Open source · self-hosted

When your AI agent gets stuck,
tap a human on the shoulder.

ShoulderTap routes an agent's unanswerable questions to the right expert on Slack, captures the reply as structured knowledge with provenance, and writes it back — so no agent is ever stuck on that question again.

View on GitHub Read the docs

The problem

Docs can't answer everything.

Agents hit judgment calls, undefined terms, and context that lives only in someone's head. Today they fail, guess, or hallucinate. ShoulderTap gives them a fourth option: ask a person — once — and remember the answer forever.

How it works

  1. A consumer asks. An agent, RAG pipeline, or MCP-capable tool submits a question it can't resolve.
  2. ShoulderTap routes it. The engine deduplicates, picks the right expert, and drafts a clear Slack message.
  3. An expert replies. A human answers in Slack — messy, human, in their own words. They can mute or skip anytime.
  4. The reply becomes knowledge. An LLM structures the answer with full provenance and drops it in an approval queue.
  5. You approve, it's written back. The consumer gets its answer and your org permanently learns.

Who it's for

Agent & RAG builders

Add a reliable human-in-the-loop fallback for the questions your retrieval layer can't cover.

MCP-capable agents

Wire in expert escalation over MCP with no custom integration to build or maintain.

Data & knowledge teams

Turn scattered tribal knowledge into a system of record with attribution and provenance.

Anyone shipping to prod

Safety guardrails, rate limits, quiet hours, and escalation timeouts keep experts unbothered.

Get started

Running in three commands.

Self-hosted and free. Test the whole loop from your terminal — no Slack credentials required.

# install
$ pip install shouldertap

# configure your experts & transports
$ shtap init

# run the loop locally — no Slack needed
$ shtap serve --transport console