wshobson/agents
TL;DR
wshobson/agents is Seth Hobson's MIT-licensed plugin marketplace of subagents, skills, slash commands and multi-agent workflow commands, written once in Markdown for Claude Code and converted by adapters for Codex, OpenCode, Cursor, Copilot, Antigravity and Pi. It suits developers who want ready-made specialist agents inside the coding tool they already use.
Key facts
| Type | Orchestrator |
|---|---|
| Languages / SDKs | Markdown, Python |
| License | MIT |
| Pricing model | Open source, free |
| Orchestration pattern | Pipeline |
| GitHub stars | 40,104 (as of 2026-09-30) |
| GitHub forks | 4,281 |
| Last push | 2026-09-29 |
| Latest release | No release published |
| Repository | wshobson/agents |
| Website | sethhobson.com |
| Documentation | github.com |
| Last verified | 2026-09-30 |
Key features
- Claude Code plugin marketplace:
/plugin marketplace add wshobson/agents, then install single-purpose plugins that each bundle their own agents, commands and skills. (source) - Adapters generate harness-specific files from the same
plugins/source for Codex CLI, Cursor, OpenCode, Antigravity CLI, Copilot and Pi, with a published capability matrix. (source) - Workflow orchestrator commands chain specialist subagents in a fixed order, for example
/full-stack-orchestration:full-stack-featurerunning architecture, database, frontend, test, security, deployment and observability agents. (source) - The agent-teams plugin spawns preset teams (review, debug, feature, research, security, migration) on Claude Code's experimental Agent Teams feature. (source)
- Agents declare a model tier (opus, sonnet, haiku or inherit), which the adapters map to each harness's own model names. (source)
- Skills can be installed one at a time into any supported agent with
gh skill installornpx skills add, without cloning or adding the marketplace. (source) - plugin-eval scores plugin and skill quality with a static lint plus optional LLM-judge and Monte Carlo layers. (source)
Architecture and orchestration pattern
Pattern: Pipeline
The repository is content, not a runtime. Each folder under plugins/ holds Markdown agent definitions, slash commands and skills plus a plugin manifest; Claude Code reads this tree directly through .claude-plugin/marketplace.json. Python adapters under tools/adapters/ translate the same source into other harness formats: Codex manifests are committed, while OpenCode, Antigravity, Copilot and Pi trees are generated locally with make generate and kept out of git.
Orchestration runs inside the host agent. Orchestrator plugins are slash commands whose instructions tell the host to call specialist subagents one after another, each step writing its output to a file before the next begins. The agent-teams plugin instead uses Claude Code's experimental Agent Teams, with a team lead and parallel reviewers, debuggers or implementers. On harnesses without a spawn tool the adapters degrade the instructions; Codex, for example, needs the agent named in prose to delegate.
There is no memory store of its own. Orchestrator commands such as full-stack-feature keep progress in a state.json file and step outputs in a working folder, which lets a later session resume. Cross-session memory is left to Pensyve, an external plugin listed in the marketplace.
Human in the loop
The orchestrator commands include explicit approval points. full-stack-feature, for example, gathers requirements one question at a time, stops at each PHASE CHECKPOINT and offers approve, revise or pause (pause saves progress to state.json), and halts to ask the user when any step fails. When it finds an in-progress session it asks whether to resume or start fresh. The governance category adds a review-agent-governance plugin for human approval before agents post reviews or comments, merge, or write to CI. Tool access can be limited per agent through tools: allowlists, which the OpenCode adapter turns into a permission: block.
Harnesses it can drive
- Claude Code (evidence)
- Codex (evidence)
- OpenCode (evidence)
Protocols
| Protocol | Support | Note |
|---|---|---|
| MCP | Unknown | The repo ships no MCP server or client of its own (no .mcp.json in the file tree); the harness matrix lists MCP only as a host capability, and the external Pensyve plugin is described as MCP-backed. |
| A2A | Unknown | No mention of A2A / Agent2Agent in the README, docs/harnesses.md, docs/architecture.md, docs/usage.md, docs/plugins.md or the repository file tree. |
| AG-UI | Unknown | No mention of AG-UI in the README, docs or repository file tree; not listed in the AG-UI README. |
Best for
- Claude Code users who want language- or domain-specific subagents and skills, installed one plugin at a time. (shortlist)
- Developers who switch between Claude Code, Codex and OpenCode and want the same agent definitions in each. (shortlist)
- Multi-phase feature work where the agent must stop for approval between design, implementation and deployment. (shortlist)
- Adding policy checks or signed audit records around agent tool calls through the governance plugins.
