╔══════════════════════════════════════════════════════════════╗
║                                                              ║
║                T  R  I  A  G  E  M  C  P                     ║
║                                                              ║
║    expose Triage's priority queue as MCP tools for agents    ║
║                                                              ║
╚══════════════════════════════════════════════════════════════╝

An MCP server that turns Triage into a set of tools an AI agent can call directly. Add tasks, recompute priorities, explain why a task ranks where it does — all from inside a Claude Code (or any MCP-compatible) session.

Mirrored on both GitHub and Codeberg. Issues on either forge are welcome; commits land on both.

---

What is Triage?

Triage is a meta-scheduler: a self-aware priority queue. Signals push facts about the world; rules turn facts into priority deltas; the queue reorders itself on every tick. The human sets the goals — Triage decides the order, and tells you exactly why.

It already has a clean Python API + a CLI (triage list, triage why, triage tick). TriageMCP is the bridge that makes that API addressable from inside an agent's tool-use loop.

If you're new to Triage, read its README first — TriageMCP is a thin wrapper, not a re-implementation.

---

Tools exposed

Every tool maps to a method already available on triage directly, returning structured data (not stringified CLI output) so the agent can reason about it:

MCP tool	Read/Write	Mirrors CLI	Returns
list_tasks	read	triage list	priority-ordered task list with full rule contributions
get_task	read	triage show	one task's full record
why_task	read	triage why	rule-by-rule contribution deltas explaining a task's priority
status	read	triage status	top-3 + tag counts + active-signal counts
add_task	write	triage add	the new task + its assigned id
remove_task	write	triage rm	confirms removal
tick	write	triage tick	recomputes priorities; returns new top-3 + signal counts
inject_signal	write	triage signal manual	writes a manual signal that rule_manual_bump picks up

The agent works on the same ~/.triage store the Triage CLI does — so anything you add via triage add from the terminal is visible to the agent on the next call, and vice versa.

---

Install

# 1. Install Triage (TriageMCP imports from it directly — no subprocess shell-out)
pip install git+https://github.com/CryptoJones/Triage.git
# or, for an editable clone:
git clone https://github.com/CryptoJones/Triage.git && pip install -e ./Triage

# 2. Install TriageMCP
git clone https://github.com/CryptoJones/TriageMCP.git
cd TriageMCP
pip install -e .

Once a Triage release is on PyPI, this collapses to a single pip install triagemcp.

Wire into Claude Code

Edit (or create) ~/.claude/mcp.json:

{
  "mcpServers": {
    "triagemcp": {
      "command": "triagemcp"
    }
  }
}

Restart Claude Code. The eight tools above will be available to the agent. Ask it to "what should I do next?" and it can call list_tasks + why_task to give you a real answer.

Override the store location

By default TriageMCP uses Triage's own resolution: $TRIAGE_HOME → ~/.triage. To point an agent at a different store (e.g. one task queue per agent), set TRIAGEMCP_HOME:

{
  "mcpServers": {
    "triagemcp": {
      "command": "triagemcp",
      "env": { "TRIAGEMCP_HOME": "/home/you/.triage-agent-A" }
    }
  }
}

---

Why direct import, not subprocess

Some MCP servers (including KaliMCP, its security-tooling sibling) shell out to wrapped binaries. TriageMCP does not:

• Speed — no fork/exec overhead per tool call.
• Fidelity — results come back as typed structures with full rule contribution arrays, not text that the agent (or this server) has to parse.
• Determinism — both this server and the triage CLI use the same scheduler module, so what the agent sees is exactly what the CLI would compute, no more and no less.

The cost: TriageMCP requires Triage as a Python dependency. That's why the install step has two commands. Once a PyPI release of Triage exists this collapses to one.

---

Sample agent transcript

USER: I just merged the linter fix PR. What's next?

AGENT: [calls tick()] OK, recomputed priorities.
       [calls list_tasks(limit=5)] Your top 5 are:
         1. Fix the auth bug             [120]
         2. Rotate the staging cert       [85]
         3. Reply to legal about contract [50]
         4. Open release PR               [12]
         5. Update README badges           [3]

       [calls why_task(top.id)] Auth bug is at 120 because:
         + base_score: +10
         + manual_bump: +100 (from runpodboss signal — pod is idle)
         + deadline_decay: +10 (deadline in 3 days)

Every priority is explainable end-to-end via the rule-contribution arrays the tools return. No black box.

---

Status

Version	Feature	Status
v0.1	eight tools (list / get / why / status / add / remove / tick / inject_signal); direct in-process import of triage; matching tests for each	shipped
v0.2	streaming watch via triage watch mode once Triage ships that (v0.10 milestone)	planned
v0.3	per-task update_task partial-edit tool (description / base_score / tags only — id stays stable)	planned
v0.4	Dockerfile parity with KaliMCP so the server can run in a sealed image	planned

---

Related projects

• Triage — the meta-scheduler this server wraps.
• claude_skill-Triage — Claude Code skill repository: TaskPriorityReorder (manual override) + forthcoming triage skill (signal-driven recommender).
• KaliMCP — sibling MCP server; same authorial conventions, different tool surface (Kali security tools).
• RunPodBoss — RunPod credit guardrail with a documented integration recipe into Triage via extra_notify_command.

---

License

Apache 2.0. See LICENSE.

Proudly Made in Nebraska. Go Big Red! 🌽 https://xkcd.com/2347/