# Midas > Self-hosted AI agent for safe LLM workflows: approvals, receipts, budget controls, local dashboard, CLI, Ollama, and MCP. - **Type:** MCP server - **Install:** `agentstack add mcp-omarkhandji-commits-midas` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [omarkhandji-commits](https://agentstack.voostack.com/s/omarkhandji-commits) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [omarkhandji-commits](https://github.com/omarkhandji-commits) - **Source:** https://github.com/omarkhandji-commits/midas - **Website:** https://github.com/omarkhandji-commits/midas#how-it-works ## Install ```sh agentstack add mcp-omarkhandji-commits-midas ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # MIDAS [](https://github.com/omarkhandji-commits/midas/actions/workflows/ci.yml) [](https://github.com/omarkhandji-commits/midas/actions/workflows/cross-platform.yml) [](https://github.com/omarkhandji-commits/midas/actions/workflows/codeql.yml) [](https://github.com/omarkhandji-commits/midas/actions/workflows/lighthouse.yml) [](https://github.com/omarkhandji-commits/midas/actions/workflows/a11y.yml) [](LICENSE) [](pyproject.toml) [](#testing) [](pyproject.toml) > **Demo (15 s):** *coming — see [tools/record_demo.md](tools/record_demo.md) to record it locally and drop the file at `docs/assets/midas-demo.gif`.* Local-first AI agent for approval-gated automation and verifiable LLM workflows. MIDAS is a self-hosted AI agent with a local dashboard, CLI, signed receipts, budget controls, Ollama support, and MCP tooling. It helps you draft, review, and verify agent actions before they change files, call services, publish content, or use external tools. For non-developers: download the repo, open the folder, then double-click `Launch MIDAS.bat` on Windows. On macOS, use `Launch MIDAS.command`. On Linux, run `./launch-midas.sh`. See [docs/INSTALL_FOR_EVERYONE.md](docs/INSTALL_FOR_EVERYONE.md). Read [DISCLAIMER.md](DISCLAIMER.md) before using MIDAS with external accounts, generated content, automation, or third-party tools. ## Who It Is For - Users who want a local-first AI agent with a dashboard instead of terminal-only setup. - Developers building LLM workflows that need approvals, receipts, and budget controls. - Operators who want an audit trail before connecting email, files, MCP tools, or external APIs. - Teams testing Ollama, cloud model providers, and self-hosted agent workflows. ## What It Does - Plans and drafts work through a bounded LLM agent loop. - Stores every step as an Ed25519-signed receipt in a hash chain. - Routes file writes, code execution, spreadsheet writes, outbound sends, Stripe intents, media files, and external MCP calls through the approval queue. - Preserves untrusted taint across agent steps so fetched pages, PDFs, emails, and third-party tool output cannot become instructions. - Uses a lightweight skill index and loads `SKILL.md` only when needed. - Links receipts to operator-recorded outcomes so runs can be reviewed later. ## Screenshots | Chat | Approvals | Proof Ledger | |---|---|---| | | | | ## Status - ✅ **Code gates** — 718 tests, mypy strict (157 files), ruff, import-linter contract `core ↛ flagship`, bandit medium+, ShipVitals `READY p0=0 p1=0`. - ✅ **Secret scan** — clean. - ⏳ **Human UAT** — see [docs/UAT_CHECKLIST.md](docs/UAT_CHECKLIST.md), unsigned at first release. - ⏳ **Lighthouse + axe-core (WCAG 2.2 AA)** — workflows wired, first results pending after `v0.1.0`. - ⏳ **Cross-platform** — Linux/macOS/Windows × Python 3.11/3.12 matrix runs via the [Cross-platform](https://github.com/omarkhandji-commits/midas/actions/workflows/cross-platform.yml) workflow. - ❌ **Independent third-party security review** — not yet performed. Until that lands, MIDAS is *local-first, approval-gated, with security defaults*. It is **not** described as "secure", "certified", or "production-ready". ## Use Cases - Run a self-hosted LLM agent from a local dashboard or CLI. - Draft files, media plans, and code changes behind approval cards. - Use Ollama locally, or connect a cloud provider with your own API key. - Route MCP tools through an approval workflow and receipt ledger. - Verify agent activity with signed receipts and an independent verifier. - Check local capabilities before installing or enabling extra tools. ## Quickstart ### No-terminal start Windows: ```text Double-click: Launch MIDAS.bat ``` macOS: ```text Double-click: Launch MIDAS.command ``` Linux: ```bash chmod +x launch-midas.sh ./launch-midas.sh ``` The launcher creates a private `.venv`, installs MIDAS, prepares local state, opens the dashboard, and prints a rescue login link if the browser does not open. ### Developer start ```bash git clone https://github.com/omarkhandji-commits/midas.git cd midas python -m venv .venv .venv\Scripts\pip install -e ".