Open Jobsite

Open Jobsite is an open, local-first construction operations server for goose and other Model Context Protocol (MCP) clients. It turns field evidence into reviewable calculations, daily logs, estimates, and change orders while keeping source references, unit math, assumptions, exclusions, and human approval boundaries visible.

This repository is a working v0.1 grant-proof MVP. It uses synthetic data, makes no network calls, and never sends a message, submits a price, places an order, or approves an artifact.

Why it exists

Field work starts with rough notes, photos, sketches, measurements, receipts, and conversations—not clean database rows. General-purpose agents can draft polished answers while hiding where numbers came from. Open Jobsite takes the opposite approach:

• Evidence first: records retain a user-supplied source reference and privacy status.

• Deterministic math: quantity tools use decimal arithmetic and expose formulas and intermediate values.

• Draft by default: every estimate and change order is explicitly unapproved.

• Local and portable: JSON files stay in a user-controlled folder and the tool surface is standard MCP.

• Field-shaped workflows: portable skills cover daily logs, scoped estimates, and change orders.

Related MCP server: Keystone MCP Server

Working flow

1. Create a local project.

2. Record a note or measurement with a source reference.

3. Run a deterministic quantity calculation.

4. Draft an artifact linked to the evidence ID.

5. Review and edit the draft outside the server before any external action.

Tools

Financial and job artifacts always return:

{
  "status": "draft",
  "requires_human_approval": true,
  "external_action_performed": false
}

Install and test

Requirements: Python 3.12+ and uv.

uv sync --extra dev
uv run pytest
uv run open-jobsite --version

Run the MCP server over stdio:

uv run open-jobsite

Synthetic demo

The demo/ folder contains a synthetic end-to-end workflow, a prompt, fixed expected results, a Windows goose CLI launcher, and a preflight that launches the real STDIO server. Run the preflight with:

uv run python demo/run_demo.py

The preflight is also part of GitHub Actions. It verifies the same evidence, calculation, estimate, daily log, and approval-gate flow intended for a future screen recording. A recording is not included in v0.1.

Data defaults to .open-jobsite-data/. Select another local folder with either --data-dir PATH or the OPEN_JOBSITE_DATA_DIR environment variable.

Connect to goose Desktop

Build and test the repository first. Then add a custom STDIO extension in goose Desktop with:

• Command: the absolute path to uv (where uv on Windows)

• Arguments: run --directory C:\absolute\path\to\open-jobsite open-jobsite

• Environment: optionally set OPEN_JOBSITE_DATA_DIR to a private job-data folder

Use only synthetic data in a public demo. The synthetic prompt has been run through goose CLI 1.46.0 on Windows; see the verification record. A Berd-local MCP run has not yet been verified.

Verified v0.1 evidence

As of 2026-08-27:

• 28 automated tests pass on Windows.

• all five synthetic benchmark cases pass, including four exact numeric cases and three approval invariants.

• the real STDIO preflight produces 108 square feet, four sheets, a CAD 344.00 subtotal, a CAD 378.40 total, and 6.00 labor hours.

• an isolated Windows goose CLI run produced two evidence records and two draft artifacts with the same expected values.

These are narrow software checks, not claims about field accuracy, estimating accuracy, code compliance, or user outcomes.

Example MCP arguments

Complex inputs use typed arrays so goose can see the required fields in the MCP schema. For draft_estimate, a minimal line_items value is:

[
  {
    "description": "Synthetic wallboard",
    "quantity": 3,
    "unit": "sheet",
    "unit_cost": 18.5,
    "evidence_ids": ["ev-example"]
  }
]

The evidence ID must already exist in the same project.

Repository map

• src/open_jobsite/: deterministic core, local store, artifacts, MCP tools

• skills/: portable field-workflow instructions for goose

• benchmark/: synthetic, machine-readable evaluation cases

• demo/: reproducible preflight, Windows goose launcher, and recording plan

• examples/: a synthetic example project record

• tests/: core and in-process MCP protocol tests

• docs/architecture.md: component and trust-boundary design

• docs/threat-model.md: current risks and mitigations

Safety and limitations

This alpha is not estimating, legal, contract, tax, structural, code, or safety advice. It does not inspect site conditions, validate plans, optimize cuts, confirm prices, or replace a qualified professional. A correct calculation can still be based on a wrong measurement or assumption. Review all output against current drawings, contracts, codes, manufacturer instructions, and field conditions.

Do not put tenant names, addresses, credentials, private photos, client records, or proprietary price data into public examples or bug reports. See SECURITY.md and docs/threat-model.md.

Roadmap

• Q1: stable schema, 25 benchmark cases, CI, Windows goose/Berd demo, upstream goose contribution

• Q2: voice/photo/PDF adapters, bilingual workflows, provenance review MCP App, three safely documented pilots

• Q3: revision history, change detection, local price-book adapters, security hardening, three external pilots

• Q4: v1.0, 50+ benchmark tasks, usability study with five field users, maintainer documentation

See CONTRIBUTING.md to participate. Apache-2.0 licensed.