Run SocialHub from Claude Code
SoCode is the SocialHub integration for Claude Code (and Claude Desktop / Cursor): a downloadable skill plus a governed MCP server. Operate SocialHub — and ask about your data — in plain language.
› ask in plain language
you › GMV last week by store, top 3▋
↳ governed metric · gmv · by store · last_7d
from the semantic layer · no SQL · audited · illustrative
What is SoCode?
SoCode is how SocialHub plugs into your AI assistant. It's two pieces working together — so Claude can operate SocialHub and answer questions about your data, in plain language, without you leaving your editor.
flash-operations
A downloadable Claude skill that teaches Claude how to drive SocialHub safely — which tool to use, confirming writes by member name/email, and only quoting governed metric numbers.
Governed MCP server
A Model Context Protocol endpoint that exposes SocialHub's tools — loyalty, members, coupons, campaigns and your governed metrics — authenticated by a tenant API key and authorized per tool.
Why SoCode?
Your data and operations, reachable from the AI tools your team already uses — without trading away governance.
Natural language, no dashboards
Ask in plain English; Claude picks the right governed tool and runs it. No clicking through screens, no SQL to write.
Governed & safe by design
Metric numbers come from SocialHub's semantic layer (the AI never invents figures). Writes require a write scope and pass a server-side guard. Every key is scoped to your team; every call is audited.
Works where you already work
Claude Code, Claude Desktop, Cursor — any MCP client. Connect once with a single command and operate SocialHub without leaving your assistant.
Start read-only
Begin with the metrics:read preset (safest for AI). Add write scopes only when you need to award points or issue coupons — you stay in control.
What wasn't possible before SoCode
Each of these used to need an analyst, an engineer, or a dashboard you don't have — so the person with the question rarely got the answer in time. A governed agent inside your own AI toolis what makes them possible: business-grade answers and actions, self-serve, without giving up control of your data. That's the unlock — not a faster dashboard, a different operating model.
“Find the members slipping toward churn, launch a win-back, and keep tuning it to my goal.”
BeforeThree teams and three tools across weeks: an analyst pulls the data, a marketer builds the campaign, someone checks results next month — by then the moment has passed and the loop never actually closes.
With SoCodeOne agentic loop runs analyze → execute → improve: it reads the governed metrics, launches the journey you approve, measures the lift, and re-tunes — continuously, against the goals and guardrails you set.
“GMV last week by store — and why are redemptions down this month?”
BeforeFile a request and wait for an analyst, or stitch it together across dashboards — the answer arrives after the decision is already made.
With SoCodeA governed, trustworthy answer in seconds — the call gets made on data, not on a hunch.
Expose SocialHub's governed metrics to your own assistant or a Cursor workflow over MCP.
BeforeA custom integration, hand-written queries, and governance re-implemented from scratch — months of work, and a new place for numbers to drift.
With SoCodeOne scoped key and the governed layer is callable from your stack. The integration that used to be a project is a config line.
“Did active members drop vs last month? Break it down by tier.”
BeforeYou'd notice it in the monthly review — after the month was already lost.
With SoCodeSelf-serve, multi-step diagnosis the moment you wonder — catch the leak while you can still act on it.
How SoCode works
Three steps: download the skill, install the MCP server, and use it in natural language.
1. Download SoCode (the skill)
The SoCode skill (flash-operations) teaches Claude how to drive SocialHub safely. It pairs with the SocialHub MCP server you'll add in step 2.
Get the skill folder
# Pull just the SoCode skill into Claude Code
npx degit huangchunbo2025/flash/skills/flash-operations \
~/.claude/skills/flash-operationsOr browse / download it on GitHub → skills/flash-operations. Copy the folder into ~/.claude/skills/ (personal) or your project's .claude/skills/.
The skill is optional but recommended — the MCP server works without it; SoCode's skill just makes Claude use it well.
2. Install the MCP server
a) Create an API key
In SocialHub, go to Settings → API Keys → Create Key and pick the "Read-only" preset (includes metrics:read — safest for AI use; add write scopes only when you need to award points or issue coupons). The key is shown once — copy it.
b) Add the server to Claude Code
Export the key as an environment variable first, so the plaintext key never lands in your shell history or a config file:
export FLASH_API_KEY="fl_live_your_key_here"
claude mcp add --transport http flash \
https://flash.socialhub.ai/api/mcp \
--header "Authorization: Bearer ${FLASH_API_KEY}"Prefer config? Add a project-scoped .mcp.json at your repo root — reference the env var, never paste a plaintext key:
{
"mcpServers": {
"flash": {
"type": "http",
"url": "https://flash.socialhub.ai/api/mcp",
"headers": {
"Authorization": "Bearer ${FLASH_API_KEY}"
}
}
}
}c) Verify
- Run
claude mcp list—flashshould show as connected. - Ask Claude to "list available SocialHub tools"— you'll only see the tools your key's scopes allow.
- The server name must be exactly
flash, andFLASH_API_KEYmust be exported in the shell that launches Claude Code. An invalid key shows as a failed connection.
3. Use it
Just ask Claude in plain language. For example:
Query your data
- “List the metrics I can query.”
- “GMV last week by store.”
- “Why did redemptions drop this month?”
Run operations
- “Award 200 points to Alice (alice@…).”
- “List active coupon pools.”
- “Issue an ambassador code for this member.”
Stay in control
- Writes are confirmed before running.
- Server-side caps block oversized writes.
- See every call under API Keys → Recent MCP Activity.
Safe by design.Metric numbers come from SocialHub's governed semantic layer (the AI never writes SQL or invents figures). Write operations require a write scope and pass a server-side guard (per-call value cap + rate limit) — chat confirmation is courtesy, the server is the boundary. Every key is scoped to your team only.