projects have uncommitted work, what is on port 3000, what your agents spent in tokens this week, and what breaks if a package goes.
/plugin marketplace add HMAKT99/Agasthy # Claude Code: skill + MCP server
npx skills add HMAKT99/Agasthy # Codex, Cursor, Gemini CLI and more
hermes skills install HMAKT99/Agasthy/skills/agasthyOr point your agent at llms.txt: "Install AGASTHY for me by reading https://raw.githubusercontent.com/HMAKT99/Agasthy/main/llms.txt".
It runs agasthy, the read-only command-line tool attached to each release, and leaves every change to you. Setup for each agent: integrations/agent-skill.
Prefer MCP? agasthy mcp is a read-only MCP server with nine tools (status, find, entity, impact, shared dependencies, agents, processes, token usage, telemetry) for Claude Desktop, Claude Code, Cursor, VS Code, Codex and any other MCP client:
claude mcp add agasthy -- ~/.local/bin/agasthy mcpSetup for each client: integrations/mcp.
Install
Expect one security prompt on first launch. AGASTHY is not yet signed with a paid Apple Developer ID or a Windows code-signing certificate. That does not mean anything is wrong with the download; it means the operating system cannot check who published it.
macOS: "Apple could not verify AGASTHY is free of malware" β System Settings β Privacy & Security β Open Anyway.
Windows: "Windows protected your PC" (SmartScreen) β More info β Run anyway.
Linux: no prompt; just make the AppImage executable.
You only do this once; updates install without it. To check a download yourself, compare it with the .sha256 file next to it on the release page (shasum -a 256 <file> on macOS, Get-FileHash <file> on Windows, sha256sum <file> on Linux).
macOS (Apple silicon, macOS 12+)
With Homebrew:
brew install --cask hmakt99/tap/agasthy # the app
brew install hmakt99/tap/agasthy-cli # optional: the read-only CLI and MCP server for agentsOr by hand:
Download
AGASTHY-β¦-apple-silicon.dmgfrom Releases.Drag AGASTHY to Applications.
First launch: macOS says it cannot verify the developer (AGASTHY is not notarized by Apple yet). Open System Settings β Privacy & Security, scroll down and click Open Anyway. You only do this once.
Windows (x64, preview)
Download AGASTHY_β¦_x64-setup.exe from Releases and run it. SmartScreen warns because the installer is not code-signed yet: More info β Run anyway.
Linux (x64, preview)
Download AGASTHY_β¦_amd64.AppImage, then chmod +x it and run it; or install the .deb with sudo apt install ./agasthy_β¦_amd64.deb. Built and launch-tested in CI on Ubuntu; feedback from other distributions welcome.
AGASTHY lives in the menu bar: closing the window keeps it running, and Quit is in the menu-bar menu. Updates are offered in the app, signature-checked, and installed only when you click Install and restart.
Supported platforms
Platform | Status |
macOS 12+ (Apple silicon) | Supported. Built and tested on macOS 15 / M2. |
macOS (Intel) | Expected to work from source; not tested, no download yet. |
Windows 10/11 (x64) | Preview. Installer with each release; built and tested in CI. Known gaps: graceful Stop (Force quit works), WinGet packages, GPU details. Reports welcome. |
Linux (x64) | Preview. AppImage and |
Everything it does
Area | What AGASTHY does |
Live resources | CPU (total and per core), memory, swap, disk per volume, load average, process count, top processes by memory and CPU, GPU model. Anything it cannot measure is shown as unavailable with a reason, never as 0. |
Repositories | In folders you choose: finds Git working trees and reports branch, HEAD commit, last commit time, working-tree status (staged / modified / untracked / conflicted), local branch count, remotes (credentials stripped), GitHub origin, languages, frameworks, licence, manifests, size on disk, and duplicate clones of the same project. |
Applications | Bundles in |
CLI tools & runtimes | A curated catalogue of about 75 developer tools. For each copy on PATH: version, which copy is active and which are shadowed, and how it was installed (for example, it resolves into a Homebrew keg). |
Packages | Read-only inventories from Homebrew (formulae and casks), npm and pnpm globals, uv tools, pip (default |
Provenance | Every fact carries evidence and a confidence level (verified, strong, probable, unknown). A Git remote read from the repository's own config is verified; an upstream taken from a Homebrew formula's homepage is strong. Nothing is guessed from names. |
Dependencies | Reads |
Runtime requirements & impact | Links projects to the runtimes they declare ( |
Machine Graph | An interactive graph of projects, runtimes, dependencies, packages, binaries and upstream projects, with focus, depth and kind filters, search, and the evidence behind every edge. |
