A Slack Agent With MCP and Skills#

Most Slack bots are annoying because they are not workflows.

They are command handlers with a personality.

An agentic Slack workflow should feel different. It should understand the request, gather context, use the right tools, ask for approval when needed, and leave a clear handoff.

That takes more than a bot token.

The job#

Pick one job first.

Not "help the whole company."

Something like:

• Engineering triage.
• Support escalation.
• Customer risk summary.
• Incident update drafting.
• Billing exception routing.
• Release note drafting.

The narrower the job, the more useful the first version will be.

The agent shape#

I would start with four pieces.

Intake

Reads the Slack thread, classifies the request, and decides whether the workflow should run.

Context

Pulls the sources the human would open anyway: docs, tickets, GitHub, dashboards, customer records, runbooks.

Draft

Creates the next useful artifact: reply, summary, owner route, bug report, incident update, or checklist.

Approval

Blocks risky actions and shows the human exactly what will happen.

That is enough.

Do not build a dramatic agent swarm when the first useful product is better triage.

Where MCP fits#

MCP is a clean way to expose tools and data sources to the agent.

Useful MCP-backed surfaces might include:

• Docs search.
• Issue search.
• GitHub context.
• Workflow automation.
• Customer data lookup.
• Observability reads.
• Runbook retrieval.

The point is not to expose everything.

The point is to expose the right tool surface for the workflow.

Where Skills fit#

Skills are the reusable playbooks.

Examples:

• Incident update style.
• Billing escalation policy.
• Support tone rules.
• PR review checklist.
• Security review steps.
• Customer-facing message rules.

Without Skills, the agent's behavior lives in one giant prompt.

With Skills, the workflow can load the right instructions for the job.

That makes the agent easier to review and easier to improve.

What the Slack reply should do#

A good agent reply is compact.

It should include:

• What it found.
• Sources.
• Confidence.
• Suggested owner.
• Recommended next step.
• Approval request if needed.

It should not dump a novel into the thread.

Slack is where work moves. Keep the reply shaped for work.

Approval examples#

Require approval before:

• Posting customer-facing text.
• Creating a ticket.
• Changing status.
• Issuing a refund.
• Triggering a deploy.
• Running any tool with side effects.

Drafts are cheap. Writes are expensive.

Make that difference visible.

Where Codelit fits#

Codelit can design the Slack agent before you wire Slack, MCP, and credentials together.

You can map:

• Slack trigger.
• Agent roles.
• MCP tools.
• Skills.
• Approval gates.
• Model routes.
• Evals.
• Runbook.
• Production handoff.

That turns "we should add a Slack AI agent" into an actual implementation plan.

Build it in Codelit#

Try this:

Build a Slack agent workflow for engineering triage using MCP tools and Skills. Include thread intake, context gathering, GitHub and docs lookup, draft response, owner routing, human approval, eval cases, and repo handoff.

Build the Slack agent workflow

A Slack agent is not valuable because it talks in Slack. It is valuable because it reduces the time from vague request to clean next step.