microsoft/add-workiq
Adds Work IQ (M365 Copilot Search) to a Power Apps code app via the Work IQ Copilot MCP connector (shared_a365copilotchatmcp), then wires up a production-ready McpSession wrapper for AI-powered, knowledge-grounded search and chat. Use when integrating Microsoft 365 Copilot search/chat. The CopilotChat tool searches internal Microsoft 365 content (documents, emails, chats, sites, files) across your organization — prefer workload-specific tools (SharePoint, OneDrive, Teams, Mail) when the workload is explicit; do not use it for general knowledge, news, public web, or external information.
npx skills add https://github.com/microsoft/power-platform-skills --skill add-workiq
📋 Shared Instructions: shared-instructions.md - Cross-cutting concerns.
Work IQ is accessed through the dedicated Work IQ Copilot MCP connector (shared_a365copilotchatmcp; shown as "Work IQ Copilot MCP (Preview)" in the maker portal). It exposes an MCP (Model Context Protocol) endpoint whose CopilotChat tool performs AI-powered, knowledge-grounded search and conversation over Microsoft 365 content.
> This is the purpose-built Work IQ connector and generates a WorkIQCopilotMCPService with a single mcp_m365copilot operation. Every command and code reference in this skill is specific to shared_a365copilotchatmcp (connection commands, and the generated WorkIQCopilotMCPService / WorkIQCopilotMCPModel files). Do not run this skill for a different connector. If you instead need the broader shared_a365mcpservers "Microsoft 365 MCP Servers" bundle (Mail/Teams/SharePoint MCP servers), add it via /add-connector — its generated service is MicrosoftMCPServersService, so you must adjust the Step 2 commands and the wrapper imports accordingly.
> ⚠️ Work IQ uses MCP. A JSON-RPC initialize handshake runs before the first tools/call. This connector's server (Microsoft.MCPPlatform.WebApi) is stateless-tolerant — do NOT send an Mcp-Session-Id on initialize (the server treats it as a session lookup and returns -32001 Session not found). Drive the connector through the McpSession wrapper below, which runs the handshake, sends no session id by default, sequences JSON-RPC ids, auto-retries on Session not found, and parses the nested response for you.
Check for memory-bank.md per shared-instructions.md.
The Power Apps code-app CLI (@microsoft/power-apps-cli, invoked as pa or power-apps — resolve via cli-binary.md) creates the connection and generates the typed service itself. Make sure the CLI is installed (npm install) and you are signed in (pa auth status, or power-apps auth-status on power-apps-only projects; it shares the same auth as the rest of the code-app skills).
Check for an existing connection first:
pa connection list
Look for a Work IQ Copilot MCP (Preview) connection (api id shared_a365copilotchatmcp) in the output. If one is listed, note its Connection ID and skip to "Add the Data Source" below.
Otherwise create one with the native create-connection verb:
pa connection create --connector shared_a365copilotchatmcp
power.config.json — you do not need to pass an environment id.STOP HERE — interactive sign-in required:
If create-connection fails:
pa auth status (or power-apps auth-status on power-apps-only projects), sign in if needed, and retry.As a fallback, the user can create the connection manually in the maker portal: https://make.powerapps.com/environments/<environment-id>/connections → + New connection → search for "Work IQ Copilot MCP" → Create, then re-run list-connections.
Once the connection exists, add it to the code app (this is what generates the typed service + model):
pa app add data-source --connector shared_a365copilotchatmcp -c <connection-id>
This is a non-tabular connector — only --connector (api id) and -c (connection id) are needed.
After adding the connector, confirm the generated service is present. This is a small, single-operation service, so you can read it directly or grep it:
Grep pattern="async \w+" path="src/generated/services/WorkIQCopilotMCPService.ts"
The WorkIQCopilotMCPService exposes exactly one operation — mcp_m365copilot ("Work IQ Copilot (Preview)"), plus a Getmcp_m365copilot GET variant used only for connection verification. Work IQ / CopilotChat is driven through mcp_m365copilot.