AI ecosystem | Finds AI coding assistants, desktop AI apps and local inference engines from evidence on disk; the providers they use (with how strongly each is confirmed); MCP servers in every known config; local models; and AI instruction files ( |
Processes | Every process with CPU, memory, listening ports, the project it runs in, and why it is protected if it is. Stop or force-quit your own processes after a reviewed confirmation; system and other users' processes are refused. |
Storage & cleanup | Measures dependency folders and build output in your projects (only those Git ignores and that contain no tracked files) and a fixed list of developer caches. You pick items, review a dry run, and they move to a quarantine you can restore from; space is freed when you delete them permanently. Docker build cache and dangling images can be pruned; volumes never are. |
Impact analysis | "What if I remove X?" for packages, runtimes, dependencies, tools and apps: Homebrew packages that need it (from what installed kegs actually link to), executables and runtimes removed with it, and every project whose runtime would break or fall back, with unknowns and a suggested plan AGASTHY does not run. |
Safe Update Lab | Try a dependency update (npm, pnpm, Cargo) in a separate Git worktree. Install and test steps run only when you approve them (install scripts off by default), and results are copied back only on request. |
Snapshots | Record installed software, repositories (commit, branch, remote) and environment variable names; compare snapshots; list the commands that would bring back what is missing. Metadata only, not a backup. |
Security & lifecycle | Outdated Homebrew packages, tools of unknown origin, and, if you opt in, OSV.dev advisories for npm, PyPI, crates.io and Go packages with exact versions. |
AI Operator | Ask questions in plain language: what is using my RAM, free 10 GB safely, which repos have I not touched in 90 days, which MCP servers are configured, what breaks if I remove openssl@3. A built-in mode answers common questions with no AI model and sends nothing anywhere. Optionally connect a local model (Ollama, LM Studio, vLLM, LocalAI) or, after a separate opt-in, a cloud model (Anthropic, OpenAI, Gemini, OpenRouter, any OpenAI-compatible endpoint). The model can only read through AGASTHY's typed tools and propose changes; every proposal opens the same dry-run review as the rest of the app. |
Agents | Every running coding-agent session (Claude Code, Codex, Gemini CLI, Cursor Agent, Aider, Hermes Agent, OpenClaw and more, plus any process that starts MCP servers) with its process tree, memory, ports and MCP/dev servers; what is still running after an agent exits, with one reviewed action to stop it; Git worktrees agents work in. |
Token Usage | Tokens, requests and estimated cost per day, session, project folder and model, read from Claude Code and Codex logs and Hermes Agent's database. Counters only, never message text. |
Insights & menu bar | A ranked list of what needs attention (low disk, swap pressure, processes agents left running, token spikes, stale data, security findings), and a menu-bar item with the same at a glance. |
Audit & privacy | An append-only audit log of every action, and a privacy page listing exactly which programs AGASTHY may run. |
Repositories, installed packages, binaries on PATH and applications are kept as separate kinds. A local clone is never treated as installed software.
Repositories, installed packages, binaries on PATH and applications are kept as separate kinds. A local clone is never treated as installed software. Open items are in ROADMAP.md.
Under the hood
Rust workspace, thin Tauri 2 shell, React UI. 290+ Rust tests, Playwright UI tests, CI on macOS, Windows and Linux.
apps/desktop/ React + TypeScript UI; src-tauri/ is a thin command layer
crates/core/ domain model, SQLite store (FTS5), migrations, audit log
crates/platform-adapters/ OS adapters (macOS / Windows / Linux) and the guarded command runner
crates/discovery/ repositories, software, packages, AI ecosystem, token usage
crates/graph/ graph views, shared dependencies, impact analysis
crates/process-manager/ processes, ports, coding-agent sessions
crates/cleanup/ disk analysis, safety checks, quarantine, Docker usage
crates/actions/ typed actions: plan (dry run) β confirm β execute β verify β audit
crates/ai-engine/ AI Operator: typed tools, built-in answers, model providers
crates/snapshots/ environment snapshots, comparison, reconstruction plans
crates/security/ outdated software, unknown provenance, opt-in OSV.dev advisories
crates/update-lab/ Git-worktree labs for trying dependency updates
crates/cli/ `agasthy` headless CLI (JSON output)More in docs/architecture.md.
