Add Function Agent — Opt-in Business-Function Agent Scaffolding

Scaffold a single business-function agent backed by a verified-write MCP integration. Each function is opt-in. The plugin does NOT pretend that "scaffolding a marketing folder" gives you a marketing agent — it gives you nothing until you explicitly add one via this skill, with the MCP wired up and the scope written down.

What this skill creates per function

For each --function you add:

1. An agent definition at <project-or-company>/.claude/agents/<function>-agent.md — with tools restricted to the relevant MCP, model selection rationale, and a system prompt that describes scope.
2. An MCP server entry merged into .mcp.json — with environment variables for credentials (never hardcoded).
3. A SCOPE.md file at <project-or-company>/.claude/agents/<function>-scope.md — explicit "what this agent will / will NOT do" + recommended human-in-the-loop checkpoints + escalation rules.
4. A note in CLAUDE.md at the appropriate level — telling other agents this function-specific agent exists and when to defer to it.

Available functions (April 2026)

Only functions with a verified write-capable MCP are supported. If a function isn't in this list, it's because the integration doesn't exist or is read-only.

Verified write-capable

--function	MCP	Vendor status	Agent can	Agent will NOT
billing	Stripe MCP (official)	GA	Look up customers, create/void invoices, issue refunds (within limit), pull payment history, manage subscriptions	Issue refunds > configured limit; modify pricing; touch Connect / Treasury without explicit human approval
bookkeeping	QuickBooks MCP (official Intuit) OR Xero MCP (official)	GA	Reconcile transactions, categorize expenses, generate reports, post journal entries to draft state	Finalize close; modify chart of accounts; file taxes; alter prior-period entries
support-tier1	Intercom MCP (official)	GA	Look up users, search conversations, draft replies, tag/route tickets, escalate to human queue	Send replies autonomously (drafts only by default); refund or compensate without human; close tickets touching billing/legal
recruiting	Greenhouse MCP (community, mature) OR Lever MCP (community)	Community	First-pass screening notes, schedule interviews, sync candidate status, generate offer-letter drafts	Make hiring decisions; send rejection emails autonomously; modify job requisitions
compliance	Vanta MCP (official open-source) OR Drata MCP (official, hosted)	GA	Pull evidence, check control status, generate audit reports, flag missing evidence	Mark controls as compliant; modify policies; respond to auditors
crm-ops	HubSpot MCP (official) OR Salesforce MCP (official, hosted)	GA / Beta	Update contact/deal records, log activities, run reports, segment lists	Send autonomous outbound; modify pipeline definitions; bulk-delete records
contracts	DocuSign MCP (official, beta)	Beta	Send pre-approved templates, check signing status, query Navigator agreements, trigger Maestro workflows	Send custom contracts without human review; modify signed documents; legally bind without explicit human send
analytics-reporting	PostHog OR Mixpanel OR Amplitude OR GA4 MCP (all official)	GA	Run queries, build dashboards, export reports, drill into events	Modify tracking implementation; delete data; share PII outside approved destinations
marketing-email	Mailchimp MCP (community, mature) OR Customer.io MCP (official)	GA / Community	Build segments, draft campaigns, schedule sends to test segments, pull metrics	Send to full list without human approval; modify suppression list autonomously; touch GDPR / unsubscribe records
web-cms	Webflow MCP (official)	GA	Read/write CMS items, manage Sites API, publish (with confirmation)	Publish without human approval; modify Designer canvas without MCP Bridge App
social-drafts	(no MCP for Buffer/Hootsuite — drafts only, written to filesystem)	N/A	Draft posts to a social-drafts/ folder for human review and manual posting	Post anything autonomously (no integration exists in 2026 for write-capable scheduling)

Explicitly NOT supported (and why)

These functions don't get an agent because the integration isn't there:

Function	Status	Why no agent
payroll	No MCP for Gusto/Rippling write; Deel is read-only	Cannot reliably execute payroll without write APIs
banking-ops	No MCP for Mercury/Brex/Ramp	Cannot move money out of operating accounts
support-tier2-zendesk	Zendesk is MCP client, not server	Cannot let Claude act in Zendesk
support-helpscout	No MCP	No write integration
clm	No Ironclad MCP (only DocuSign signing)	Contract lifecycle management is human-only
social-scheduling	No Buffer/Hootsuite MCP	Drafts only — see social-drafts
apollo-outbound	No first-party Apollo MCP	Use crm-ops if you sync Apollo data into HubSpot/Salesforce first

Workflow

Step 1: Determine target

Find the right .claude/agents/ directory. If --project <path> is given, use that. Otherwise:

1. Check current directory for .claude/agents/
2. Walk up to find a company root (look for CLAUDE.md + .claude/)
3. Ask the user

Step 2: Validate function choice

Confirm --function is in the supported list above. If not, surface the "not supported" table above and explain why — never silently fall back to a generic template.