Its generated signature is:
public static async mcp_m365copilot(
Mcp_Session_Id?: string,
queryRequest?: QueryRequest
): Promise<IOperationResult<void>>
Mcp-Session-Id parameter). Leave it undefined — this server assigns/needs no client session id, and sending one on initialize returns -32001 (see below).QueryRequest ({ jsonrpc?, id?, method?, params?, result?, error? }) — exported from src/generated/models/WorkIQCopilotMCPModel.ts.IOperationResult<void>; the actual JSON-RPC / SSE body arrives in result.data at runtime.Facts you can rely on (do not try to discover them at runtime):
CopilotChat (case-sensitive). Do not substitute query, search, or find.message — not query, prompt, or question.Mcp-Session-Id on the initialize handshake. This connector's server treats an incoming id as an existing-session lookup and returns -32001 Session not found. The server is stateless-tolerant: initialize and tools/call both succeed with no session id, so the McpSession wrapper below tracks none by default.> Session-id handling (verified during testing). This connector's MCP server is stateless-tolerant. initialize with no Mcp-Session-Id returns 200 with server capabilities, and a subsequent tools/call with no id returns the Copilot reply — no session id needs to be tracked or echoed. Do not generate a client id and send it on initialize: the server treats it as a lookup and returns 404 / -32001 Session not found (this was a real bug in an earlier version of this wrapper). Note that the code-apps data layer's IOperationResult<TResponse> exposes only { success, data, error, skipToken, count, fileName } — it does not surface response headers — so if a future MCP server returns a session id only in the Mcp-Session-Id response header, it would be unreadable here. The wrapper still defensively adopts a server id if one ever appears inside result.data.
⚠️ CRITICAL: MCP session handling and response parsing are intricate. Copy the production-ready McpSession class below exactly. It runs the initialize handshake (sending no session id), sequences JSON-RPC ids, auto-retries on "Session not found", persists the conversation id, and parses the deeply nested response.
Create src/connectors/mcpClient.ts:
// src/connectors/mcpClient.ts
import type { IOperationResult } from '@microsoft/power-apps/data'
import { WorkIQCopilotMCPService } from '../generated/services/WorkIQCopilotMCPService'
import type { QueryRequest } from '../generated/models/WorkIQCopilotMCPModel'
export interface JsonRpcRequest {
jsonrpc: '2.0'
id?: string
method: string
params?: Record<string, unknown>
}
export interface JsonRpcResponse {
jsonrpc?: string
id?: string
result?: Record<string, unknown>
error?: { code?: number; message?: string; data?: unknown }
}
type CopilotConversationMessage = {
text?: string
attributions?: Array<{
attributionType?: string
providerDisplayName?: string
seeMoreWebUrl?: string
}>
}
type CopilotConversation = {
messages?: CopilotConversationMessage[]
}
function parseRpc(result: IOperationResult<unknown>): JsonRpcResponse {
if (!result.success && result.error) {
return { error: { message: result.error.message } }
}
const data: unknown = result.data
if (data == null) return {}
if (typeof data === 'object') return data as JsonRpcResponse
if (typeof data === 'string') {
// Handle SSE framing: "event: message\ndata: {JSON}"
const dataLines = data
.split(/\r?\n/)
.filter((line) => line.startsWith('data:'))
.map((line) => line.slice(5).trim())
// Per the SSE spec, multiple `data:` lines are joined with newlines — a single
// JSON object can be split across lines, so joining with '' would corrupt it.
const payload = dataLines.length ? dataLines.join('\n') : data
try {
return JSON.parse(payload) as JsonRpcResponse
} catch {
return { result: { raw: data } }
}
}
return { result: { raw: data } }
}
export class McpSession {
private nextId = 1
// MCP Streamable HTTP: the client must NOT send a session id on `initialize` —
// the server assigns one. Sending a client-generated id makes this connector's
// server return `404 / -32001 Session not found`. This server is stateless-
// tolerant, so we send no id at all; `extractSessionId` still adopts a server
// id if one ever surfaces in the response body.
private sessionId: string | undefined = undefined
private conversationId: string | undefined
private initialized = false
private extractSessionId(raw: IOperationResult<unknown>): string | undefined {
const container = raw as unknown as Record<string, unknown>
const dataObj =
raw.data && typeof raw.data === 'object'
? (raw.data as Record<string, unknown>)
: undefined
const resultObj =
dataObj?.result && typeof dataObj.result === 'object'
? (dataObj.result as Record<string, unknown>)
: undefined
const candidates: Array<unknown> = [
dataObj?.['Mcp-Session-Id'],
dataObj?.mcpSessionId,
dataObj?.sessionId,
resultObj?.['Mcp-Session-Id'],
resultObj?.mcpSessionId,
resultObj?.sessionId,
container['Mcp-Session-Id'],
container.mcpSessionId,
container.sessionId,
]
const found = candidates.find((value) => typeof value === 'string' && value.length > 0)
return typeof found === 'string' ? found : undefined
}
private isSessionNotFound(res: JsonRpcResponse): boolean {
const message = (res.error?.message ?? '').toLowerCase()
return message.includes('session not found') || res.error?.code === -32001
}
private resetSession(): void {
// Drop any session id and re-handshake from scratch (no id on `initialize`).
