artemis - Moses Yebei

artemis

Abstract

ARTEMIS turns natural-language instructions into reliable Android automation. It automates end-to-end workflows, captures logs for debugging and analysis, and integrates seamlessly with AI coding assistants such as Antigravity, Codex, and Claude Code. On the AndroidWorld benchmark, ARTEMIS achieves a success rate of over 99%.

The Problem

Traditional Android UI automation relies on brittle selectors (XPath/ID) that break with interface revisions, requiring constant maintenance. ARTEMIS addresses this fragility with a "Dynamic-First, Coordinate-Fallback" multimodal locating engine designed to remain resilient to UI redesigns and resolution drift.

What This Does

The core artemis/ package (531 Python files) processes natural-language instructions into executable test plans. It integrates via the Model Context Protocol (mcp_server/, 24 files), enabling assistants like Antigravity to drive physical test devices. A key architectural feature is the "Optimistic Asynchronous Pipeline," routed through artemis/__main__.py, which decouples UI interaction from heavy LLM reasoning to achieve 3–5s step execution. The system captures logcat stacks and keyframe screenshots for debugging and achieved 99%+ task completion on the AndroidWorld benchmark.

How It Is Wired

Control flows from an MCP request into the artemis execution engine. The mcp_server/ receives the natural language prompt and triggers plan generation. Execution is managed by the pipeline in artemis/, which interacts with the device and writes results to the database managed by apps/admin_console/database/repositories/. The apps/admin_console provides the administrative UI for task management, routing through apps/admin_console/routers/ and handling state via apps/admin_console/core/state.py. The mcp_server acts as the primary hub connecting external AI assistants to this internal flow, touching the device, the filesystem (screenshots), and the network (logcat capture).

How To Use It

Setup:

git clone https://github.com/moses-y/artemis.git
cp .env.example .env

Configuration: Required environment variables and keys are defined in .env.example. The playground/artemis_container/docker-compose.yml provides a pre-configured Docker setup for rapid deployment.

Running it:

# Via Docker Compose
docker compose up --build -d

# Or directly from source
python -m artemis

The entry point is artemis/__main__.py.

Real-World Use

A QA engineer describes a multi-app workflow via Antigravity: "Navigate to Google Maps, set a driving route to the airport, then open YouTube and play a specific artist." ARTEMIS generates a step-by-step plan, executes the actions across the Android device, and returns a diagnostic report containing logcat stacks and keyframe screenshots, enabling immediate triage without manual script maintenance.

Code Health & Issues

Static analysis of the repository structure found no red flags. The project includes 217 test files, a CI configuration at .github/workflows/ci.yml, an Apache 2.0 license (LICENSE), and package lockfiles (package-lock.json, pyproject.toml). All expected SDLC artifacts are present.

The Bottom Line

ARTEMIS is a capable framework for autonomous Android automation with strong benchmark performance and unique MCP integration for AI assistant workflows. The Python/TypeScript split and dual-container deployment profile indicate an enterprise-ready tool, though the setup complexity may be overkill for simple scripting needs. It is well-suited for teams requiring reliable cross-app testing without selector maintenance.

What the analyser found

Deployment readiness

7/7

Container image
CI pipeline
Lockfile committed
Test suite
README
License
No committed secrets

Composition

692 files

Python531
TypeScript40
Markdown25
JSON20
HTML11
SCSS10

ReactAngularDjangoFlaskExpressDocker

Architecture

Top-level areas of the codebase, sized by module count. Arrows show how many imports cross from one area into another.

graph TD g0["artemis 253 modules"] g1["tests 208 modules"] g2["apps 72 modules"] g3["mcp_server 22 modules"] g4["packages 8 modules"] g5["scripts 6 modules"] g6["docs 2 modules"] g1 -->|515| g0 g2 -->|46| g0 g1 -->|32| g2 g3 -->|24| g0 g1 -->|20| g3 g5 -->|19| g0 g0 -->|4| g3 g0 -->|2| g2 g2 -->|1| g3 g5 -->|1| g2

Module dependencies

The 10 most depended-upon modules of 572, from static import analysis. Red outline marks a module in an import cycle.

