cloud.abridge.services

Getting started

Updated 2026-08-29

The MCP is the day-to-day front door. No Frankie-specific install — you need Teleport's tsh client and an active Teleport session; no 1Password, Google token, gcloud, HTTP token, or local checkout.

Connect the MCP (about a minute)

  1. Sign in to Teleport and confirm you can see the Frankie MCP:
tsh login --proxy=abridge.teleport.sh:443
tsh mcp ls   # must list: prod-frankie-mcp-teleport

If it doesn't list prod-frankie-mcp-teleport, request Teleport access via Serval — you need the Google Groups acl-teleport and builders-all.

  1. Register an MCP named frankie-mcp:
# Codex
codex mcp add frankie-mcp -- tsh mcp connect prod-frankie-mcp-teleport

# Claude Code
claude mcp add frankie-mcp --scope user -- tsh mcp connect prod-frankie-mcp-teleport

For Cursor, add this inside the existing mcpServers object in ~/.cursor/mcp.json (don't replace your other servers):

{
  "mcpServers": {
    "frankie-mcp": {
      "command": "tsh",
      "args": ["mcp", "connect", "prod-frankie-mcp-teleport"]
    }
  }
}
  1. Restart your coding agent (or reload MCPs). If your session expires later, run tsh login --proxy=abridge.teleport.sh:443 again — don't add a second MCP.

Hand over your first PR

  1. Tell your agent what you want to change, then say "hand this over to Frankie." The agent calls frankie_handover. With a PR URL it triggers the run directly; without one, the server routes lite vs openspec and returns the plan recipe (a draft PR plus the READY TO BUILD label). You review the plan in the PR.
  2. Watch progress in the Apex Console — yellow = a human owes a decision, red = genuinely broken. Same data in your editor via frankie_status (no args = fleet list; with a PR URL = full detail, timeline, and a diagnosis when stuck).
  3. Already have code that just needs CI + review cleanup? Ask to "finish this PR" → frankie_finish takes it to green without re-implementing.
  4. Redirect a run in any state with frankie_update ("actually, use the v2 endpoint") — it classifies steer/reply/fix server-side. Want it local? frankie_takeover cancels the run and hands you the branch recipe.

The one trigger is the label AI FACTORY: READY TO BUILD — applying it by hand on GitHub works too. Retry = re-apply the label (or "hand it over again").

Other front doors

  • Linear: make the ticket build-ready, add repo: abridgeai/<name>, then delegate it to the Frankie app.
  • Slack: start a coding request with @APEX; APEX clarifies it and prepares the plan-only PR, then Frankie builds once the factory label is applied.

All doors converge on the same pull request, label lifecycle, build loop, and human merge decision.

CLI (repo onboarding only)

The CLI is only for repository operations — installing managed repo resources and seeding factory labels. Needs Node 22+ and PNPM, plus NPM access (pnpm login; file a Serval ticket if you lack it).

# set up a repo for the factory
pnpx @abridge/frankie repo install --target-path /path/to/your/repo

# seed the AI FACTORY labels on GitHub
pnpx @abridge/frankie pr labels add --target-path /path/to/your/repo

Re-run frankie repo install to upgrade a repo to the latest Frankie version.