Not for
- Building agents in code; this is prompt and configuration content for existing coding agents, not an SDK or runtime.
- Setups that need tagged releases to pin; the repository publishes none and installs track main.
- Codex users who depend on long skills or per-agent tool restrictions: skills over 8 KB are truncated and tool allowlists are dropped.
Quickstart
/plugin marketplace add wshobson/agents# Inside Claude Code
/plugin marketplace add wshobson/agents
/plugin install full-stack-orchestration
/full-stack-orchestration:full-stack-feature "user settings page" --stack react/fastapi/postgres
# Parallel review team (needs CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
/plugin install agent-teams@claude-code-workflows
# Codex CLI, from a terminal
npx codex-marketplace add wshobson/agents
# OpenCode: clone, generate and symlink into ~/.config/opencode/
gh repo clone wshobson/agents ~/agents && cd ~/agents
make install-opencode
# A single skill into any supported agent
npx skills add wshobson/agents --skill python-testing-patterns -a claude-code
Common pitfalls
- The repository publishes no GitHub releases, so every install tracks
main. - OpenCode, Antigravity and Pi have no one-step install: clone the repo and run
make generate/make install-<harness>; the generated trees are gitignored and must be regenerated after pulling changes. - Codex truncates skills over 8 KB at load, converts slash commands to skills, and ignores per-agent
tools:allowlists (a read-only sandbox heuristic is used instead). - Skills-only installers (
gh skill,npx skills) bring no agents, commands or hooks. - The agent-teams plugin needs
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. - Installing Pi skills globally with
make install-piand also runningpiinside the clone makes every skill appear twice.
Pros
- Installing a plugin loads only that plugin's agents, commands and skills into context, not the whole marketplace. (source)
- One Markdown source produces native artifacts for seven harnesses, with the per-harness differences documented in a capability matrix. (source)
- Orchestrator commands write each step to disk and track progress in state.json, so runs can pause at checkpoints and resume later. (source)
- Includes governance plugins for policy enforcement, signed audit trails and human approval of agent writes. (source)
- MIT licensed with no paid tier; individual skills can be installed without adopting the whole marketplace. (source)
Cons
- No tagged releases: installs follow the main branch, so content can change between two installs. (source)
- Support outside Claude Code is degraded in documented ways, such as Codex's 8 KB skill cap, dropped tool allowlists in Codex and Cursor, and no TodoWrite outside Claude Code and OpenCode. (source)
- OpenCode, Antigravity and Pi require a local clone plus a make-based generate step instead of a marketplace install. (source)
- The agent-teams plugin depends on Claude Code's Agent Teams feature, which is experimental and must be switched on with an environment variable. (source)
- plugin-eval's LLM-judge and Monte Carlo layers are marked experimental and not validated against human labels. (source)
Alternatives
FAQ
Does wshobson/agents support MCP?
It ships no MCP server or client of its own. MCP servers are a feature of the host harness, and the external Pensyve memory plugin in the marketplace is described as MCP-backed.
Which coding agents can use it?
Claude Code natively, Codex CLI through npx codex-marketplace add wshobson/agents, OpenCode through make install-opencode, and also Cursor, GitHub Copilot, Antigravity CLI and Pi through generated files.
Is it free?
Yes. The repository is MIT licensed and no paid offering is documented; model usage is billed by whichever harness and provider you use.
Does it run its own orchestration process?
No. Orchestrator plugins are slash-command instructions that the host agent follows, calling specialist subagents in order and saving each step to files. The agent-teams plugin relies on Claude Code's Agent Teams feature.
What is the difference between the marketplace and the skills-only install?
A marketplace plugin brings its agents, commands and skills together. gh skill install or npx skills add installs individual skills only, without agents, commands or hooks.