Skip to main content
Glama
open-jobsite

open-jobsite

by raoulstongea

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: construction-supervision-mcp

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

Tool

Purpose

External side effect

create_project

Create a local JSON job record

Local file only

record_site_evidence

Record a sourced note, measurement, or media reference

Local file only

get_project

Read the complete local record

None

calculate_surface_area

Area plus stated waste factor

None

calculate_concrete_volume

Rectangular volume in cubic yards

None

calculate_sheet_count

Whole sheets plus stated waste

None

calculate_linear_pieces

Whole stock pieces plus stated waste

None

draft_daily_log

Draft labor summary linked to evidence

Local file only

draft_estimate

Draft priced scope with unit math

Local file only

draft_change_order

Draft scope/cost/schedule change

Local file only

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.

Available Tools

10 tools
calculate_concrete_volumeC

Calculate cubic yards for a uniform rectangular pour; not engineering advice.

Inputs: depth_in (required), width_ft (required), length_ft (required), waste_percent. Returns: object

calculate_linear_piecesC

Calculate whole stock pieces for a linear requirement.

Inputs: piece_length_ft (required), required_length_ft (required), waste_percent. Returns: object

calculate_sheet_countB

Calculate whole sheet count with explicit sheet dimensions and waste.

Inputs: area_sq_ft (required), waste_percent, sheet_width_ft, sheet_length_ft. Returns: object

calculate_surface_areaA

Calculate rectangular area in square feet, including a stated waste factor.

Inputs: width_ft (required), length_ft (required), waste_percent. Returns: object

create_projectB

Create a new local project. Use a lowercase, non-sensitive project_id.

Inputs: name (required), project_id (required), description. Returns: object

draft_change_orderC

Save a draft change order with structured lines and an approval gate.

Inputs: title (required), reason (required), line_items (required), project_id (required), +5 more. Returns: object

draft_daily_logC

Save a draft daily log with worker identifier, role, and hours entries.

Inputs: summary (required), workers (required), work_date (required), project_id (required), +3 more. Returns: object

draft_estimateA

Save an auditable draft estimate; never send, approve, or accept it.

Each line item requires description, quantity, unit, and unit_cost, with an optional evidence_ids array.

Inputs: title (required), line_items (required), project_id (required), evidence_ids (required), +5 more. Returns: object

get_projectA

Read the complete local project record.

Inputs: project_id (required). Returns: object

record_site_evidenceC

Record sourced evidence locally; this performs no upload.

evidence_type must be document, measurement, note, photo, receipt, sketch, or voice_transcript. publication_status must be private, synthetic, or permission_cleared.

Inputs: content (required), project_id (required), evidence_type (required), source_reference (required), +1 more. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 10 tools

    Disambiguation5/5

    Each tool targets a distinct construction or project-management action: four different calculation types, project creation/reading, evidence recording, and three distinct drafting operations. There is no meaningful overlap between the calculation tools or draft tools because the descriptions clearly differentiate their inputs and outputs.

    Naming Consistency5/5

    All tool names follow a consistent snake_case verb_noun pattern, such as calculate_concrete_volume, create_project, record_site_evidence, and draft_estimate. This makes the tool set predictable and easy for an agent to navigate.

    Tool Count5/5

    Ten tools is well within the ideal range and fits the server's scope: calculations, project records, evidence, and draft documents. Each tool has a clear purpose and none feel redundant or excessive.

    Completeness4/5

    The core workflow is well covered: create and read a project, record evidence, draft logs/estimates/change orders, and perform common material calculations. However, there are no update/delete project operations or direct evidence retrieval/list endpoints, which leaves minor workflow gaps.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Provides construction-supervision tools for small-to-medium building projects, including stage guidance, inspection checklists, progress tracking, next-stage recommendations, and quotation analysis. It enables local, standards-based MCP interaction over stdio without third-party MCP SDKs.
      5
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables local, private receipt management for users and their agents, with all BillCheck MCP tools for creating, updating, and sharing receipt data without requiring an account, cloud database, or model API key.
      AGPL 3.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables people and general-purpose AI agents to inspect, model, edit, verify, and export local 3D projects through MCP tools, including print preparation, without requiring a hosted 3D service key.
      10
      MIT