// agents · google/mantis

mantis

Google's modular toolkit that lets AI coding agents hunt, reproduce, and patch vulnerabilities autonomously.

★ 2.4k stars⑂ 201 forksApache-2.0updated 2026-10

▦ visual — this project's README ships no embeddable screenshot,
so here's a placeholder. The guide below is complete regardless.

// project facts

Project facts

mantis is an open-source ai agents project by google (google/mantis on GitHub). It has 2,380 GitHub stars and 201 forks, is released under the Apache-2.0 license, and was last updated on GitHub on 7 October 2026.

Repository: google/mantis
GitHub stars: 2,380
Forks: 201
License: Apache-2.0
Last updated on GitHub: 7 October 2026
Category: AI Agents
Listing: Full install guide
Official site: Official site

// how it works

How it works

Clarify your goal
Decide which codebase or file you want reviewed, and prepare an isolated, restricted environment with no production access, as the README strongly requires.
Get the project
Clone the repository and move into the reference harness directory.
Install and configure
Run the install script, authenticate Google Cloud Application Default Credentials if you use Vertex AI, then let the configuration tool auto-detect capabilities or use its interactive wizard.
Run your first task
Run the preflight validation, then launch a review campaign against a file or repository, optionally focusing it with a plain-language hint.

// features

Why people love it

Threat modelling and code indexing

Mantis reviews codebase history for past vulnerabilities, builds a semantic index or summaries for navigation, and generates a threat model before scanning.

Hypothesis-driven agent research

Individual agents research hypotheses or scan files systematically, then deduplicate the findings they produce.

Triage and calibration against hallucination

Findings are critiqued, statically checked for production viability, and calibrated with a rubric so severity is not inflated.

Reproduction attempts

Suspected vulnerabilities are reproduced at whatever depth the environment allows, from a reasoned exploit sketch to a unit test or a mock server.

Adversarial patch verification

Mantis drafts patches and uses an adversarial loop to check the vulnerability is genuinely fixed.

Custom research graphs and focus hints

You can steer a run with --focus, remove spend limits for long unattended runs with parallel workers, or have it design a new agent-graph topology from an objective.

// run it

Run it yourself

Clone and install the reference harness
Install the venv tooling first if needed, then run the bundled installer.
```
git clone https://github.com/google/mantis.git
cd mantis/reference && ./install.sh
```
Configure and validate
Authenticate ADC for Vertex AI, auto-configure, and run the fast preflight plus reachability probe.
```
gcloud auth application-default login
python3 scripts/configure.py --auto
python3 scripts/configure.py --test --probe
```
Launch a review campaign
Point run.sh at a file or repository, optionally with a focus hint.
```
./run.sh path/to/code
./run.sh path/to/code --focus "look for IDOR"
```

// use cases

What you can build

Running an agent-driven security review over a service before a release
Checking a codebase for repeats of vulnerability patterns found in its history
Triaging and reproducing suspected findings before a human files a report
Adapting the included skills to specialised domains such as firmware, infrastructure as code, or ML pipelines

// faq

Questions, answered

Who maintains it?

It is developed by Google and published under the google GitHub organisation, though the README notes it is not an officially supported Google product.

Is it free to use?

Yes, the code is Apache-2.0 licensed and free to use and adapt. Running it still costs whatever the underlying AI models charge.

Does it need a paid API key?

In practice yes: the reference harness is built around Google Cloud and Gemini/Vertex AI access, so you need Google Cloud credentials (ADC) and you pay for model usage, with budget controls in the tool.

How does it compare to a well-known alternative?

Traditional SAST tools like CodeQL or Semgrep match deterministic rules; Mantis uses agents to form hypotheses, attempt reproductions, and draft patches, so it explores behaviour rules cannot express, but its findings need human verification.

How do I get started?

The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if mantis is not the right fit.

openclaw★ 391.6k
A personal AI that really does things — on any OS, with your data staying yours.
superpowers★ 294.9k
An agentic skills framework & software development methodology that works.
ECC★ 275.2k
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
hermes-agent★ 252.0k
The agent that grows with you — a Python framework for adaptive AI agents.
deepseek-harness★ 245.5k
DeepSeek's plugin-based agent harness where everything is a plugin.

Get mantis

Free and open — grab the source, star the repo, and start creating.

View on GitHub ↗← More AI Agents tools