this.sessionId = undefined
this.initialized = false
}
private async send(
method: string,
params?: Record<string, unknown>,
allowRetry = true
): Promise<JsonRpcResponse> {
const req: JsonRpcRequest = { jsonrpc: '2.0', id: String(this.nextId++), method, params }
// Send no session id on `initialize` (server treats any id as existing-session lookup).
const sessionId = method === 'initialize' ? undefined : this.sessionId
const raw = (await WorkIQCopilotMCPService.mcp_m365copilot(
sessionId,
req as unknown as QueryRequest
)) as unknown as IOperationResult<unknown>
const negotiatedSessionId = this.extractSessionId(raw)
if (negotiatedSessionId) {
this.sessionId = negotiatedSessionId
}
const parsed = parseRpc(raw)
if (allowRetry && method !== 'initialize' && this.isSessionNotFound(parsed)) {
this.resetSession()
await this.initialize()
return this.send(method, params, false)
}
return parsed
}
async initialize(): Promise<JsonRpcResponse> {
const res = await this.send('initialize', {
protocolVersion: '2025-06-18',
capabilities: {},
clientInfo: { name: 'Power Apps Code App', version: '1.0.0' },
})
// Fail fast: surface a failed handshake instead of continuing to `tools/call`.
if (res.error) {
throw new Error(`MCP initialize failed: ${res.error.message ?? JSON.stringify(res.error)}`)
}
this.initialized = true
return res
}
/** Discover the available MCP tools (optional; the tool name is fixed as `CopilotChat`). */
async listTools(): Promise<JsonRpcResponse> {
if (!this.initialized) await this.initialize()
return this.send('tools/list', {})
}
async callCopilotChat(message: string): Promise<{ text: string; conversationId?: string }> {
if (!this.initialized) await this.initialize()
const raw = await this.send('tools/call', {
name: 'CopilotChat',
arguments: {
message,
...(this.conversationId ? { conversationId: this.conversationId } : {}),
},
})
// Surface MCP/connector errors as thrown exceptions so callers can `try/catch`
// instead of string-matching the returned text.
if (raw.error) {
throw new Error(raw.error.message ?? JSON.stringify(raw.error))
}
const parsed = extractCopilotText(raw)
if (parsed.conversationId) {
this.conversationId = parsed.conversationId
}
return { text: parsed.text, conversationId: parsed.conversationId }
}
}
function extractContentText(res: JsonRpcResponse): string | undefined {
const content = res.result?.content as Array<{ type?: string; text?: string }> | undefined
if (!Array.isArray(content)) {
return undefined
}
const textBlocks = content
.filter((c) => c.type === 'text' && typeof c.text === 'string')
.map((c) => c.text!.trim())
.filter((value) => value.length > 0)
if (textBlocks.length === 0) {
return undefined
}
// Prefer the JSON payload block. Some responses append metadata text blocks
// such as "CorrelationId: ..." that should not be concatenated.
const jsonBlock = textBlocks.find(
(block) => block.startsWith('{') && /"conversationId"|"rawResponse"|"reply"/.test(block)
)
if (jsonBlock) {
return jsonBlock
}
const nonMetadata = textBlocks.find((block) => !/^CorrelationId\s*:/i.test(block))
return nonMetadata ?? textBlocks[0]
}
export function extractCopilotText(res: JsonRpcResponse): {
text: string
conversationId?: string
} {
const rawText = extractContentText(res)
if (!rawText) {
return { text: res.result ? JSON.stringify(res.result, null, 2) : '(no content returned)' }
}
try {
const inner = JSON.parse(rawText) as {
conversationId?: string
reply?: string
message?: string
rawResponse?: string
}
if (typeof inner.rawResponse === 'string') {
try {
const convo = JSON.parse(inner.rawResponse) as CopilotConversation
const messages = Array.isArray(convo.messages) ? convo.messages : []
// messages[0] echoes the user query; the AI reply is the one carrying
// attributions (citations). Fall back to index 1, then the last message.
const attributed = messages.find(
(m) => Array.isArray(m.attributions) && m.attributions.length > 0
)
const selected = attributed ?? messages[1] ?? messages[messages.length - 1]
const replyText = selected?.text?.trim()
if (replyText) {
return { text: replyText, conversationId: inner.conversationId }
}
} catch {
// Fall through to simple reply extraction.
