Agent clients

Xerg has one canonical skill file and one hosted MCP endpoint. Client-specific differences are setup details, not separate skill files. An agent host runs setup. An audit source is the data Xerg analyzes. Codex is a setup host, not a native audit source. QM can be both host and source, but an operator must perform one-shot collection outside Slack and provide an authorized sanitized snapshot to the private scope. A Codex user audits whichever source xerg doctor finds: OpenClaw, Hermes, Claude Code, a Cursor export, or an ingest payload.

Install the skill (all clients)

For a terminal-capable agent, the universal cold-fetch prompt is:

set up https://xerg.ai/skill.md

The agent must ask before writing to a persistent or global skill directory. The deterministic alternative for Agent Skills clients is: For any agent that supports the Agent Skills standard (Claude Code, Codex, Cursor, and others), install the skill with one command:

npx skills add xergai/skills

Then ask your agent to audit your AI spend. The client-specific sections below cover manual skill paths and hosted MCP configuration.

QM in Slack

The same cold-fetch prompt can tell QM what operator prerequisites are missing, but it cannot grant itself PostgreSQL privileges, persist an administrator skill pack, or ask you to paste secrets:

set up https://xerg.ai/skill.md

QM support requires an operator to create a sanitized snapshot through strict direct or Fly-contained collection. An administrator-only private QM scope may receive an authorized snapshot and use an exact CLI that the operator provisioned with explicit approval, run doctor before auditing it, interpret results, and ask before activation or push. Current Fly Sprites resume persistent runtimes without applying the configured sandbox OCI image, so operators must verify the CLI inside that private Sprite instead of assuming an image pin installed it. The Sprite receives no database URL, identity key, Fly token, or Xerg credential by default. The skill explains pricing and detector coverage plus QM’s one-hour activity limit, refuses secrets and deployment-wide output, and does not claim it can initiate live database or Fly collection. Continuous follow capture, durable tool-history capture beyond QM’s retention window, and live Slack-triggered collection are not currently supported. QM uses the CLI rather than native Xerg hosted MCP to audit the authorized snapshot:

xerg doctor --runtime qm --qm-snapshot <snapshot.jsonl>
xerg audit --runtime qm --qm-snapshot <snapshot.jsonl> --since 7d

An explicit approved push may then make only the pseudonymous summary available to the dashboard, Slack, Ask Xerg, and hosted MCP. See secure QM setup and QM limitations.

Public agent surfaces

https://github.com/xergai/skills is the public skill repo (the npx skills add install source).
https://xerg.ai/skill.md is the canonical public SKILL.md mirror.
https://xerg.ai/.well-known/agent-skills/index.json is the v0.2.0 discovery index.
https://xerg.ai/.well-known/skills/index.json is the legacy Hermes discovery index.
https://xerg.ai/llms.txt is the short machine-readable site map.
https://xerg.ai/llms-full.txt is the longer agent reference.
@xerg/cli is the public npm CLI package.
@xerg/schemas is the MIT-licensed public wire-contract package.

Codex

Use SKILL.md for repeatable local audit workflows and hosted MCP for workspace data. Install with npx skills add xergai/skills, or copy or link the bundled skill directory into a Codex skill search path manually:

node_modules/@xerg/cli/skills/xerg

For user-wide Codex installs, use:

~/.agents/skills/xerg

For repo-scoped Codex installs, use:

.agents/skills/xerg

Codex hosted MCP config in ~/.codex/config.toml:

[mcp_servers.xerg]
enabled = true
url = "https://mcp.xerg.ai/mcp"

[mcp_servers.xerg.http_headers]
Authorization = "Bearer <XERG_API_KEY>"

Use a workspace API key from the hosted dashboard. Restart Codex after editing ~/.codex/config.toml.

Cursor

Run the guided setup:

npx @xerg/cli@latest mcp-setup

For project config, use Cursor’s remote MCP shape with the Xerg hosted endpoint and a Bearer token:

{
  "mcpServers": {
    "xerg": {
      "type": "http",
      "url": "https://mcp.xerg.ai/mcp",
      "headers": {
        "Authorization": "Bearer <XERG_API_KEY>"
      }
    }
  }
}

Claude Code and Claude Desktop

Use the standard remote MCP JSON shape:

{
  "mcpServers": {
    "xerg": {
      "type": "http",
      "url": "https://mcp.xerg.ai/mcp",
      "headers": {
        "Authorization": "Bearer <XERG_API_KEY>"
      }
    }
  }
}

Antigravity

Use the generic streamable HTTP MCP settings if your Antigravity install supports remote MCP servers with Bearer headers. Xerg does not publish Antigravity-specific setup steps yet because the product-specific import/config format still needs to be verified against official Antigravity documentation.

Generic MCP

Use this server endpoint:

https://mcp.xerg.ai/mcp

Use this auth header:

Authorization: Bearer <XERG_API_KEY>

Metadata endpoint:

https://mcp.xerg.ai/server.json

Local CLI for agents

Agents can run local audits without hosted auth:

npx @xerg/cli@latest audit --json
npx @xerg/cli@latest audit --compare
npx @xerg/cli@latest audit --fail-above-waste-rate 0.30

Hosted push and hosted MCP require an active workspace credential. To create a free workspace first, start at Get Started. Normal CLI setup uses npx @xerg/cli@latest activate; hosted MCP clients still use an API key from Settings.

Concepts

Hosted workspace

Automate and share

Releases

Agent clients

Install the skill (all clients)

QM in Slack

Public agent surfaces

Codex

Cursor

Claude Code and Claude Desktop

Antigravity

Generic MCP

Local CLI for agents

​Install the skill (all clients)

​QM in Slack

​Public agent surfaces

​Codex

​Cursor

​Claude Code and Claude Desktop

​Antigravity

​Generic MCP

​Local CLI for agents

Install the skill (all clients)

QM in Slack

Public agent surfaces

Codex

Cursor

Claude Code and Claude Desktop

Antigravity

Generic MCP

Local CLI for agents