Skip to content

Getting started

How Merlin works

Follow a request from a team surface to a local agent and back.

Merlin connects requests in Slack, Linear and GitHub to Claude Code or Codex running on a connected machine, then returns progress and the final result to the source conversation.

Follow one request

Suppose Mia asks @Merlin in a Slack channel to investigate a checkout regression and prepare a fix.

  1. The Slack thread becomes the conversation for that request.
  2. Merlin checks who made the request, the conversation's audience and which runtimes that person may use.
  3. The selected runtime runs Claude Code or Codex with the files, credentials and tools available on its connected machine.
  4. Progress and the final result return to the same Slack thread.
  5. An eligible teammate can reply in that thread with a missing detail, a correction or the next requested revision.
Conversation, same thread
Device and runtime
Workspace
Conversation
Source conversation
Slack thread, Linear issue, GitHub thread or an automation
Workspace admission
Who asked, the conversation's audience, the runtimes they may use
Runtime selection
The runtime already doing the work, a preferred one, or another eligible runtime
Local execution
Claude Code or Codex with the files, credentials and tools on that machine
Progress and result
Structured progress and the final answer return to the source thread
Teammate follow-up
A missing detail, a correction or the next requested revision
Follow-up re-enters

The same shape applies when work starts in Linear, GitHub or an automation. The source and available context change, but work still moves through an eligible runtime and returns to its originating conversation.

The three-legged model

There are three parts to a working setup:

PartWhat it does
WorkspaceBrings together the team's members, integrations, machines, skills and settings.
Device and runtimeConnects a machine and exposes an available local agent such as Claude Code or Codex.
ConversationKeeps a request, progress, source links and follow-ups together on the surface where work began.

The desktop app and CLI daemon connect the machine. The runtime uses the files, credentials and tools available in that local environment. Merlin does not create a separate hosted machine for every member.

From request to result

  1. A registered or otherwise admitted requester mentions or delegates to Merlin in Slack, Linear or GitHub, or an automation's schedule or event fires.
  2. Merlin records the surface and audience, checks workspace admission and looks for an eligible runtime.
  3. The selected runtime receives the request, bounded source context, saved guidance and any memory that surface may read.
  4. Claude Code or Codex performs the work with the capabilities available on that machine.
  5. Merlin sends structured progress and the final answer back to the source thread or session.
  6. A follow-up in that conversation can steer active work or continue it later.

How a task finds a machine

Merlin looks for a runtime the requester may use and that has capacity. Existing conversations prefer continuity with the runtime already doing the work. For new work, an explicit or preferred runtime influences selection. Otherwise, Merlin can choose an eligible runtime owned by the requester or shared with the workspace.

Private runtimes accept only their owner's requests. Public runtimes can accept eligible requests from other members. An automation runs only on the runtime its creator chose for it, one they own; when that runtime is offline or busy, the run is skipped.

A connected device is not sufficient by itself. Its runtime must be enabled, signed in and available. See Devices and runtimes.

The execution boundary

A runtime belongs to a person, while the team may be allowed to send it work. Runtime access is therefore the execution boundary.

When an owner makes a runtime public, authorized tasks can use that runtime's enabled local skills and MCP tools. Publishing a skill or showing MCP metadata in the web app does not independently grant execution access. Review Security before sharing a machine.

Context that travels with work

  • The task and follow-ups supply immediate context.
  • Workspace and personal guidance supply lasting preferences without granting access.
  • Skills supply reusable procedures and supporting files.
  • When memory is on, every task reads notes saved in public channels, and a private-channel task also reads that channel's notes.
  • Surface connectors provide bounded Slack, Linear or GitHub context and tools.

Use Custom instructions to choose the right form of instruction and Shared skills for workflows that should be installed and reused.

What persists between follow-ups

KindWhat persists
Source conversationThe request, progress, results and replies used to continue the same work.
Saved guidancePersonal or workspace preferences that apply within their configured scope.
SkillA reusable procedure and supporting files installed for a local agent.
MemoryEligible channel-scoped facts, not a complete transcript.
Local capabilityFiles, credentials and tools available to the runtime selected for the task.

The source conversation preserves continuity, but the machine must still be connected and an eligible runtime must still be available. A different runtime may have different local files, credentials, skills and MCP tools.