Iron Bridge Construction — MCP Equipment Safety Server

Company & Problem

Iron Bridge Construction manages heavy equipment (cranes, excavators, scaffolding) across multiple job sites.
Before this system, site workers used paper checklists. Nothing stopped an uncertified operator from taking a crane, and nothing forced a human supervisor sign-off when the work was near power lines.

The fix: an MCP server that gives an LLM scoped, safe access to equipment data. The model never talks to the database directly. Every write that carries real risk is gated by capability checks, role changes, elicitation, and sampling.

Protocol Concerns Mapping

#	Concern	How it appears in this system
1	Capability negotiation	Server declares elicitation + sampling. Client checks before offering high-risk tools.
2	Notifications	Worker starts with read-only tools. After authenticate_supervisor, server pushes tools/list_changed and approval tools appear.
3	Elicitation	request_equipment for high-risk items (crane / near power lines) pauses mid-call and asks a human supervisor for explicit confirmation.
4	Resources	Safety policies (lifting-safety, electrical-proximity) are exposed as resources, not tools.
5	Prompts	Reusable template prepare-equipment-receipt that takes {request_id}.
6	Sampling	Before final approval of a high-risk request the server asks the client's model to draft a short risk summary; result is stored in audit log.
7	Progress tracking	generate_site_compliance_report walks every request and reports 25/50/75/100 %.
8	Defensive tool design	Strict JSON Schema (additionalProperties: false), server-side jsonschema validation, parameterized SQL, and handler-level authorization.

Transport

• Development: stdio (default).
• Demo / multi-site: Streamable HTTP (MCP_TRANSPORT=streamable-http).

Commit history shows the transition from stdio-only to HTTP support.

Quick Start

python -m venv .venv
source .venv/bin/activate   # or .venv\Scripts\Activate.ps1 on Windows
pip install -r requirements.txt
python db/init_db.py
python agent/client.py      # starts stdio server automatically

HTTP mode:

export MCP_TRANSPORT=streamable-http
export MCP_HOST=127.0.0.1
export MCP_PORT=8000
python -m mcp_server.server
# endpoint: http://127.0.0.1:8000/mcp

Tool Comparison

Tool	Read/Write	Needs elicitation?	Needs sampling?	Notes
check_worker_certification	read	no	no	always available
get_equipment_status	read	no	no	always available
request_equipment	write	yes if high-risk	yes if high-risk	creates PENDING request
authenticate_supervisor	write (session)	no	no	triggers tools/list_changed
approve_equipment_request	write	no	no	only after supervisor auth
generate_site_compliance_report	read (report)	no	no	progress updates

If a client connects without elicitation or sampling capability, high-risk requests return NOT_SUBMITTED and nothing is written to the database.

Folder Layout

db/                 schema, seed, ERD, init script
mcp_server/         server, schemas, service layer, database helpers
agent/              demo client that performs the full handshake
docs/               progressive code parts for team commits
issues/             ready-to-paste GitHub Issue bodies (one per concern)

Team Workflow (4 sequential parts per concern)

See docs/ and issues/. Each concern is an independent GitHub Issue.
Code for each concern is delivered in 4 progressive commits so every teammate has a visible contribution.