Skip to main content
ShoulderTap is the missing layer between AI agents and the humans who hold undocumented knowledge. When an agent hits a term it can’t resolve, a judgment call, or context that no document answers, ShoulderTap routes the question to the person who knows, asks them on Slack in a well-formed message, captures their reply as structured knowledge with full provenance, runs it through human approval, and writes it back to a system of record. The agent’s blocker is resolved, and your organization permanently learns — no agent asks that question again.
ShoulderTap is a self-hosted, open-source service you run yourself. It is distributed as the shouldertap Python package with the shtap command-line tool. There is no hosted product to sign up for — you run the engine, own the data, and point your agents at it.

​
The loop

A consumer (your agent) submits a ContextRequest. The engine deduplicates it, routes it to the right expert from a registry, and drafts a Slack DM. The expert replies in plain language. An LLM structures that reply into a ContextProposal, which lands in an approval queue. Once a human accepts it, the proposal is delivered back to the consumer to write into a system of record, and the expert is told their answer was accepted and attributed to them. If no reply arrives within escalation_after (default 2 hours), ShoulderTap re-asks the expert’s configured escalation target. If nothing arrives within give_up_after (default 24 hours), the request fails with reason timeout.

​
Who it’s for

ShoulderTap is for teams building AI agents that keep hitting the same wall: questions no document answers.
  • Agent and RAG builders who need a human-in-the-loop escape hatch for their agent’s low-confidence path, instead of guessing or giving up.
  • Data and platform teams who own a system of record — a data catalog, glossary, or knowledge base — and want expert answers captured into it with provenance, not lost in Slack threads.
  • Anyone running MCP-capable agents who wants to tap a human expert with no custom integration, via the built-in MCP server.

​
The safety promise, to your experts

Every outbound ask — with no code path that skips this — contains, in order: a self-identification (“I’m ShoulderTap, an automated assistant”; it never pretends to be a person), who’s asking and why, the question (drafted to be answerable from memory), effort framing (“a one-or-two sentence reply is all that’s needed”), an attribution promise (the answer is recorded with the expert’s name after human review — never written back automatically unless a consumer explicitly opts into auto_accept), and an opt-out (reply mute to stop receiving asks, or skip to pass this one along). Rate limits and quiet hours are enforced before every send, not just documented.

​
Get started

Quickstart

Run the whole loop in your terminal in about 90 seconds — no Slack required.

Concepts

The vocabulary: requests, proposals, experts, consumers, kinds, and topics.

Integrate an agent

Wire your agent’s low-confidence path to a human expert via MCP, the SDK, or webhooks.

Protocol reference

The stable HTTP API, contracts, and kind schemas for any client.