}
}
const fallbackText = inner.reply?.trim() || inner.message?.trim() || rawText
return { text: fallbackText, conversationId: inner.conversationId }
} catch {
return { text: rawText }
}
}
> Session-id note (verified during testing). McpSession starts with no session id and sends none on the initialize handshake. Testing against shared_a365copilotchatmcp (server Microsoft.MCPPlatform.WebApi) confirmed it is stateless-tolerant: initialize with no id returns 200 + capabilities, and tools/call with no id returns the Copilot reply. Sending a client-generated id on initialize instead makes the server return 404 / -32001 Session not found — that was the original bug. IOperationResult exposes only { success, data, error, skipToken, count, fileName } (no response headers), so a header-only session id would be unreadable here; extractSessionId still adopts a server id if one ever appears inside result.data, and the handshake auto-retries on Session not found for resilience.
>
> If you ever bind a *different* MCP connector that is genuinely stateful and returns its session id only in the Mcp-Session-Id response header, this SDK cannot read it — capture one raw mcp_m365copilot result in the debugger to confirm where the id surfaces, and if it appears inside result.data, add that path to extractSessionId.
Initialize once per app (e.g., in a React useEffect on boot or a module singleton) and reuse for all Work IQ calls:
import { McpSession } from './connectors/mcpClient'
// Initialize once per app session and reuse across calls.
const workIqSession = new McpSession()
export async function queryWorkIQ(userPrompt: string): Promise<string> {
try {
const { text } = await workIqSession.callCopilotChat(userPrompt)
return text
} catch (error) {
const msg = error instanceof Error ? error.message : 'Work IQ query failed'
console.error('Work IQ Error:', msg)
throw error
}
}
Key patterns:
new McpSession() per request)callCopilotChat()conversationId is automatically persisted across calls for multi-turn chatsWork IQ responds well to context-rich, structured prompts. Adapt this pattern for your scenario:
const prompt = `
You are [role/expert description].
**Context:**
- [Relevant data or background information]
- [Additional context as needed]
**Task:** [Clear, specific instruction]
**Format:**
- Use markdown with clear section headings (## Summary, ## Action Items, etc.)
- Specify limits (word count, number of items, etc.)
`.trim()
const { text } = await workIqSession.callCopilotChat(prompt)
Adaptable scenarios: meeting summaries with action items, prioritized daily action items from email, project risk analysis from documents, team performance insights, or any knowledge-grounded analysis over M365 data.
callCopilotChat() already unwraps the nested JSON-RPC / SSE response and returns clean text. Parse that text according to the format you requested.
For structured markdown output (when you asked for ## sections):
function extractSection(text: string, sectionName: string): string[] {
const regex = new RegExp(`##\\s*${sectionName}\\s*([\\s\\S]*?)(?=##|$)`)
const match = text.match(regex)
if (!match) return []
return match[1]
.split('\n')
.filter((line) => line.trim().startsWith('-'))
.map((line) => line.replace(/^-\s*/, '').trim())
.filter(Boolean)
}
const { text } = await workIqSession.callCopilotChat(prompt)
const summary = extractSection(text, 'Summary')
const actionItems = extractSection(text, 'Action Items')
For unstructured output — use text directly (display as-is or format for your UI).
Work IQ uses MCP (Model Context Protocol). Driving it correctly requires:
initialize triggers -32001 Session not found); it defensively adopts a server id only if one ever surfaces in the response body (see the session-id note above)-32001 / "Session not found")McpSession handles all of this. Bypassing it (ad-hoc per-call ids or direct calls without the handshake) leads to "Session not found" errors and failed integrations.
npm run build
Fix TypeScript errors before proceeding. Do NOT deploy yet.
Update memory-bank.md with: connector added (shared_a365copilotchatmcp), McpSession wrapper created at src/connectors/mcpClient.ts, Work IQ usage wired up, build status.
IMPORTANT — Do NOT save sensitive information:
Only save high-level progress like "Work IQ Copilot MCP connector configured" or "Work IQ CopilotChat integration implemented via McpSession".
Take microsoft/add-workiq from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference npm.
Without those the skill loads but fails at the first command.