Frank — the course MCP server

The teaching build for Claude for Engineering Teams — From Chat to Agent Pipelines, Sections 5–7. An MCP server that runs on AWS EC2 and gives every agent your team already uses — Claude Desktop, Claude Code, VS Code, GitHub Copilot — typed, audited, least-privilege tools over your infrastructure and GitHub pipeline.

One server. Every agent your team already uses.

This is a clean-room teaching build: small on purpose, every tool under ~40 lines, built live on camera in two passes.

The toolbox

Pass	Lecture	Tool	What it answers
1 — infra	5.3	instance_health()	"How's the box?"
1 — infra	5.3	service_status(name)	"Is shipit up?"
1 — infra	5.3	tail_log(service, lines)	"What's it logging?"
1 — infra	5.3	deploy_status()	"What's deployed?"
2 — pipeline	5.4	list_workflow_runs(limit)	"Is CI green?"
2 — pipeline	5.4	get_failed_jobs(run_id)	"What failed?"
2 — pipeline	5.4	get_job_log(job_id)	"Why did it fail?"
2 — pipeline	5.4	pr_checks(pr_number)	"Can this PR merge?"

Design rule (Lecture 5.3): each tool does one job with typed inputs and a small, capped output. Agents compose tools; we don't build "do everything" tools.

Security posture (Lecture 5.5)

• Credential boundary: Frank holds the IAM role and the GitHub PAT. Laptops and prompts never see a secret.
• Fine-grained PAT: one repo, Actions: read. Frank observes; the repair agent writes with its own credential.
• IAM: read-mostly, region-pinned, no wildcards (deploy/iam-policy.json).
• Transport: bearer-token auth, private bind, tunnel/VPN access only.
• Allowlists everywhere: services Frank may inspect are enumerated in config, not discovered.
• Audit: every tool call is a loggable event; systemd hardening in deploy/frank.service.

Quick start (local, no AWS needed)

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
export FRANK_AUTH_TOKEN=$(python -c "import secrets; print(secrets.token_urlsafe(32))")
.venv/bin/python -m frank.server
# then: claude mcp add --transport http frank http://127.0.0.1:8720/mcp \
#         --header "Authorization: Bearer $FRANK_AUTH_TOKEN"

instance_health, service_status, and tail_log read /proc and systemd, so they return real data on any Linux box. GitHub tools need FRANK_GITHUB_TOKEN + FRANK_GITHUB_REPO (see .env.example).

EC2 deployment, IAM, and client wiring: deploy/ec2-setup.md.

Repo layout

frank/
  config.py        # all configuration, all from env
  tools_infra.py   # pass 1 — Lecture 5.3
  tools_github.py  # pass 2 — Lecture 5.4
  server.py        # FastMCP wiring + auth — Lectures 5.5–5.6
deploy/
  frank.service    # systemd unit, hardened
  iam-policy.json  # read-mostly instance role
  ec2-setup.md     # the on-camera deployment walkthrough

Recording checkpoints

Tag the repo as you record so students can check out each lecture's state: git tag lecture-5.3 after pass 1, lecture-5.4 after pass 2, lecture-5.5 after auth + hardening.