An open-source AI agent that runs on your machine, works for hours, and won't blow your budget.
Hand Maverick a goal. Its orchestrator decomposes it, spawns specialist sub-agents — researcher, coder, writer, verifier — that work in parallel, checks their output, and returns a result. Every step runs under a hard spending cap and through a safety layer, on the models you choose.
- 🧠 Long-horizon swarm. Recursive multi-agent coordination built for hours-long tasks, not one-shot replies.
- 🛡️ Safe by default. Agent Shield screens every prompt, tool call, and output — the full SDK scores F1 0.988 on prompt-injection, and a built-in rule set ships out of the box (fail-open, never a hard dependency).
- 🎛️ Your models, your budget. 12 providers, routable per role (plan on Opus, summarise on Haiku). Set a dollar + wall-clock + tool-call cap the kernel refuses to exceed — no surprise bills.
- 💻 Runs anywhere. Desktop app, one-line script, Docker, VPS, or a phone companion across 12 channels. MIT-licensed, no telemetry, no paid tier.
pipx install 'maverick-agent[installer]'
maverick init # four questions, safe defaults
maverick start "Research the top 3 CRMs for a 5-person startup and write a recommendation"Prefer no terminal? Grab the double-click desktop installer. New here? See docs/getting-started.md.
Alpha, but installable today: all six packages are on PyPI, the one-line installers work on Windows/macOS/Linux, and a native double-click installer builds for all three. See docs/getting-started.md for the full flow.
| Component | v0.1 (today) | Planned (v0.2+) |
|---|---|---|
| Install | Native installer (.exe / .dmg / .AppImage), one-line bootstrap (install.ps1 / install.sh), pipx, or from source |
Code-signed bundles + auto-update |
| GUI | Native installer app + local web dashboard (maverick dashboard) + chat at /chat |
Native Tauri shell for the agent itself + iOS/Android |
| Sandbox | Local subprocess, Docker, SSH, Podman, devcontainer, Firecracker, Kubernetes | Modal, Daytona |
| AI providers | Anthropic (full), OpenAI, OpenRouter, Ollama, Gemini, DeepSeek, Bedrock, Azure, xAI, Moonshot, TGI, vLLM (per-role routable) | Cohere |
| Channels | All 12 wired — Telegram, Discord, Slack, Signal, Email, Matrix, Bluesky, Mastodon, Voice; WhatsApp/SMS (need Twilio), iMessage (macOS-only) | Push notifications |
| Safety | Shield wired at 3 chokepoints; agent-shield SDK if installed, else a built-in rule set | Agent-shield full ~115 patterns |
| Distribution | PyPI (6 packages), GHCR image, PyInstaller binaries, native installers on Releases | Code signing; Homebrew tap |
| Tests | 2000+ tests, ruff + pytest on Py 3.10/3.11/3.12 | Integration suite + benchmark RESULTS.md |
Grab the installer for your OS from the latest release ›, double-click it, then press Install Maverick:
| OS | File on the release |
|---|---|
| Windows | Maverick_*_x64-setup.exe |
| macOS | Maverick_*_aarch64.dmg |
| Linux | Maverick_*_amd64.AppImage |
It's unsigned for now, so the first launch shows an "unknown developer" prompt — on Windows click More info → Run anyway; on macOS right-click the app → Open. The app installs Python and Maverick for you, then you're set.
If you already have Python 3.10+, install the published package instead of running a remote bootstrap script:
pipx install 'maverick-agent[installer]'
maverick initFor source-based desktop bootstrapping, download deploy/desktop/install.sh or deploy/desktop/install.ps1 from a commit or release you trust, verify it, and set MAVERICK_REF to a full 40-character commit SHA. The bootstrap scripts intentionally reject mutable branch/tag refs by default.
The PyPI distribution name is maverick-agent (the maverick name is squatted on PyPI). The [installer] extra pulls the wizard into the same pipx environment so maverick init resolves.
