mcpField
J
junai Pipeline MCP

junai Pipeline MCP

Local
github.com/saajunaidAI & Agents0Updated 1d ago

About junai Pipeline MCP

caddis is a Claude Code harness of commands, skills, subagents, hooks, and shared project conventions for development workflows. Its portable resource pool also supports other AI CLIs, while session handoffs and memory help carry context forward.

Key features

  1. 01Claude Code commands and skills
  2. 02Subagents for review and testing
  3. 03Session relay and memory
  4. 04Hooks for safety and linting
  5. 05Portable bundles for other CLIs

Use cases

  1. 01Plan and implement features in phases
  2. 02Review changes with another model vendor
  3. 03Resume work across coding sessions

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"saajunaid-junai": {
"command": "uvx",
"args": [
"junai-mcp"
],
"env": {
"PIPELINE_STATE_PATH": "<PIPELINE_STATE_PATH>",
"JUNAI_WORKSPACE_ROOT": "<JUNAI_WORKSPACE_ROOT>"
}
}
}
}
Run commanduvx junai-mcp
Environment variables to set
PIPELINE_STATE_PATHPath to pipeline-state.json. Defaults to .github/pipeline-state.json relative to the workspace root (detected by searching upward from the current directory for a .github folder).
JUNAI_WORKSPACE_ROOTExplicit workspace root path override. Use if automatic detection (searching upward for .github) does not find the correct root.

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
pypijunai-mcp0.1.1

Tools

8

Saved tool list · updated 3w ago

ParameterTypeRequiredDescription
artefact_pathanyNo
result_statusstringYes
result_payloadanyNo
stage_completedstringYes

Details

From the project README

caddis — a Claude Code harness

License: MIT

Browse every command, skill, agent and hook -> saajunaid.github.io/caddis-plugin — the reference site, rebuilt on every publish.

caddis is a Claude Code harness — a plugin (plus a shared pool of skills, subagents, slash-commands, and hooks) that makes a single Claude Code session dramatically more capable and much harder to derail. It's agent-agnostic: the same pool is exported to other AI CLIs (Codex, Antigravity, Copilot), and CLAUDE.md↔AGENTS.md are mirrors so every agent reads the same conventions.

Before you start

You needWhy
Claude Code, installed and signed incaddis is a plugin for it
Python 3.11+ on PATHthe hooks and the cross-review tool are Python
Gitseveral commands read repository state
(optional) an API key for DeepSeek or GLMonly for cross-vendor review and the OSS lanes

Honest about the platform: caddis is developed and used daily on Windows with PowerShell. The hooks are pure Python and cross-platform, and the skills, commands and agents are plain markdown that work anywhere — but some launchers ship as both .ps1 and .sh, the deployment skill is Windows-specific, and the publishing machinery (sync.ps1) is PowerShell-only. Nothing here is hostile to macOS or Linux; it is simply less travelled, so expect rough edges and please report them.

Your first five minutes

claude plugin marketplace add saajunaid/caddis-plugin
claude plugin install caddis@caddis

Then, in a Claude Code session, confirm it actually loaded:

/caddis:version

If that prints a version, you are installed. From there, the shortest path to seeing what caddis is for:

Try thisWhat it does
/caddis:feature-plan add CSV exportwrites a phased, TDD-structured plan to .caddis/plans/ — the durable spine the rest of the harness reads
/caddis:implement <plan>executes that plan phase by phase, committing each, and stops only at a real boundary
/caddis:cross-reviewhas a different vendor's model review your diff, because a same-vendor reviewer shares your blind spots
/caddis:handoffwrites the resume doc, so the next session starts with zero re-discovery

Everything else is optional. If you only ever use feature-plan → implement → cross-review, you have most of the value.

This repository is the published marketplace mirror: it hosts the caddis and caddis-extras Claude Code plugins, the portable per-harness bundles under bundles/, and the shared pool. It is generated and pushed by the caddis authoring repo's exporter — issues are welcome here, but code changes land upstream and are synced in.


Install (Claude Code)

# 1. Register this repo as a plugin marketplace (once)
claude plugin marketplace add saajunaid/caddis-plugin

# 2. Install the core plugin (always-on, lean context)
claude plugin install caddis@caddis

# 3. (optional) Add the long-tail skill library — cloud/data/media/etc. Disabled by default,
#    so it costs zero always-on context until you enable it.
claude plugin install caddis-extras@caddis

Then, in any repo:

/caddis:setup-project-ai     # deploy the harness into the project

A normal working loop looks like:

/feature-plan      # (or /prd first) → writes .caddis/plans/<slug>.md, the durable spine
/implement         # executes the plan phase-by-phase, TDD, commit per phase
/ship              # express lane: commit → push → CI (hotfixes); feature work: /ship-pr then /ship-merge
/handoff           # ALWAYS end a session with this — writes the resume doc

The one habit that matters most: end every session with /handoff. It writes .caddis/relay.md, which is re-injected automatically at the next session start — you resume with zero re-discovery instead of a cold, forgetful session.

See USERGUIDE.md for the full command/skill/hook reference and the mental models.


What you get

  • Slash-commands — /feature-plan, /prd, /implement, /tdd, /ship · /ship-pr · /ship-merge, /handoff, /kb, /digress · /resume, /cross-review, /usage-review, /mermaid-db · /excalidraw-db, /setup-project-ai, /migrate-dir.
  • Subagents — lean, own-context helpers: anchor (evidence-first verification), code-reviewer, preflight (plan-vs-codebase validation), tester, debug, codebase-audit, security-analyst, data-engineer, sql-expert, knowledge-transfer.
  • A four-layer memory model — session relay (relay.md), automatic Dream Memory (memory.jsonl), a per-repo knowledge base (.caddis/kb/ + DOC-MAP.md), and cross-repo memory.
  • Hooks & safety — a PreToolUse guard for secret/destructive writes, auto-lint on edit, and the SessionStart/Stop hooks behind relay + memory.
  • Tiered skills — the core dev set ships in caddis; the long tail (cloud, data, media, productivity) lives in caddis-extras, off by default.

Installing outside Claude Code (Codex, Antigravity, …)

The knowledge layer (skills + AGENTS.md conventions) works in any harness. Per-harness bundles are published under bundles/ and installed with caddis-init:

# from a checkout of this repo (the installer ships in plugin/scripts/)
python plugin/scripts/claudster_init.py --target codex        --dest /path/to/project
python plugin/scripts/claudster_init.py --target antigravity  --dest /path/to/project

Safe by design: a sha256 manifest tracks what the installer wrote; re-runs update only unmodified files, and anything you edited locally is reported as a conflict, never overwritten.


Repo layout

PathWhat it is
.claude-plugin/marketplace.jsonThe marketplace manifest (caddis + caddis-extras)
plugin/The core caddis Claude Code plugin (agents, commands, hooks, core skills)
plugin-extras/The caddis-extras skill library (disabled by default)
bundles/Portable per-harness exports (codex, antigravity) for caddis-init
.github/The shared pool: skills, agents, prompts, instructions, tools
sync.ps1, export_runtime_resources.py, validate_pool.pyThe build/sync machinery

A note on names

caddis was previously published as claudster (and, before that, this repo hosted the junai Copilot pipeline — see the pre-1.2 entries in CHANGELOG.md). Repos set up under the old name have a .claudster/ artifact dir; everything reads both .caddis/ and .claudster/ and writes where the repo already lives, so nothing breaks. Run /caddis:migrate-dir when you want a repo converted. The ~/.claudster/ user scope (keys, install records) is permanent.


License

MIT. The published plugin bundles (caddis / caddis-extras) carry the same license in their manifests.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.