# Superpowers: features, protocols, quickstart

## TL;DR

Superpowers is a set of skills plus a session-start hook that gives a coding agent a fixed workflow: brainstorm a design, write a plan, execute it with subagents and reviews, then finish the branch. Jesse Vincent and Prime Radiant maintain it. It installs as a plugin into many harnesses and suits developers who want a defined process.

## Key facts

| Field | Value |
| --- | --- |
| Type | Orchestrator |
| Languages / SDKs | Markdown, Shell, JavaScript, TypeScript |
| License | MIT |
| Pricing model | Open source, free |
| Orchestration pattern | Supervisor |
| GitHub stars | 293,256 (as of 2026-09-30) |
| GitHub forks | 26,242 |
| Last push | 2026-09-27 |
| Latest release | v6.4.2 |
| Repository | [obra/superpowers](https://github.com/obra/superpowers) |
| Documentation | [github.com](https://github.com/obra/superpowers/blob/main/README.md) |
| Last verified | 2026-09-30 |

## Key features

- A library of 15 skills covering brainstorming, planning, test-driven development, systematic debugging, code review, worktrees, branch finishing, and writing new skills. ([source](https://github.com/obra/superpowers/blob/main/README.md))
- A session-start hook injects the using-superpowers bootstrap, so skills are checked before each task without the user invoking them. ([source](https://github.com/obra/superpowers/blob/main/hooks/hooks.json))
- subagent-driven-development dispatches a fresh implementer subagent per plan task, runs a spec-and-quality review after each, and ends with a whole-branch review. ([source](https://github.com/obra/superpowers/blob/main/skills/subagent-driven-development/SKILL.md))
- dispatching-parallel-agents tells the host to assign one agent per independent problem area, such as separate failing test files, and run them concurrently. ([source](https://github.com/obra/superpowers/blob/main/skills/dispatching-parallel-agents/SKILL.md))
- using-git-worktrees creates an isolated workspace before plan execution, preferring the harness's native worktree tool and falling back to plain git. ([source](https://github.com/obra/superpowers/blob/main/skills/using-git-worktrees/SKILL.md))
- Model-selection guidance for subagent roles: cheap models for mechanical tasks, the most capable model for design and the final review, and an explicit model on every dispatch. ([source](https://github.com/obra/superpowers/blob/main/skills/subagent-driven-development/SKILL.md))
- Per-harness install paths for Claude Code, Codex, OpenCode, Hermes Agent, Cursor, Gemini CLI, Copilot CLI, Pi and others. ([source](https://github.com/obra/superpowers/blob/main/README.md))
- A diagnosing-superpowers skill reads a session transcript, reports what went wrong with line-level evidence, and can package a scrubbed bundle for a bug report. ([source](https://github.com/obra/superpowers/blob/main/README.md))

## Architecture and orchestration pattern

Pattern: Supervisor.

Superpowers is content, not a runtime. The repository holds Markdown skills, per-harness plugin manifests (`.claude-plugin`, `.codex-plugin`, `.hermes-plugin`, `.opencode/plugins`, `.pi/extensions` and others) and a session-start hook. Nothing runs as a separate process; the host coding agent loads the skills and follows them.

Multi-agent work is delegated to the host. The `subagent-driven-development` skill makes the main session act as a controller: it reads an approved plan, dispatches one implementer subagent per task, dispatches a reviewer after each task, and finishes with a broad branch review. The subagents are launched by the host's own subagent tool. The README notes that Pi has no built-in subagent tool, so subagent and task-list tools stay optional companion packages there. `executing-plans` is the single-context alternative with one final review.

State lives in files: design documents, the plan, and a ledger kept in a plan workspace that the controller updates as tasks finish. There is no long-term memory store; the bootstrap is re-injected at session start, and again after compaction on harnesses that offer such a hook.

### Human in the loop

Approval gates sit before implementation. The brainstorming skill carries a hard gate: the human approves the design (a short in-chat design for small work, a written spec for architectural work) before a plan is written or code is touched. Once execution starts, the subagent-driven skill is written to run without pausing between tasks and to record its own rulings in a ledger. It stops for irreversible or destructive actions, security-sensitive actions, side effects outside the worktree such as merges or pushes, and plans so broken every path is a guess. finishing-a-development-branch presents merge, PR, keep or discard options for the user to choose.

