AI coding agent comparison matrix
Every agent we track, in one table you can filter. 13 systems across 15 attributes — including the memory and CPU figures for the 5 we have actually run on a server, rather than read off a README.
How to read this. A blank cell means the project does not state it and we did not guess — it is not a “no”. Public facts come from each project’s own repository and carry the date we read them; the RAM and CPU figures are ours, measured on real servers. We host 5 of these 13, marked with a bolt, and the sort order never takes that into account.
Showing 13 of 13 agents
| Agent | License | Language | Interface | Model-agnostic | Providers | Local models | MCP | Code execution | Browser/computer use | Self-host difficulty | Min RAM (measured) | Idle RAM (measured) | Deploy | Pricing | Maturity |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OpenClaw | MIT | TypeScript | CLI, Desktop app, API | Yes | Anthropic, OpenAI, Other | Via proxy | Native | Native | Native | Moderate | 1 GB | 278 MB | Native, Binary | Free / OSS | Beta |
| OpenCode | MIT | TypeScript | CLI, TUI, Desktop app, Web UI | — | — | — | — | Native | — | Easy | — | — | Native, Binary | Free / OSS | — |
| Hermes Agent | MIT | Python | CLI, Desktop app, API | Yes | OpenAI, OpenRouter, Other | Via proxy | Native | Native | — | Easy | 1 GB | 120 MB | Native | Free / OSS | Beta |
| LangChain | MIT | Python, TypeScript | SDK, API | Yes | OpenAI, Anthropic, Google, Azure OpenAI, AWS Bedrock, Mistral, Cohere, Ollama, Other | Full | Native | Via plugin | Via plugin | — | — | — | Native, Docker, Serverless, Kubernetes | OSS + paid cloud | Production |
| Claude Code | Proprietary | — | CLI, SDK, IDE extension | No | Anthropic, AWS Bedrock, Google | None | Native | Native | Via plugin | Easy | 1 GB | None resident | Native, Binary | Usage-based | Production |
| Pi | MIT | TypeScript | CLI, TUI, SDK | Yes | OpenAI, Anthropic, Google, Other | — | — | Native | — | Easy | — | — | Native, Binary, Docker | Free / OSS | — |
| Codex | Apache-2.0 | Rust | CLI | No | OpenAI | None | — | Native | None | Easy | 1 GB | None resident | Native, Binary | Usage-based | Production |
| OpenHands | MIT | Python | Web UI, CLI, API | Yes | Anthropic, OpenAI, Google, Other | Full | Native | Native | Native | Moderate | 2 GB | 846 MB | Docker, Kubernetes, Native | OSS + paid cloud | Production |
| Cline | Apache-2.0 | TypeScript | IDE extension, CLI | Yes | Anthropic, OpenAI, Google, AWS Bedrock, Azure OpenAI, Groq, OpenRouter, Ollama, Other | Full | Native | Native | Native | Easy | — | — | Native | Free / OSS | Production |
| goose | Apache-2.0 | Rust | CLI, Desktop app, API | Yes | Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure OpenAI, AWS Bedrock, Other | Full | Native | Native | — | Easy | — | — | Binary, Native | Free / OSS | — |
| Aider | Apache-2.0 | Python | CLI | Yes | OpenAI, Anthropic, Google, Mistral, Cohere, Ollama, OpenRouter, Other | Via proxy | — | Native | None | Easy | — | — | Native | Free / OSS | Production |
| Grok Build | Apache-2.0 | Rust | TUI, CLI, IDE extension | — | — | — | Native | Native | — | Easy | — | — | Binary, Native | Free / OSS | — |
| Cursor CLI | Proprietary | — | CLI | — | — | — | — | — | — | — | — | — | — | Paid license | — |
Head-to-head
The matrix answers “which of these fits my constraints”. When the choice is down to two and the specs do not separate them, these pages argue it out:
Claude Code vs Codex
Source availability, not capability. Codex ships an open client; Claude Code does not. Both still send your work to one company's API, which is why neither is the answer if vendor independence is the goal.
Read the verdictHermes Agent vs Claude Code
Resident versus invoked. We measured Hermes idling at 120 MB on a 1 vCPU / 1 GB server - a process waiting for work. A box with Claude Code installed sits at its OS baseline with nothing resident, because there is nothing running between sessions.
Read the verdictOpenClaw vs Hermes Agent
Measured cost, not capability. On the same 1 vCPU / 1 GB server: Hermes idles at 120 MB and takes about five minutes to install; OpenClaw idles at 278 MB and installs inside a minute. OpenClaw also grows to fill a bigger box - 293 MB on a 4 GB tier - so its footprint tracks the machine you give it.
Read the verdictOpenCode vs Claude Code
Who can take it away. OpenCode's source is yours and its provider is your choice. Claude Code is a product you are licensed to use, on models from one company. Google retiring Gemini CLI in June 2026 and replacing it with a closed successor is what makes that abstract-sounding difference concrete.
Read the verdictgoose vs Claude Code
Governance. Foundation stewardship means no single company can retire it, relicense it, or change its terms unilaterally. That reads like a technicality right up until it is not - which is exactly what Gemini CLI users found out this year.
Read the verdictOpenCode vs goose
Community size against governance guarantees. OpenCode wins on people; goose wins on who is holding the keys. Both are permissively licensed, so neither locks you in.
Read the verdictWhere these numbers come from
- Public attributes — licence, language, stars, interfaces — read from each project’s repository and documentation, with the date recorded per cell.
- Memory and CPU are ours. We provision a server, install the agent through our own one-click preset, and read memory thirty seconds after a live model request. Only agents we host carry these figures.
- No quality ranking. Nothing here says which agent writes better code. We have not benchmarked that, and a matrix that implied an answer would be the least trustworthy thing on the page.
- Blank is honest. Where a project does not state something, the cell stays empty rather than being filled with a plausible guess.