mcpx

mcpx turns registered MCP servers into an agent-friendly command surface. Use it when a task needs MCP-backed services such as PostHog, Sentry, Cloudflare, or another server registered in the user's global mcpx registry.

Execution Rules

• Run mcpx directly.
• Start with the server inventory instead of dumping every registered schema.
• Call MCP tools as mcpx <server> <tool> --input <json>.
• Pass all tool arguments through --input; do not invent flags for MCP tool fields.
• Use focused schema selectors before calling tools on large MCP surfaces.
• Do not hand-edit ~/.agents/mcpx/servers.json or token cache files unless the user explicitly asks for registry surgery.
• Treat MCP server registration and OAuth setup as human-owned configuration. If a required server is missing or unauthenticated, stop and ask the user to configure it.

First Move: Discover The Surface

At the start of an MCP task, run bare mcpx to list configured servers:

mcpx

Choose the likely server from the user's request, then inspect only that server's schema:

mcpx --schema=.sentry
mcpx --schema=.posthog

After that, narrow to the specific tool or a short candidate list:

mcpx --schema=.posthog.projects-get
mcpx --schema='.posthog.{projects-get,alerts-list,alert-create}'

Avoid mcpx --schema unless the user explicitly asks for the complete command surface. On a real registry it may expand every server schema and waste the agent context.

Call Tools

Use inline JSON/JSON5 for tiny inputs:

mcpx posthog projects-get --input '{}'
mcpx sentry search-issues --input '{"query":"is:unresolved"}'

For non-trivial payloads, prefer @file or heredoc input:

mcpx <server> <tool> --input @payload.json

mcpx <server> <tool> --input @- <<'JSON'
{
  "query": "is:unresolved"
}
JSON

Default output is optimized for agents (YAML for structured content, direct text passthrough, media saved as file saved <path>). Use --raw to disable that optimization. When a tool emits notifications, --raw output may also include a JSON envelope — see Notifications below.

Server Transports

mcpx supports two MCP transports, both configured through mcpx @add:

• HTTP — default. Remote MCP services like PostHog or Sentry.
• Stdio — local processes started by mcpx, for example

mcpx @add --name fs --transport stdio --command bunx --args -y --args @modelcontextprotocol/server-filesystem --args /tmp/fs-sandbox.

HTTP bearer auth uses --bearer. Prefer env:NAME so secrets stay out of the registry. Repeat --bearer for multiple credentials; mcpx uses round-robin by default:

mcpx @add --name posthog --url https://mcp.posthog.com/mcp \
  --bearer env:POSTHOG_TOKEN_A \
  --bearer env:POSTHOG_TOKEN_B

After registration both behave identically from the agent's view: same schema discovery, same mcpx <server> <tool> --input ... call pattern, same output shape. You do not need to differentiate when calling tools.

Notifications

Most tools emit no notifications and this section never applies.

When an MCP server emits events during a call (progress reporting, server-side state changes, custom events), mcpx merges them into default structured output under an injected $notifications field:

count: 1
$notifications:
  - method: notifications/progress
    params:
      progressToken: '...'
      progress: 3
      total: 4
      message: step 3

For non-JSON text, binary, or mixed content, mcpx falls back to the trailing sentinel line:

<tool result lines>
$notification: [{"method":"notifications/progress","params":{...}}]

Notifications are objects with these fields:

• method — MCP notification method name
• params — method-specific payload (shape depends on method)
• aggregatedCount — only on the last notifications/progress entry per

progress token; indicates that N intermediate progress entries were collapsed to keep output bounded. The first and last entries are preserved verbatim.

Notification methods you may encounter:

In --raw mode with a structured result and non-empty notifications, the trailing sentinel line is replaced by a JSON envelope on stdout:

{ "result": <tool-result>, "notifications": [ ... ] }

Text results keep the trailing $notification: line even under --raw, because text content is not JSON and wrapping it would break consumers.

Practical guidance:

• If your task does not depend on progress or server events, ignore notifications.
• Check $notifications, the sentinel line, or the raw envelope only when present.
• Do not assume notifications appear; the common case is no $notification: at all.

Project-Local Skills

When the user has already configured the relevant global MCP servers and wants a project to expose only that approved set to agents, generate a project-local mcpx skill:

mcpx @skill --servers posthog,sentry

This writes:

.agents/skills/mcpx/SKILL.md

Use the generated project skill as the project-specific router. It should name servers the user has already made available; do not use this flow to register or authenticate new MCP servers.

For temporary agent guidance without writing .agents/skills, print a single-server skill to stdout:

mcpx @skill --show slack

Use this when the user asks for a one-off service action during the current agent session and does not want persistent project skills.

References

mcpx maintains a background daemon (mcpxd) for session reuse and notification buffering. You normally do not need to interact with it. If a tool call behaves unexpectedly (stuck session, surprising eviction, schema drift), see references/daemon.md before filing an issue.

Feedback

If mcpx behaves unexpectedly, capture the exact command, schema selector, input, and structured error, then file an issue:

gh issue create -R AIGC-Hackers/mcpx --title "bug: <summary>" --body "<repro steps and error output>"