Development setup
Requirements: Rust (stable, 1.90+), Node.js 20+, pnpm 10, and on macOS the Xcode Command Line Tools.
pnpm install
pnpm dev # runs the desktop app (Vite + Tauri)Checks:
cargo fmt --all --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace
pnpm typecheck && pnpm lint && pnpm test
pnpm e2e # Playwright UI tests (uses installed Chrome locally)
pnpm check:types # generated TS types match the Rust modelHeadless discovery (prints JSON, changes nothing):
cargo run -p agasthy-cli -- scan ~/code --depth 4
cargo run -p agasthy-cli -- software
cargo run -p agasthy-cli -- packages
cargo run -p agasthy-cli -- telemetry
cargo run -p agasthy-cli -- impact runtime:/usr/local/bin/python3 --db <file>Build a DMG (unsigned unless Apple credentials are set; see docs/releasing.md):
pnpm release:mac # output: dist/AGASTHY-<version>-apple-silicon.dmgDetails: docs/development.md.
Privacy
No account, telemetry or analytics. The only network requests: an update check against this repository's Releases (can be turned off; sends nothing about your machine), OSV.dev advisory checks if you turn them on, and an AI model endpoint if you configure one (cloud models need a separate opt-in). Repositories are scanned only in folders you add. External programs run only from an allowlist, with fixed arguments, never through a shell. Full details: docs/privacy.md.
Contributing
Contributions are welcome: see CONTRIBUTING.md, CODE_OF_CONDUCT.md and the ROADMAP. Found a bug? Settings β About β Report a problem prepares a redacted report. Security issues: SECURITY.md, not public issues.
Name
AGASTHY is the product name. A suggested backronym, not a historical meaning: Autonomous Governance, Analysis, Systems, Tooling & Hybrid intelligence.
License
Apache License 2.0. See LICENSE.
Please note: the first launch asks for your OK once. AGASTHY is free and not yet signed with a paid Apple or Microsoft certificate, so your system shows a warning the first time. This is expected, and the install steps show the one click to get past it. Every download is checksummed, built in public by GitHub Actions, and in-app updates are signature-checked.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelogβ¦
This server cannot be deployed
Maintenance
Related MCP Connectors
The Cortex MCP server provides read-only access to real-time engineering context from the Cortex developer portal, allowing AI coding assistants to answer natural language questions about your organization's catalog (microservices, libraries, domains, teams, infrastructure), scorecards (engineering standards and best practices), initiatives (goals and deadlines), and Engineering Intelligence metrics. It includes tools for querying documentation, tracking personal entities, and accessing AI-assisted insights across the entire Cortex ecosystem.
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
Free, source-checked directory of AI agent tools, MCP servers and skills, with read-only search.
Search and inspect source-pinned agent skills, plugins and tools through a free public MCP.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceAn MCP server that gives LLMs structured, read-only insight into local code repositories β directory tree, languages, dependencies, scripts, and config files.MIT
- AlicenseNot gradedqualityCmaintenanceA read-only MCP server that provides AI agents with live, structured workspace awareness, including project listing, git status, and budgeted context packing, minimizing token usage.3 npm2MIT
- AlicenseAqualityBmaintenanceA read-only MCP server that exposes local coding-agent session logs as three tools for introspection of recent work, debugging tool failures, and tracking token usage and estimated cost without parsing log files.3MIT
- AlicenseAqualityBmaintenanceA secure, local-first MCP server for read-only inspection and troubleshooting of development environments, exposing narrow, typed, auditable capabilities for repository inspection, log summarization, Docker review, and security scanning without granting unrestricted machine access.8MIT