### Harnesses it can drive

- Claude Code ([evidence](https://github.com/obra/superpowers/blob/main/README.md))
- Codex ([evidence](https://github.com/obra/superpowers/blob/main/README.md))
- OpenCode ([evidence](https://github.com/obra/superpowers/blob/main/README.md))
- Hermes ([evidence](https://github.com/obra/superpowers/blob/main/README.md))

## Protocols

| Protocol | Support | Evidence | Note |
| --- | --- | --- | --- |
| MCP | Unknown (checked 2026-09-30) | — | Searched the README, the repository code and the skills for MCP. The only hits are a reference to the separate mcp-builder skill and a generic note in the porting guide; Superpowers ships no MCP server or client. |
| A2A | Unknown (checked 2026-09-30) | — | A2A / Agent2Agent search of the README and repository code found no mention (the one code-search hit is a substring inside an image URL). |
| AG-UI | Unknown (checked 2026-09-30) | — | No AG-UI mention in the README or repository code search; not checked against a listing in the AG-UI README. |

## Best for

- Developers who want one repeatable design, plan, implement and review process across whichever coding agents they use. ([shortlist](https://multiagentguide.top/best/coding-agents.md))
- Plan-driven feature work where a controller session hands each task to a fresh subagent and reviews the result. ([shortlist](https://multiagentguide.top/best/coding-agents.md))
- Teams that want human sign-off on a design before an agent writes code, then a mostly unattended execution phase.
- Debugging and test-first discipline enforced through skills rather than through prompts typed each session. ([shortlist](https://multiagentguide.top/best/coding-agents.md))

## Not for

- Building agents in code; there is no SDK or runtime, only skills and hooks for existing coding agents.
- Users who want to start coding immediately with no design step; the workflow is written to block that.
- Cross-harness setups that need one install; the README says to install separately for each harness.

## Quickstart

```sh
/plugin install superpowers@claude-plugins-official
```

Install not yet verified by this site.

```bash
# Claude Code (inside the app)
/plugin install superpowers@claude-plugins-official

# Codex CLI: open the plugin search, search "superpowers", choose Install Plugin
/plugins

# Hermes Agent
hermes plugins install obra/superpowers --enable

# Pi
pi install git:github.com/obra/superpowers

# OpenCode: send this instruction to the agent
# Fetch and follow instructions from https://raw.githubusercontent.com/obra/superpowers/refs/heads/main/.opencode/INSTALL.md
```
Start a new session and ask for something to build; the brainstorming skill should run before any code is written.

### Common pitfalls

- Install once per harness; installing in one does not enable it in another.
- Hermes has no post-compaction hook, so a long session that compacts past its first turn can lose the bootstrap; start a fresh session if skills stop triggering.
- Restart running sessions (Hermes, Muse) after installing so the session-start hook takes effect.
- The brainstorming visual companion loads a logo from the vendor's site that reports the Superpowers version; set `SUPERPOWERS_DISABLE_TELEMETRY` to disable it.
- Pi has native skills but no subagent tool in the package; subagent-driven work needs a separate Pi companion package.

Official quickstart: https://github.com/obra/superpowers/blob/main/README.md#installation

## Pros

- The brainstorming skill blocks implementation until the human approves the design, and separates approval of the conversation from approval of the written spec. ([source](https://github.com/obra/superpowers/blob/main/skills/brainstorming/SKILL.md))
- Documents install paths for more than a dozen harnesses, including the official Claude and Codex plugin marketplaces. ([source](https://github.com/obra/superpowers/blob/main/README.md))
- Subagent roles get written model-selection rules, with a mid-tier floor for reviewers and a stronger model for the final review. ([source](https://github.com/obra/superpowers/blob/main/skills/subagent-driven-development/SKILL.md))
- Zero-dependency by design: the contributor guide rejects PRs that add third-party dependencies unless they add a new harness. ([source](https://github.com/obra/superpowers/blob/main/AGENTS.md))
- MIT license and no paid tier in the repository; frequent tagged releases with release notes. ([source](https://github.com/obra/superpowers/releases))