graph LR n0_artemis_context["context"] n1_artemis_utils_logger["logger"] n2_artemis_config___init__["__init__"] n3_artemis_graph_state["state"] n4_artemis_data_engine_trace["trace"] n5_artemis_tools_base["base"] n6_artemis_runtime___init__["__init__"] n7_artemis_drivers_base["base"] n8_artemis_tools_tool_wrapper["tool_wrapper"] n9_artemis_services_llm["llm"] n0_artemis_context --> n2_artemis_config___init__ n0_artemis_context --> n1_artemis_utils_logger n4_artemis_data_engine_trace --> n0_artemis_context n3_artemis_graph_state --> n1_artemis_utils_logger n9_artemis_services_llm --> n2_artemis_config___init__ n9_artemis_services_llm --> n0_artemis_context n9_artemis_services_llm --> n4_artemis_data_engine_trace n9_artemis_services_llm --> n1_artemis_utils_logger n5_artemis_tools_base --> n7_artemis_drivers_base n5_artemis_tools_base --> n1_artemis_utils_logger n8_artemis_tools_tool_wrapper --> n0_artemis_context n8_artemis_tools_tool_wrapper --> n4_artemis_data_engine_trace n8_artemis_tools_tool_wrapper --> n3_artemis_graph_state classDef cyc stroke:#ef4444,stroke-width:2px; class n0_artemis_context,n4_artemis_data_engine_trace cyc;

Dependencies

13 of 64 declared dependencies are a major version behind. Resolved 64 of 64 against the registry.

Package	Declared	Latest	Behind
`websockets`	>=12.0	17.0.1	5 major
`posthog`	>=3.5.0	7.39.1	4 major
`jasmine-core`	~5.6.0	7.0.1	2 major
`google-genai`	>=0.1.1	2.18.1	2 major
`rich`	>=13.7.0	15.0.0	2 major
`pillow`	>=10.0.0	12.3.0	2 major
`psutil`	>=5.9.0	7.2.2	2 major
`@types/jasmine`	~5.1.0	6.0.0	1 major

What would stop me shipping this

Ranked by severity × confidence × production reach. Reach is the honest discriminator across a collection that is mostly other people's code: the same finding matters more in something that ships.

high1

medium6

Pin third-party GitHub Actions to a commit SHAhigh

.github/workflows

astral-sh/setup-uv@v6

A tag can be moved, so the action running with your token and secrets is whatever its owner last pushed; this is how tj-actions/changed-files leaked secrets from thousands of repos.

Fix: Replace each @vN with the 40-character commit SHA, keep # vN as a comment, and let Dependabot bump the SHAs.

Enable Dependabot or Renovatemedium

3 manifest(s), no update bot configured

Without a bot a published advisory sits unpatched until someone audits by hand, which across 1,322 repositories means never.

Fix: Commit .github/dependabot.yml covering the repo ecosystems plus github-actions.

Pin the container base image by digestmedium3 occurrences

Dockerfile

node:22-bookworm-slim, python:3.12-slim, python:3.12-slim

An untagged or mutable base means today's build and last month's contain different libc and a different CVE set, with no record of which shipped.

Fix: Use image:tag@sha256:<digest> and enable Dependabot's docker ecosystem.

Gate pull requests on a dependency vulnerability scanmedium

.github/workflows

no dependency scan in CI

This is the one gate that would catch a known-vulnerable package before it reaches a build, and no repository in the sample had it.

Fix: Add dependency-review-action on pull_request, or osv-scanner on push and a schedule.

Move large binaries to Git LFS or out of the repositorymedium2 occurrences

tests/tools/inputs/recording.mp4

2 blobs over 5MB: recording.mp4 6.7MB, demo.gif 5.8MB

One repository here carries twenty blobs over 5MB including a 10.9MB spreadsheet, so every clone and every CI checkout pays for data nobody diffs.

Fix: Track those extensions with LFS, or move datasets to object storage and fetch them in a setup step.

Set persist-credentials: false on checkoutmedium

.github/workflows/ci.yml

checkout keeps the token, then dependencies are installed

The token stays in .git/config for every later step, so a malicious postinstall script reads a pushable credential without one ever being passed to it.

Fix: Add with: persist-credentials: false, and pass an explicit token only to the step that pushes.

Add a non-root USER to the imagemedium

Dockerfile

CMD or ENTRYPOINT with no USER directive

A process running as root in the container is root against every mounted volume, and it turns any container escape or writable-mount mistake from a contained problem into a host one.

Fix: Create an unprivileged user, chown what it needs, and end the Dockerfile with USER.

Checked deterministically against the repository tree and a bounded set of its files: committed credentials, unpinned actions and base images, missing lockfiles and update bots, workflows that discard failures, published advisories against the declared dependencies, runtime configuration, licensing and notebook reproducibility. No language model is involved in this section.