If you already installed the kernel without the extra, inject the wizard:
pipx inject maverick-agent maverick-installergit clone https://github.com/cdayAI/maverick
cd maverick
pip install -e ./packages/maverick-core
pip install -e ./apps/installer-cli
# Optional sister packages:
pip install -e ./packages/maverick-shield
pip install -e ./packages/maverick-channels
pip install -e ./packages/maverick-dashboard
pip install -e ./packages/maverick-mcp
maverick init # interactive wizard
maverick start "Plan a 2-week trip" # one-shot goal
maverick chat # interactive REPL
maverick dashboard # web UI at http://127.0.0.1:8765
maverick serve # channel server (Telegram/Discord/...)
maverick mcp # MCP server (Claude Code / Cursor)
maverick doctor # health check
maverick version # installed package versions| Command | What |
|---|---|
maverick init |
Interactive setup wizard with preflight + API-key validation |
maverick doctor |
Green / yellow / red health check + remediation hints |
maverick version |
Installed package versions + runtime info |
maverick config show / path / edit |
Show / locate / edit ~/.maverick/config.toml |
maverick start TITLE [--template NAME --param k=v] |
Run a goal once |
maverick chat |
Interactive REPL (each line = a goal) |
maverick serve |
Channel server (reads [channels.*] from config) |
maverick dashboard [--host --port --token] |
Local web UI + REST API |
maverick mcp |
MCP server on stdio for Claude Code / Cursor / etc. |
maverick logs / status / answer / resume |
Inspect + control running goals |
maverick fact / facts |
Get / set persistent facts |
maverick skills |
List installed + distilled skills |
maverick skill install / remove / info |
Manage the skill marketplace |
maverick template list / show |
Goal templates with {{ var }} substitution |
maverick budget |
Total + per-run cost history |
packages/
maverick-core/ Python agent kernel: recursive swarm, persistent world
model (SQLite + FTS5 + schema v8), 12 LLM providers, 7
sandboxes, MCP client, skills, templates, persona,
background runner, budget tracking
maverick-shield/ Agent Shield integration + built-in fallback rule set
maverick-channels/ 12 channel adapters: Telegram, Discord, Slack, Signal,
Email, Matrix, Bluesky, Mastodon, Voice, WhatsApp, SMS,
iMessage (WhatsApp/SMS need Twilio; iMessage is macOS-only)
maverick-dashboard/ Local FastAPI web UI + REST API at /api/v1 + OpenAPI
docs at /docs. Live progress streaming via short-poll.
maverick-mcp/ MCP server (stdio JSON-RPC) -- exposes Maverick to Claude
Code, Cursor, Claude Desktop as a tool. The agent kernel
can also CONSUME external MCP servers as its own tools.
apps/
installer-cli/ Interactive Python TUI wizard (`maverick init`)
installer-desktop/ Tauri-based GUI installer scaffold (signing in v0.2)
deploy/
docker/ vps/ desktop/ Dockerfile, install.sh, systemd unit, Caddyfile
docs/
getting-started.md Install + first run
configuration.md Full config schema reference
deployment.md Desktop / Docker / VPS / Phone-companion targets
safety.md Shield chokepoints and built-in rule set
api.md REST API reference + curl examples
benchmarks/
longhorizon/ Reproducible long-horizon evaluation tasks
example-skills/ Curated SKILL.md files for the marketplace
example-templates/ Reusable goal-template files
Maverick's kernel is Python, but its wire surface is the Model Context Protocol. Any MCP-speaking language can drive the swarm from outside Python:
- TypeScript / JavaScript → docs/clients/typescript-quickstart.md
- Go → docs/clients/go-quickstart.md
- Rust → docs/clients/rust-quickstart.md
- C# / .NET → docs/clients/csharp-quickstart.md
- Java / JVM → docs/clients/java-quickstart.md
Each is a 20-line program: spawn maverick mcp, list tools, call one.
Why this and not a separate @maverick/core port?
Language Bindings — Council Decision.
Run the swarm inside any repo's GitHub Actions — on a PR, a schedule, or on demand — under a hard spend cap:
- uses: cdayAI/maverick/deploy/github-action@v0.1.6
with:
goal: "Summarize this PR and flag anything risky."
max-dollars: "0.50"
anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}| Axis | Maverick |
|---|---|
| Target user | General consumer -- no AI expertise required |
| Wedge | Long-horizon depth + true multi-agent coordination |
| Safety | First-class. Every input, tool call, and output passes through Agent Shield. |
| Control | You pick the models. Per-role. Multi-provider. |
| Deploy | Desktop / Docker / VPS / Phone (12 channels) |
| Privacy | All detection runs locally. Your data never leaves your machine unless you choose a cloud LLM. |
MIT. See LICENSE.