## Cons

- An open issue reports a simple plan run with subagent-driven development consuming a full five-hour Codex token budget. ([source](https://github.com/obra/superpowers/issues/1152))
- An open bug report says the orchestrator sometimes writes out subagent dialogues itself instead of making real subagent calls. ([source](https://github.com/obra/superpowers/issues/1749))
- On OpenCode, the bootstrap is reported to be injected into task subagents, which then restart the controller workflow (open issue). ([source](https://github.com/obra/superpowers/issues/2160))
- Each harness needs its own install, and on Hermes the bootstrap can be lost after compaction because there is no post-compaction hook. ([source](https://github.com/obra/superpowers/blob/main/README.md))
- The README says new skills are generally not accepted as contributions, and any skill change must work across all supported agents. ([source](https://github.com/obra/superpowers/blob/main/README.md))

## Alternatives

- [BMad Method](https://multiagentguide.top/tools/bmad-method.md)
- [gstack](https://multiagentguide.top/tools/gstack.md) ([Superpowers vs gstack](https://multiagentguide.top/compare/superpowers-vs-gstack.md))
- [ECC](https://multiagentguide.top/tools/ecc.md)
- [wshobson/agents](https://multiagentguide.top/tools/wshobson-agents.md)
- [agency-agents](https://multiagentguide.top/tools/agency-agents.md)

## FAQ

### Does Superpowers support MCP?

No MCP support was found. It ships skills, hooks and plugin manifests, not an MCP server or client, so the field is recorded as unknown. MCP use depends on the host harness.

### Does Superpowers run its own agents?

No. The skills instruct the host coding agent to dispatch subagents with the host's own tools. Superpowers defines the controller, implementer and reviewer roles and the order they run in.

### Which harnesses does it install into?

The README documents Claude Code, Codex, OpenCode and Hermes Agent among others, including Cursor, Gemini CLI, Copilot CLI, Pi, Kimi Code and Qwen Code. Install separately for each.

### Is Superpowers free?

The repository is MIT licensed. The README also offers commercial support and additional tooling through a sales contact but describes no hosted product, so it is recorded as oss-free.

### How is it different from a subagent pack such as wshobson/agents?

A subagent pack supplies many specialist role definitions. Superpowers supplies one workflow of skills, and the controller and reviewer roles are part of that workflow.

## Sources

- [Superpowers README](https://github.com/obra/superpowers/blob/main/README.md)
- [Superpowers hooks.json (SessionStart hook)](https://github.com/obra/superpowers/blob/main/hooks/hooks.json)
- [subagent-driven-development skill](https://github.com/obra/superpowers/blob/main/skills/subagent-driven-development/SKILL.md)
- [dispatching-parallel-agents skill](https://github.com/obra/superpowers/blob/main/skills/dispatching-parallel-agents/SKILL.md)
- [using-git-worktrees skill](https://github.com/obra/superpowers/blob/main/skills/using-git-worktrees/SKILL.md)
- [brainstorming skill](https://github.com/obra/superpowers/blob/main/skills/brainstorming/SKILL.md)
- [Superpowers contributor guidelines (AGENTS.md)](https://github.com/obra/superpowers/blob/main/AGENTS.md)
- [Superpowers releases](https://github.com/obra/superpowers/releases)
- [Issue 1152: Codex token budget with subagent-driven development](https://github.com/obra/superpowers/issues/1152)
- [Issue 1749: orchestrator fabricates subagent dialogues](https://github.com/obra/superpowers/issues/1749)
- [Issue 2160: OpenCode bootstrap injected into task subagents](https://github.com/obra/superpowers/issues/2160)
- [Superpowers README, installation section](https://github.com/obra/superpowers/blob/main/README.md#installation)
- [obra/superpowers GitHub repository](https://github.com/obra/superpowers)

## Unknown fields

protocols.mcp: README, skills and repo code search show no MCP server or client (only a reference to a separate mcp-builder skill). protocols.a2a: no mention (only a spurious substring hit in an image URL). protocols.agui: no mention found. Commercial services (sales contact for support, tooling, managed spending) are described in one README paragraph without a product page, so pricing_model is oss-free with that caveat.

Corrections or removal requests: support@multiagentguide.top

---

Data as of 2026-09-30. Not affiliated with listed projects. HTML version: https://multiagentguide.top/tools/superpowers