Your agent runs on one model.xen puts the others on call.

A single binary for Claude Code or any MCP client. Frontier models from several vendors, one call away.

see it run

three models argue, one reconciles

backend — claudeexample session · council output captured 2026-09-24

should we put redis in front of postgres as a read cache? get the council on it

xen - council (MCP)(prompt: "Should we put Redis in front of Postgres as a read cache? Context: a typical web API on a single Postgres database…")

Convening three seats… (esc to interrupt)

degraded: false … +15 lines (ctrl+o to expand)

Three seats answered. The council leans against adding Redis now.

stance	model	verdict	conf.
for	`glm-5.2`	a sound architectural decision for this workload	7/10
against	`gemini-3.1-pro-preview`	an expensive and premature architectural leap	8/10
neutral	`grok-4.7`	only after measurement shows the database is the bottleneck	8/10

Synthesis gpt-6-astra: Do not commit to Redis yet. Approve a measured, narrowly scoped evaluation—not a default architectural rollout. …

review the retry helper in our api client before I merge it

xen - codereview (MCP)(step: "Review charter for src/retry.ts, a retry helper used by our TypeScript API client (Node 22, strict mode)…", relevant_files: ["/home/julien/projects/xen-website/.superpowers/hero-captures/retry.ts"], model: "gpt-6-astra", review_type: "full")

Reviewing with gpt-6-astra… (esc to interrupt)

code_review_complete: true … +39 lines (ctrl+o to expand)

gpt-6-astra raised six issues, and all six reproduce. It overruled one of my notes.

severity	where	finding
high	`retry.ts:38–39`	A retried POST can create six orders
high	`retry.ts:23–30`	A 401 is retried like a timeout
high	`retry.ts:20–30`	Abort mid-backoff keeps retrying
medium	`retry.ts:25–30`	It sleeps after the last attempt

I ran each one. My own note that the bare final error was a defect did not hold, so I dropped it.

our db gets bursts of five identical queries and I don't see why. think it through properly with a stronger model

xen - deepthink (MCP)(step: "Our API's database query rate is spiky and I don't understand why. A read-through cache sits in front of a slow query (200 ms)…", relevant_files: ["/home/julien/projects/xen-website/.superpowers/hero-captures/stampede.ts", …], model: "gpt-6-astra", thinking_mode: "high")

Thinking with gpt-6-astra… (esc to interrupt)

{"complete_deepthink":{"files_examined":[ … (ctrl+o to expand)

gpt-6-astra traced the bursts to the cache writing its entry only after the query returns.

claim	outcome	why
cache write	confirmed	written after the query returns
first burst	not ten: 15	timer ordering admits another batch
Map safety	ruled out	the race is async, not the Map

Its fix: share one in-flight query per key. I applied it and reran: 145 requests, 3 database calls instead of 35.

see it run

+13 tools

three models argue, one reconciles

The Operator: a robot wired to four AI model nodes

[ 01 // tools ]

Sixteen tools in one server.

Your agent calls xen's tools directly. Six skills, coming soon, will tell it which tool fits the job.

Skillscoming

instruction files your agent reads

xen:orchestrationxen:review-gatexen:challengexen:deepthinkxen:councilxen:brainstorm

When to call xen, which tool, which model. Once built, it will point your agent at all sixteen of xen's tools.

An outside review before a diff ships. Once built, it will point your agent at precommit, codereview and secaudit.

Re-examine a claim instead of caving. Once built, it will point your agent at challenge, chat, deepthink and council.

Investigate first, then one expert read. Once built, it will point your agent at deepthink and chat.

A one-way-door decision, argued blind. Once built, it will point your agent at council and consensus.

Ideas from several model families, narrowed. Once built, it will point your agent at chat and deepthink.

Pick a skill to preview the tools it will point to.

Your agentreads the skill, calls the tools

Included in xen

seven make no model call

Reason and challenge

councilModels debate blind, then one verdict
consensusModels weigh in one at a time, with stances
deepthinkOne strong model checks your investigation
challengeno model callYour agent re-examines a disputed claim

Review, debug and secure

codereviewQuality, security, performance, design
precommitChecks a diff before it is committed
debugRoot cause of one failure
secauditOWASP Top 10 and threat modelling

Understand and plan

plannerno model callYour agent builds a plan, step by step
analyzeArchitecture, performance, maintainability
traceno model callYour agent maps call chains
docgenno model callYour agent writes docstrings and API docs

Utilities

chatAsk one model; attach files and images
docslookupno model callYour agent looks up current API docs
listmodelsno model callYour providers, models and aliases
versionno model callServer version and configuration

[ 02 // install ]

Add xen to your agent.

One command installs xen and adds it to Claude Code.

Terminal

install xen and add it to Claude Code

curl -fsSL https://get.xenmcp.dev/install.sh | bash

One static binary, no runtime to install. Linux and WSL (on Windows, run it in WSL); needs Claude Code.

The installer sets up Claude Code only. The quickstart covers your key, updating and uninstalling.

Free to use, including at work, under the xen license.