[llm,web,dev]" midas init midas dashboard ``` Open the local dashboard, connect a model, then run one mission. The dashboard is loopback-only and uses a one-time login token. `midas init` detects local Ollama, or accepts one cloud API key: ```bash midas init # running Ollama, no key needed midas init --key sk-... # OpenAI midas init --key sk-ant-... # Anthropic midas init --key sk-or-... # OpenRouter ``` ## CLI ```bash midas earn "" # scan, prepare, queue midas capabilities scan # detect local tools, no install midas capabilities plan "make a video with voice" midas approvals list midas approvals approve midas execute midas roi midas outcome record "" -m value= midas proof export out.html --run-id midas repo-map src/ midas blog-lint path/to/post.md midas course "topic" --modules 5 midas drain # queue due scheduled posts ``` Run as an MCP server: ```bash midas mcp serve ``` ## Media MIDAS never downloads tools silently. `midas capabilities scan` checks for `ffmpeg`, Node/Remotion, Edge TTS, Kokoro, Piper, XTTS/Coqui, NeuTTS, Ollama, Docker/Podman, Git, and MCP adapters. `midas capabilities plan ""` returns the local/free path, setup gaps, approval needs, privacy notes, cost notes, and fallback. Current media tools: - `image.draft`: offline PNG placeholder or opt-in provider. - `voice.synthesize`: deterministic offline WAV and opt-in provider hooks. - `video.script` and `video.storyboard`: pure planning tools. - `remotion.project.draft`: approval-gated ZIP with a minimal Remotion project. ## Security Defaults - Default-deny Sentinel policy. - Approval metadata: risk, estimated cost, expiry, hash preview when available. - Drift checks for approved file writes and `code.run`. - Per-task, daily, monthly, per-skill, and per-persona budget gates. - Remote skills are queued for review; they are not installed automatically. - Secrets stay out of receipts, logs, prompts, fixtures, and screenshots. - Kill switch blocks tool execution. See [SECURITY.md](SECURITY.md), [docs/SECURITY.md](docs/SECURITY.md), [docs/THREAT_MODEL.md](docs/THREAT_MODEL.md), and [docs/SECURITY_RELEASE_NOTES.md](docs/SECURITY_RELEASE_NOTES.md). ## Verify Receipts ```bash pip install ./tools/verify midas keys export-public python -m midas_verify .midas/receipts.jsonl --public-key ``` Flip one byte in the ledger and rerun. Verification reports the corrupted sequence index. ## Testing ```bash ruff check . mypy src lint-imports bandit -r src -ll pytest midas eval cd web && npm run lint && npm test && npm run build python -m build twine check dist/* ``` ShipVitals is used as a final release-readiness evidence pack. It does not replace tests, security review, Playwright checks, or human review. ## Project Layout ```text src/midas/core/ sentinel, budget fuse, receipts, memory, router src/midas/flagship/ CLI, dashboard, agent loop, tools, eval suites, MCP config/ policy and provider templates docs/ architecture, security, recipes, receipt spec tests/ unit, security, eval, fixtures tools/verify/ standalone receipt verifier web/ React dashboard ``` ## License MIT. See [LICENSE](LICENSE). ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [omarkhandji-commits](https://github.com/omarkhandji-commits) - **Source:** [omarkhandji-commits/midas](https://github.com/omarkhandji-commits/midas) - **License:** MIT - **Homepage:** https://github.com/omarkhandji-commits/midas#how-it-works Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-omarkhandji-commits-midas - Seller: https://agentstack.voostack.com/s/omarkhandji-commits - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.