Open Source Coding Agents, Compared

12 reviewed agents compared

Every agent below ships under a licence GitHub could identify, and every one carries the same FARS review, so the numbers are comparable rather than borrowed from each project's own README. The table leads with the review; the sections under it carry each agent's own positioning, including where its documentation says it is not the right fit.

# Agent FARS Trust Stars License Language Last commit
1 elizaOS 88 24/25 19.2k MIT TypeScript 2026-09-02
2 Unlazy Completion Discipline 86 25/25 3.0k MIT JavaScript 2026-08-29
3 Better Harness 85 23/25 2.1k MIT JavaScript 2026-09-02
4 Apache Maka 82 22/25 4.6k Apache-2.0 TypeScript 2026-09-02
5 JS Reverse MCP 82 21/25 2.7k Apache-2.0 TypeScript 2026-08-25
6 NotebookLM Python Automation 81 21/25 19.1k MIT Python 2026-09-02
7 ADK Agent Recipes ✓ Google 81 23/25 10.3k Apache-2.0 Python 2026-09-01
8 Kungfu 81 18/25 4.5k Apache-2.0 C++ 2026-09-02
9 Osaurus 80 22/25 7.8k MIT Swift 2026-09-02
10 Chrome DevTools MCP 79 17/25 50.6k Apache-2.0 TypeScript 2026-09-02
11 LangWatch 78 19/25 3.5k Apache-2.0 TypeScript 2026-09-02
12 Harmonist 78 21/25 2.3k MIT Python 2026-06-09
How this ranking is built

The order is the FARS total; stars only break ties, because popularity measures attention rather than whether something holds up. Two limits are worth knowing before you read the order. These are static reviews — source, documentation and repository metadata, with no runtime reproduction — which caps reliability, effectiveness and verifiability and puts the achievable maximum at 77 rather than 100. And a gap of a few points sits inside the noise of that method. Read the table as a shortlist worth investigating, not as a verdict; every score links to its full dimension breakdown and the evidence behind it. Review method →

1. elizaOS

Where it differs

Compared with the optional Eliza Cloud path, the local runtime and direct-provider configuration remain first-class options: Eliza Cloud adds accounts, authentication, hosted model routing, application and agent deployment, remote connectivity, and cross-device services, while the other paths preserve more direct runtime and provider control. For whole-device delivery, this monorepo retains Eliza application shells and native runtime bridges, whereas the separate elizaOS/os repository owns bootable Linux and AOSP distributions, installers, release manifests, and OS toolchains.

Full review → GitHub →

2. Unlazy Completion Discipline

Full review → GitHub →

3. Better Harness

Full review → GitHub →

4. Apache Maka

Full review → GitHub →

5. JS Reverse MCP

Where it differs

Unlike exposing the Chrome DevTools API to a model unchanged, JS Reverse MCP groups scripts, breakpoints, traffic, WebSockets, state, and local files into primitives for sustained analysis. The default Patchright mode uses the installed Google Chrome and retains access to the Web Store, extensions, and sync; --cloak instead uses a CloakBrowser binary with C++-level fingerprint patches but no Google services or Web Store. The project positions anti-detection as support for the debugging chain, not as a general-purpose crawling framework.

Full review → GitHub →

6. NotebookLM Python Automation

Where it differs

Compared with the Gemini Notebook web interface, notebooklm-py emphasizes scripted control and local extraction. It supports batch downloads, structured quiz and flashcard exports, mind-map JSON, CSV data tables, source full text, individual slide revision, and programmatic sharing. That additional reach comes with greater maintenance risk because the client relies on undocumented endpoints rather than an official Google SDK.

Full review → GitHub →

7. ADK Agent Recipes

Full review → GitHub →

8. Kungfu

Where it differs

Compared with manually copying chats and re-explaining decisions when moving among Codex, Claude, OpenCode, or Amp, Kungfu keeps the Work objective, evidence, progress, and Attempt history outside those sessions. It complements rather than replaces these agents, and its TUI and GUI are optional sidecar views rather than substitutes for native agent consoles.

Full review → GitHub →

9. Osaurus

Full review → GitHub →

10. Chrome DevTools MCP

Where it differs

Full mode targets deep debugging with automation, performance, network, console, and optional specialist categories; --slim exposes only navigation, script execution, and screenshots for basic browser work and a smaller tool surface. The MCP server is intended for structured integration with compatible coding clients, while the included CLI covers use without MCP. Launching a dedicated Chrome profile provides separation from everyday browsing, whereas attaching to an existing Chrome preserves logins and manual testing state at the cost of greater data-exposure risk.

Full review → GitHub →

11. LangWatch

Where it differs

Compared with assembling separate tracing, dataset, evaluation, and prompt-optimization systems, LangWatch explicitly combines them into a trace → dataset → evaluate → optimize → re-test loop. Its OpenTelemetry foundation also allows compatible tracing libraries beyond the directly named framework integrations.

Full review → GitHub →

12. Harmonist

Where it differs

Harmonist positions itself as a project-local governance layer for AI-assisted coding, not as a replacement for LangChain, CrewAI, AutoGen, or MetaGPT. The repository describes those projects as orchestration primitives for building agentic applications, while Harmonist uses IDE hooks to govern review, memory, and completion behavior in the coding workflow; the two approaches can coexist.

Full review → GitHub →