Slack GitHub MCP Bot

A Slack bot that answers natural language questions about your GitHub repository using AWS Bedrock (Claude) and GitHub's official MCP server.

Ask it anything in Slack:

• @bot what bugs are open?
• @bot which PRs have been waiting for review the longest?
• @bot were any critical issues opened after last week's merges?

The bot also posts a structured morning summary to a Slack channel every day at 9 AM.

---

How It Works

Slack (@mention)
      ↓
AWS Bedrock — Claude (Agent Loop)
      ↓
MCP Servers (GitHub + any other)
      ↓
Tool results → reasoning → answer → Slack

Claude receives all available GitHub tools at startup. When a question arrives, it decides which tools to call, calls them, reads the results, and loops until it has enough to answer. Your code never routes a question — Claude handles that.

---

Prerequisites

• Node.js 18+ (required for global fetch)
• AWS account with Bedrock access and Claude Sonnet 4.6 enabled
• GitHub account with a repository to query
• Slack workspace where you can install apps

---

Setup

1. Create the Slack App

1. Go to api.slack.com/apps → Create New App → From Scratch
2. Settings → Socket Mode → Enable — create an App-Level Token with connections:write scope (this is your SLACK_APP_TOKEN)
3. OAuth & Permissions → Bot Token Scopes — add:

   app_mentions:read   chat:write   chat:write.public   im:history   im:write
   

4. Event Subscriptions → Enable → Subscribe to Bot Events — add app_mention and message.im
5. OAuth & Permissions → Install to Workspace — copy the Bot User OAuth Token
6. Copy the Signing Secret from Basic Information

2. Create a GitHub Fine-Grained Token

1. github.com → Settings → Developer Settings → Personal access tokens → Fine-grained tokens
2. Select your target repository
3. Repository permissions: Issues (Read), Pull requests (Read), Contents (Read), Metadata (Read)

3. Create a Bedrock API Key

1. AWS Console → Amazon Bedrock → Model access → Enable Claude Sonnet 4.6
2. Bedrock → API keys → Create API key

Important: Claude 4.x requires the us. cross-region inference profile prefix.
Use us.anthropic.claude-sonnet-4-6 — not anthropic.claude-sonnet-4-6.
The bare ID returns: "on-demand throughput isn't supported"

4. Configure and Run

cp .env.example .env
# Fill in all values in .env

npm install
node index.js
# Bot ready — 44 tools loaded

5. Invite to a Channel

/invite @your-bot-name

---

Morning Summary

If SUMMARY_CHANNEL_ID is set in .env, the bot posts a daily summary at 9 AM covering:

• Active release PRs grouped by environment (DEV/STG/PROD)
• Yesterday's activity (PRs merged/opened, issues closed/opened)
• New bugs filed in the last 24h, classified by severity
• Unassigned critical/major bugs flagged for attention

Test it immediately:

npm run summary

---

Adding More MCP Servers

The tool registry pattern means adding a new data source is 4 lines in index.js:

// In the startup block, after loadServer(GITHUB_MCP_URL, githubHeaders):
const myHeaders = {
  Authorization:  `Bearer ${process.env.MY_SERVICE_TOKEN}`,
  'Content-Type': 'application/json',
  Accept:         'application/json, text/event-stream',
};
await loadServer(process.env.MY_SERVICE_MCP_URL, myHeaders);
// Then add routing guidance to systemPrompt() describing when to use the new tools.

Available MCP servers:

Service	URL	Adds
Linear	mcp.linear.app/mcp	Issues, projects, cycles
Cloudflare	mcp.cloudflare.com/mcp	Workers, analytics, DNS, R2
Stripe	mcp.stripe.com/mcp	Payments, customers, subscriptions
Custom	@modelcontextprotocol/sdk	Wrap any REST API as MCP tools

---

Tuning the System Prompt

The system prompt in systemPrompt() is the highest-leverage variable. Key rules to keep:

• Explicit mrkdwn syntax — *bold* not **bold**, <url|text> not [text](url)
• GitHub link format — <https://github.com/ORG/REPO/issues/N|#N> makes issue numbers clickable
• High-volume rule — when >15 results, summarise by category and show top 10
• Tool routing — when using multiple MCP servers, tell Claude which questions map to which server
• Emoji vocabulary — define a consistent set so all responses look coherent

---

Observability

Every tool call is logged to stdout as structured JSON:

{
  "event": "tool_call",
  "tool": "search_issues",
  "input": { "q": "repo:org/repo is:open label:bug" },
  "result": { ... },
  "ts": "2026-04-10T09:00:00.000Z"
}

Ship these logs to Datadog, CloudWatch, or Grafana. Tool call frequency, latency, and error rate are the key metrics for monitoring agent health.

---

Models

Model	BEDROCK_MODEL_ID	Notes
Sonnet 4.6	us.anthropic.claude-sonnet-4-6	Recommended — fast and capable
Opus 4.6	us.anthropic.claude-opus-4-6-v1	Best reasoning, slower
Haiku 4.5	us.anthropic.claude-haiku-4-5-20251001	Fastest, most cost-efficient

---

Known Constraints

• Latency — complex queries take 10–30 seconds. The thinking-indicator pattern mitigates the UX impact.
• Debugging — non-deterministic reasoning is harder to trace than deterministic code. Use the tool call logs.
• Cost — 8 tool calls per query costs more than a single API response. Monitor at scale.
• Prompt dependency — poor prompt design causes wrong tool selection and bad formatting. Test iteratively.

---

Project Structure

index.js        — Slack handlers, agent loop, MCP client, tool registry, scheduler
package.json
.env.example
.gitignore
README.md

---

License

MIT