aistack
June 7, 2026 · View on GitHub
Clean multi-agent orchestration for Claude Code. This plugin bundles the aistack MCP server together with ready-to-use slash commands and skills, so you can spawn specialized agents, share memory across them, gate risky actions behind human approval, and drive an autonomous backlog loop — all from inside Claude Code.
This plugin is a thin packaging wrapper around the
@blackms/aistack core. The MCP server is
launched on demand via npx @blackms/aistack mcp start; the plugin version is
kept in sync with the core package version.
- Repo: https://github.com/blackms/aistack
- License: MIT
- Author: blackms
Requirements
- Claude Code with the plugin system enabled.
- Node.js >= 20 on PATH (the MCP server runs via
npx). - First run downloads
@blackms/aistackvianpx(or install it globally:npm install -g @blackms/aistack).
Install (one command flow)
Add the marketplace, then install the plugin:
/plugin marketplace add blackms/aistack
/plugin install aistack
blackms/aistackresolves to this GitHub repository. The marketplace catalog lives at the repo root in.claude-plugin/marketplace.json. If that catalog has not been published yet, you can point Claude Code straight at the bundle for local testing:claude --plugin-dir ./plugin
When the plugin is enabled, the bundled aistack MCP server starts
automatically — no separate claude mcp add step is required. To register the
server standalone instead (outside the plugin), run:
claude mcp add aistack -- npx @blackms/aistack mcp start
What's in the bundle
MCP server
| Server | Start command | Tools |
|---|---|---|
aistack | npx @blackms/aistack mcp start (stdio) | agent_, identity_, memory_, task_, consensus_, session_, system_, github_ |
List the live tool set any time with npx @blackms/aistack mcp tools.
Slash commands (commands/)
| Command | What it does |
|---|---|
/aistack-init | Bootstrap aistack config + data dir in the current project |
/aistack-pm | Autonomous PM/PgM backlog loop (one issue per tick; pairs with /loop) |
/aistack-review | Adversarial review loop over the current diff (critic + steelman + verdict) |
Skills (skills/)
| Skill | What it does |
|---|---|
aistack-orchestrate | Decompose a task and route parts to specialized agents |
aistack-memory | Store / search / recall shared agent memory across sessions |
aistack-consensus | Human-in-the-loop approval gates before risky actions |
Examples
Bootstrap a project and confirm the tools are live:
/aistack-init
npx @blackms/aistack mcp tools
Orchestrate a multi-agent change — ask Claude:
"Use aistack to orchestrate this: have an architect propose the design, a coder implement it, then an adversarial agent review the diff."
The aistack-orchestrate skill maps each part to an agent type and spawns it
via the aistack MCP server (agent_spawn).
Run an autonomous backlog tick on an interval:
/loop 5m /aistack-pm
Review a branch before merging:
/aistack-review branch
Gate a deploy behind approval — ask Claude:
"Before deploying, open an aistack consensus checkpoint and wait for my approval."
The aistack-consensus skill calls consensus_check and pauses until
consensus_approve.
Versioning
The plugin version in .claude-plugin/plugin.json tracks the @blackms/aistack
core package version. Bump them together on each release.
Links
- aistack core + full docs: https://github.com/blackms/aistack
- Marketplace submission draft:
marketplace/cline-submission.md