One shared hub for every AI coding agent on your team.
Arsenal is a template repository that centralizes agent skills, engineering standards, model routing, and work tracking in one repository that Claude Code, Codex CLI, OpenCode, and GitHub Copilot all read. Clone it once, point every harness at it, and stop re-teaching each tool how your team works.
Teams adopt AI coding agents piecemeal. Each harness brings its own config file, its own skill format, and its own habits. Within a quarter you have four half-overlapping rulebooks, model spend nobody planned, tests the agents quietly stopped running, and work tracked in whatever tool the loudest engineer prefers. The agents are not the problem. The missing shared infrastructure is.
Arsenal is that infrastructure:
- One source of truth. Standards, workflows, and model policy live in
docs/; skills live in.agents/skills/and role agents in.claude/agents/. Root-level adapter files (AGENTS.md,CLAUDE.md,.github/copilot-instructions.md,opencode.json) point every harness at it. Guidance is written once and read everywhere. - A delivery loop with gates. Research, plan, execute, verify, track, publish. Each phase has entry criteria, a defined owner (human or agent tier), and a gate. Code does not move forward through a red test suite.
- Model routing by tier. Strategist models plan and review. Executor models
write code. The policy and the per-harness wiring live in
docs/models/routing.md, so the expensive models spend their tokens where they change outcomes. - Work tracking that follows the work. One command,
scripts/work-item.sh, creates and updates work items in GitHub Issues, Jira, or Azure DevOps. Agents stamp the IDs; the tracker stays truthful. - The Superpowers methodology, pinned. The open-source Superpowers skill
library (brainstorm, plan, TDD, subagent development) is installed per
harness from a pinned version recorded in
docs/superpowers/. - Validated by CI. Skill frontmatter, internal links, shell scripts, and the tracking script's own tests are checked on every push, so the hub cannot rot quietly.
| Harness | Entry file | Skills wiring | Guide |
|---|---|---|---|
| Claude Code | CLAUDE.md (imports docs/) |
.agents/skills via the committed plugin in .claude-plugin/ |
docs/harness/claude-code.md |
| Codex CLI | AGENTS.md |
.agents/skills (native) |
docs/harness/codex.md |
| OpenCode | AGENTS.md + opencode.json |
.agents/skills (native) |
docs/harness/opencode.md |
| GitHub Copilot | .github/copilot-instructions.md (+ AGENTS.md for the coding agent) |
.agents/skills (native) |
docs/harness/copilot.md |
Skills and role agents live once, in real folders: .agents/skills/ (the
cross-tool standard location) and .claude/agents/. Codex, Copilot, and
OpenCode read the skills there natively; Claude Code reads them through a
small committed plugin (.claude-plugin/, .claude/settings.json) that
points at the same folder. Claude Code and Copilot read the agents
natively, OpenCode through opencode.json pointers, and Codex from TOML in
.codex/agents/ that scripts/sync.sh generates and CI checks for drift.
No symlinks and no copies, so a clone works the same on macOS, Linux, and
Windows with no setup.
Any harness that reads AGENTS.md or the Agent Skills standard works with the
same pattern. docs/harness/README.md shows how to add the next one.
Skills and agents work from the clone: they are plain committed folders, so there is nothing to install (Claude Code asks once to trust the folder, which enables its plugin). The checklist:
- Create your repo from this template (green "Use this template" button)
or
git clone https://github.com/Alex5350/arsenal.git your-repo. - Run
scripts/adopt.sh <project-name>(--dry-runpreviews). It renamesarsenalto your name in the three Claude Code plugin files, so two projects on one machine never share a plugin registration, and copies the starters that are not there yet:docs/templates/app-ci.ymlto.github/workflows/app.ymlanddocs/templates/work-item.example.envto.env. It also replaces arsenal's ownAGENTS.mdwithdocs/templates/AGENTS.example.md(a consumer-writtenAGENTS.mdis left alone); fill in the placeholders. - Implement
make test-unitandmake test-integrationfor your stack; the inherited Makefile defines them for the hub's own checks, so replace or extend those recipes (testing standard). - Pick a tracker in
.env(WORK_BACKEND=gh|jira|ado). On GitHub, setGH_REPOor leave it unset and the script reads theoriginremote. - Install Superpowers in each harness you use, per the pinned matrix in
docs/superpowers/. make verify, then open the repo in any harness and start with "Read AGENTS.md and begin."
Optional: scripts/bootstrap.sh --user makes the skills available in your
other projects too (--dry-run to preview, --remove to undo).
Humans should start at docs/README.md. Deep design notes
live in TECHNICAL.md, and terminology in
GLOSSARY.md.
Six phases, each with a gate. Strategist models own research and planning,
executor models own implementation, and the reviewer tier owns verification
evidence. Full definitions in docs/workflows/.
docs/
README.md hub map and reading order
ARCHITECTURE.md how the hub is wired, and why
harness/ per-harness adapter guides (4 + how to add more)
standards/ git, testing, security, code review, documentation
workflows/ the six-phase delivery loop, phase by phase
models/ tiered model routing policy and per-harness wiring
skills/ skill authoring rules (library in .agents/skills/)
agents/ role agent contract (agents in .claude/agents/)
integrations/ GitHub Issues, Jira, Azure DevOps
superpowers/ pinned Superpowers methodology layer
templates/ starter CI and release workflows, tracking config, AGENTS.md
scripts/
adopt.sh rename the plugin and copy the starters into a new repo
sync.sh generate the Codex agent TOMLs (--check in CI)
bootstrap.sh optional: wire skills into your other projects
work-item.sh one command, three tracking backends (tests in tests/)
validate-skills.sh SKILL.md frontmatter linter (CI)
validate-manifests.sh plugin, marketplace, settings, opencode.json checker (CI)
check-links.sh internal markdown link checker (CI)
tests/ stubbed-CLI tests for work-item.sh (make test-unit)
Arsenal commits to the three standards the industry converged on rather than to any vendor:
- AGENTS.md for tool-agnostic instructions
- Agent Skills (
SKILL.md) for portable procedural knowledge - MCP for tool and data connections
When a harness changes, the adapter changes. The hub does not.
MIT. See LICENSE. The Superpowers skill library it installs is MIT work by Jesse Vincent and contributors; arsenal references and pins it but does not redistribute it.