Step 3: Gather function-specific context

Ask the user for:

• Credentials source — which env var name will hold the token? (e.g., STRIPE_SECRET_KEY, INTERCOM_ACCESS_TOKEN)
• Scope limits — for billing: what's the autonomous refund limit? for marketing-email: which segment is the agent allowed to send to without human approval?
• Escalation channel — Slack channel, email, or escalations/ folder where the agent dumps anything outside its scope

In --dry-run mode, skip prompts and emit the proposed file contents to stdout instead of writing.

Step 4: Generate the agent definition

Load references/function-agent-templates.md and pick the matching template. Customize:

• Model: Opus for read-heavy / judgment-heavy functions (recruiting, compliance), Sonnet for action-heavy / drafting (billing, bookkeeping, support-tier1)
• Tools: Read, Grep, Glob, Bash plus the specific MCP tool prefix (e.g., mcp__stripe__*)
• Scope limits from Step 3 inserted into the system prompt

Write to .claude/agents/<function>-agent.md.

Step 5: Write the SCOPE.md companion

Critical step. Write .claude/agents/<function>-scope.md containing:

• Will do (concrete, scoped list)
• Will NOT do (concrete, ban list — taken verbatim from the function's table row above + any user-specified limits)
• Human-in-the-loop checkpoints (when the agent must surface something for human approval)
• Escalation channel (where things outside scope go)
• Failure modes (what to do when the MCP errors, when credentials expire, when rate-limited)

The scope file is the honesty contract. The agent's system prompt references it.

Step 6: Merge MCP entry into .mcp.json

Read existing .mcp.json (or create fresh). Merge in the new server entry from references/function-agent-templates.md. Use ${ENV_VAR} references — never hardcode credentials.

If .mcp.json already has a server with the same name, ask before overwriting.

Step 7: Update CLAUDE.md to register the agent

Add a one-line entry in the appropriate CLAUDE.md (engineering/<project>/CLAUDE.md or company root, depending on scope) under an "Agents" section:

## Function-Specific Agents
- `billing-agent` — Stripe operations within scope. See `.claude/agents/billing-scope.md` for limits.

This tells other agents and humans that the specialist exists.

Step 8: Validate and report

Run the validators:

python3 ${plugin_root}/scripts/check-mcp-config.py <project>/.mcp.json
python3 ${plugin_root}/scripts/audit-structure.py <project> --skip-projects --skip-hooks

Report:

• File paths created
• Env vars the user must set before invoking the agent
• Recommended next step: test the MCP connection via /mcp in Claude Code

Step 9: Honest summary

Tell the user, verbatim:

Added <function> agent at <path>. Before this is useful:

1. Set the env var(s) listed above
2. Restart Claude Code so the MCP server picks up the new config
3. Run /mcp to verify the server is connected
4. Read <scope-file> — what this agent will and will NOT do

The agent does not run autonomously. You invoke it like any other Claude Code subagent. Recommended human-in-the-loop checkpoints are listed in the scope file.

What this skill does NOT do

• Does not configure cron / scheduled invocation of the agent (use Routines via claude.ai or the user's scheduler).
• Does not manage credential rotation — that's still your operational responsibility.
• Does not check whether the MCP is currently up — claude mcp list and /mcp are the verification tools.
• Does not verify the agent's outputs are correct — there is no automated quality gate for "did the bookkeeping reconcile correctly." That's the human-in-the-loop checkpoint's job.
• Does not give you a function that doesn't have a write-capable MCP, no matter how much you ask. See the "Explicitly NOT supported" table.

Key References

• references/function-agent-templates.md — Full agent + MCP + scope templates for each supported function
• Plugin-level scripts/check-mcp-config.py — Run after MCP changes
• Plugin-level scripts/audit-structure.py — Run after agent file changes