Quick start

Prove the first real task after installation.

Role: this page starts after herdr-mcp is installed and connected. For the executable install contract, use Agent install. For manual installation and operations, use Installation.

Do not redeploy Edge, reinstall the runtime, or add the browser extension just to follow this page. The goal is simpler: prove that Web AI actually reaches your workstation and can complete one verifiable task in a real repository.

1. Start with a read-only check#

Open a new ChatGPT conversation and ask:

CopyShow the current Herdr workspace and whether its managed project has uncommitted changes. Use one read-only baseline and do not modify anything.

A healthy path looks like:

Copyherdr_inspect
  ↓
answer from workstation facts

This proves more than a green Connector badge: the public MCP path, workstation Link, runtime, and Herdr workplace are actually connected.

If this step reports workstation_offline, zero tools, an OAuth loop, or a managed-root denial, stop before testing mutations and use Troubleshooting to locate the failing layer.

2. Make one deterministic change#

Choose a safe, easy-to-verify change such as a documentation correction, a test fixture, or a narrowly scoped configuration edit. Ask the Web planner to:

CopyReuse the current workspace. Read only the Git/file state needed for this change, make one small edit, and combine the relevant verification and final diff where safe. Do not delegate to a local agent.

The desired shape is:

Copyreuse baseline → focused read → patch → verification + diff

This proves the Web planner can perform deterministic work directly instead of turning every edit into another Coding Agent task.

3. Then try a task that deserves delegation#

Delegate only when independent reasoning, parallel investigation, or a longer execution really helps. For example:

CopyInvestigate the root cause of this failing test and implement the narrowest fix. Keep unrelated files unchanged, then re-check the Git diff and tests yourself.

The worker's final prose is not the source of truth. The Web planner should re-read repository state, tests, and runtime facts before accepting the result. Agent selection, timeout, and alternate-worker rules live in Agent delegation.

4. Add the browser extension only for long-running Web work#

Standard MCP already covers files, Git, shell, agents, and multiple workspaces. Add the extension when you specifically need:

• local working / progress / settled events to return to the correct Web conversation;
• page recovery and long-conversation handoff;
• Chrome Side Panel visibility into workspaces, panes, and agents;
• Queue, which sends an explicit next-turn user instruction after the current reply ends.

Keep Auto off on first use and verify binding and live state manually. See Browser extension for the product overview, Browser Continuity for the return path, and Browser Control Center for Side Panel operations.

5. What counts as a successful first experience#

The minimum acceptance requires all four:

1. Web AI can read the real Herdr workplace and repository;
2. a small deterministic edit is verified and visible in the real Git diff;
3. when delegation is useful, the Web planner independently verifies the worker's result;
4. if the extension is enabled, a long task can safely return to the correct conversation without replaying the original mutation.

At that point installation is over. Continue with Best practices for daily workflow and Architecture for the system model.