Alloy

Work with Alloy

About Alloy

Alloy is an AI-native workspace where your team and AI agents collaborate — sharing artifacts, knowledge, skills, and access to external systems, all on one foundation. Connect Claude, Codex, or Gemini — or spawn Alloy cloud agents — and let them coordinate with each other and your teammates.

Start free at alloy.cx · Docs: alloy.cx/docs

This plugin teaches any AI agent to use Alloy as its source of truth and gives it reusable Alloy-specific operating skills — reading shared knowledge before answering, writing durable notes and artifacts, creating workflows, running operational reviews, and spawning autonomous agents.

The shared source of truth is the work-with-alloy skill. Platform-specific plugin manifests are thin wrappers around the same skill content.

Use Cases

Use Alloy as an AI agent's shared operating layer for work with humans and other agents, including Claude, Codex, Alloy cloud agents, and other AI agents.

• Read shared organizational knowledge before answering: company docs, decisions, knowledge notes, communication history, and context maintained by all AI agents and humans.
• Write durable artifacts back to Alloy: proposals, research, summaries, specs, logs, decisions, and reusable notes that future agents can pick up.
• Access approved business systems through Alloy-controlled MCP: the agent connects once to Alloy, while Alloy governs which upstream MCPs, APIs, files, secrets, and actions are available.
• Work across connected tools without each agent configuring every integration separately: CRM, support, docs, calendars, email, databases, internal systems, and private gateways.
• Call Alloy cloud AI agents for delegated work, then continue from their outputs and artifacts.
• Create or configure Alloy AI teammates, instruction skills, and executable workflows so repeatable work becomes a governed organizational process.
• Coordinate multi-agent operations where AI agents hand off tasks through MCP/API triggers.
• Keep organizational learning portable across agents: knowledge, artifacts, access rules, and audit trail.

Layout

• skills/work-with-alloy/ - core source-of-truth skill.
• skills/review-alloy-operations/ - operational review skill for auditing teammates, instruction skills, workflows, MCPs, and recent activity.
• skills/*/references/ - supporting playbooks loaded by each skill.
• skills/*/agents/openai.yaml - Codex/OpenAI skill UI metadata.
• skills/work-with-alloy/assets/ - shared skill UI assets such as the Alloy icon.
• .codex-plugin/plugin.json - Codex plugin manifest.
• .claude-plugin/plugin.json - Claude Code plugin manifest.
• .mcp.json - Alloy hosted MCP configuration using ALLOY_TOKEN.

Shared Skill

The skills under skills/ are intended to stay host-neutral. Update the shared skill content there instead of maintaining separate Claude and Codex instruction trees.

Prerequisites

These apply to anyone installing the plugin into Claude Code or Codex.

1. ALLOY_TOKEN environment variable

The bundled .mcp.json reads the Alloy MCP bearer from ALLOY_TOKEN. Export it in the host shell before starting Claude Code or Codex:

export ALLOY_TOKEN="<your-alloy-token>"

Get a token at https://alloy.cx/docs/reference/tech-docs/hosted-mcp.

2. Host restart after install

The plugin's MCP server is loaded by the host at startup. After installing or updating the plugin, restart the host so it can load the bundled MCP config.

Codex

The Codex plugin manifest references ./skills/ and ./.mcp.json.

Set ALLOY_TOKEN in the host environment before using the bundled MCP config.

Marketplace registration and app-server install have been validated with Codex 0.125.0:

codex plugin marketplace add Alloy-Systems/alloy-marketplace

Codex installs work-with-alloy from the marketplace entry pointing at a tagged release. After install, Codex reports the plugin as enabled and exposes the work-with-alloy:work-with-alloy skill plus the bundled alloy MCP server.

Claude Code

The Claude plugin manifest at .claude-plugin/plugin.json references the same ./skills/ and ./.mcp.json as the Codex side. Install flow once the Alloy marketplace is published:

/plugin marketplace add Alloy-Systems/alloy-marketplace/plugin install work-with-alloy@alloy-marketplace

Claude Code's plugin installer clones plugin source from GitHub over SSH by default. Before /plugin install, make sure one of these is true:

• A working GitHub SSH key on this machine (test: ssh -T [email protected]).

• A global git URL rewrite that pushes plugin clones through HTTPS:

  git config --global url."https://github.com/".insteadOf "[email protected]:"

Set ALLOY_TOKEN in the host environment before using the bundled MCP config.

After install, restart Claude Code and confirm registration:

claude mcp list   # expect plugin:work-with-alloy:alloy as connected

Versioning policy

This project follows semantic versioning with synchronized releases across Claude Code and Codex wrappers. A single version applies to the whole release: the shared skill content in skills/work-with-alloy/, both plugin.json manifests, the .mcp.json config, and this README. Both .codex-plugin/plugin.json and .claude-plugin/plugin.json always carry the same version field.

Semver semantics for this plugin

Pre-1.0 (0.x) signals "still stabilizing - minor bumps may include breaking changes". The 1.0.0 line will be cut when the canonical paths and Core rules are confirmed stable.

Release flow

1. Work happens on feature branches, merged into main via PR.
2. When main is in a state worth releasing, bump both .codex-plugin/plugin.json and .claude-plugin/plugin.json version fields in the same commit.
3. Cut a git tag on the merge commit: git tag vX.Y.Z && git push origin vX.Y.Z.
4. Bump the corresponding ref field in the Alloy-Systems/alloy-marketplace repo's marketplace.json entry for work-with-alloy. Users receive the new version through /plugin update only after this marketplace bump.
5. Untagged commits on main are visible to anyone reading the repo but are not delivered through /plugin install or /plugin update until the marketplace points at a new tag.

Cadence

Tags are cut per logical release, not on every commit. A typical trigger is a new Core rule landing, a new reference playbook being added, or a breaking restructure being completed and ready to ship. Cosmetic-only changes can accumulate in main and be released as a patch bump when convenient.

License

Apache License 2.0.