Resources¶
Official documentation¶
| Tool | Docs | Source |
|---|---|---|
| Claude Code | code.claude.com/docs · cost tracking | proprietary |
| Codex CLI | developers.openai.com/codex | openai/codex |
| OpenCode | opencode.ai/docs · agents · rules · skills · permissions | open source |
| OpenCode 2 | opencode.ai/v2/docs | open source |
| Cursor | cursor.com/docs | proprietary |
| Kilo Code | kilocode.ai/docs | Kilo-Org/kilocode |
| uv (Python projects) | docs.astral.sh/uv | astral-sh/uv |
Shared conventions (they work across tools)¶
- AGENTS.md — the open format for repository instructions read by most agents.
- Agent Skills — the
SKILL.mdfolder format; examples in anthropics/skills. - Model Context Protocol — how agents connect to external tools and data.
Guides and curated collections¶
- Claude Code best practices (Anthropic engineering blog).
- awesome-claude-code — curated commands, hooks, skills and workflows.
- claude-howto — worked examples, including memory and
CLAUDE.md. - addyosmani/agent-skills — reusable skills, with setup notes for OpenCode.
- ccusage — token and cost reports from local agent logs (
npx ccusage@latest daily).
Skills, MCP servers and hooks¶
- Genesis Skills — Genesis Mission skills for science and engineering (HPC, AmSC, literature, plasma simulation, data and evaluation); upstream on GitHub.
- anthropics/skills · openai/skills · K-Dense-AI/claude-scientific-skills — skill collections.
- modelcontextprotocol/servers · MCP Registry · awesome-mcp-servers — MCP servers.
- Claude Code hooks · Codex hooks · OpenCode plugins — hooks that run commands at fixed points.
- ChemGraph — computational chemistry agent with an MCP server.
- AtomisticSkills — computational chemistry skills.
- claude-replay — turn an agent session into an HTML replay.
- ai4s-skills — skill collection for AI-for-science work.
- claude-scholar — skill collection for research.
- More, with setup for each tool: Skills, MCP servers and Hooks.
Benchmarks: how to read a leaderboard¶
| Site | What it measures |
|---|---|
| SWE-bench Pro (Scale) | real issues in maintained repositories; one fixed scaffold for every model |
| Terminal-Bench | long, multi-step shell tasks run inside real agents |
| SWE-rebench | fresh GitHub issues every month, to limit training-data leakage |
| SWE-bench-Live | continuously added tasks in many languages |
| Vals.ai | independent re-runs with cost and time per task |
| Artificial Analysis | the same tests on every model, with price and cost per task |
| ResearchClawBench | end-to-end automated research tasks (chemistry, physics and more) |
Compare scores only when the harness and the data split are the same.
DOE and ALCF resources¶
- AmSC Model Access Gateway (MAG) — model access for Genesis Mission teams: account, project, personal access token, and setup for Claude Code and Codex CLI.
- argo-proxy — multi-format gateway to ARGO for ANL-internal users; translates OpenAI, Anthropic, and Google GenAI formats.
- argo-shim — lightweight HTTP proxy + SSH tunnel for Claude Code and Codex CLI on ALCF machines.
- ALCF Inference Endpoints — open-weight models on ALCF hardware (Sophia, Metis, Minerva); Globus auth; web UI and API access.
- Custom endpoints guide — setup walkthrough for argo-proxy, argo-shim, and ALCF Inference with multiple tools.
Tutorial container¶
ghcr.io/autonomous-scientific-agents/agentic-coding— Claude Code, Codex CLI, OpenCode, uv, gh, vim, emacs and VS Code in the browser. Instructions: Run in a container (Dockerfile,run.shand.devcontainer/in this repository).
Evidence worth reading¶
- CodeRabbit, State of AI vs. Human Code Generation (Dec 2025): where agent-written pull requests drew more review findings. A vendor study; read the methodology.