1. Unified agent orchestration — not a chat box bolted on
Bidwright keeps the LLM provider adapter in packages/agent, while estimating actions live in the API, MCP server, and quote workspace services where they share the same cost intelligence, worksheet provenance, labor units, assemblies, review records, and audit trail. The agent can read a worksheet, search current market sources, pull structured cost candidates, validate labor and resource basis, propose phases, edit line items, and write review notes without drifting onto a separate legacy tool path.
AI-native construction estimating platform for intake, takeoff, pricing, scheduling, knowledge, and quotes.
ดาว
75
เครื่องมือ
—
อัปไทม์
—
Multi-provider out of the box: Anthropic, OpenAI, OpenRouter, Gemini, LM Studio. Local embeddings via Ollama. Hybrid retrieval over pgvector.
2. Bring your own coding agent — Claude Code, Codex, OpenCode, & more
Bidwright can spawn a Claude Code or Codex session against an isolated workspace seeded with the project's documents, knowledge, and CLAUDE.md context. Stream responses, read/write memory, monitor sessions, stop them. The same machinery powers the in-app review feature — and it's exposed at /api/cli/* for your own automations.
3. MCP server — Bidwright tools in your editor
The packages/mcp-server package exposes Bidwright's estimate, knowledge, model, quote, review, system, and vision tools over Model Context Protocol. Point Claude Code, Cursor, or any MCP-aware client at it and your estimating data is one tool call away from your IDE.
4. 2D takeoff and 3D model takeoff — linked to the worksheet
2D drawings: open PDFs, calibrate scale, count/linear/area annotations, symbol detection, multi-page counts, "Ask AI" on a selected region, export markups.
3D models: ingest BIM/CAD, parse element hierarchies (parent/child, material, system, discipline), extract per-element quantities with confidence/method tracking, generate filtered BOMs, and diff revisions (baseline vs. head).
Both link back to worksheet line items via TakeoffLink and ModelTakeoffLink — annotations and 3D elements stay tied to the dollars they drive, with quantity multipliers and override fields.
A dedicated Model Editor app (apps/model-editor/) ships alongside the main web app for heads-down 3D work, with bidirectional sync to the quote workspace.
5. Assemblies — reusable kits with parameters and nested sub-assemblies
Define an assembly once, drop it in a hundred quotes. Each assembly supports:
Typed parameters with defaults and units (AssemblyParameter)
Quantity expressions that bind to those parameters (quantityExpr, parameterBindings)
Knowledge Documents — hand-authored pages with structure and metadata
Datasets — structured tables built manually or extracted from books
Bind them to Estimator Personas — per-trade AI configurations with system prompts, default assumptions, productivity guidance, and review focus areas — so the agent answers like your senior estimator, not a generic LLM.
10. Plugins and dynamic tools
Drop in a plugin with a config schema and tool definitions, and it shows up in the agent's tool registry and the UI. Every execution is tracked (input, formstate, output, applied line items) so plugins are auditable, not magic.
Run Bidwright locally
The only thing you need installed is Docker Desktop. The installer
below pulls a few KB of launcher files (compose + start/stop scripts)
into ~/bidwright, then starts the stack — no source checkout.
Revision compare, package preview, branded PDF generation with configurable layouts and sections, email delivery.
AI & agents
Multi-provider models, 150+ tools, local Ollama embeddings, MCP server, Claude Code / Codex runtime sessions, plugin framework.
Quote review
Agent-driven coverage / gap / risk / competitiveness analysis with evidence-linked findings and calibration feedback.
Multi-tenant ops
Organizations, users, super-admin setup, org switching, brand profiles, brand capture from website crawl, estimator personas, data import/export, admin flows.
Inside the monorepo
apps/
api/ Fastify API — auth, quote logic, PDF/email, AI/vision routes, CLI runtime
web/ Next.js app — intake, estimating, takeoff, knowledge, performance, settings
worker/ BullMQ orchestration for ingestion and reviewable AI workflows
model-editor/ Standalone 3D model editor with bidirectional worksheet sync
packages/
agent/ Tool-backed agent runtime, provider adapters, 150+ tools
ai/ Prompt contracts and typed AI helpers
db/ Prisma schema (73 models), seeders, templates, db utilities
domain/ Shared business models and quote logic
ingestion/ Package extraction, classification, chunking, parsing
mcp-server/ MCP bridge exposing Bidwright tools to external agents
vector/ Embeddings and pgvector hybrid retrieval
vision/ PDF rendering, 2D symbol analysis, 3D model parsing
Develop on Bidwright
If you're going to edit code, you need Node.js 20+, pnpm 10+, and Docker
Desktop. Optionally an OpenAI and/or Anthropic API key.
pnpm install
cp .env.example .env
pnpm dev
pnpm dev brings up Postgres, Redis, and Ollama in Docker, generates the Prisma client, pushes the schema, sets up pgvector, and launches the web app, API, and worker together.
On Windows:
pnpm dev:windows
After startup:
Web: http://localhost:3000
API: http://localhost:4001
If no super admin exists, Bidwright opens the first-run setup wizard. From there, create your org and optionally load sample data.
Vision: Playwright, Python, OpenCV-style 2D symbol pipeline, 3D model parsing
Agent runtimes: Claude Code, Codex, MCP
Monorepo: pnpm workspaces + Turborepo
Language: TypeScript with Zod validation
Status
Bidwright is a working platform for AI-assisted construction estimating. Core estimating, takeoff (2D and 3D), assemblies, knowledge, pricing, scheduling, review, branding, multi-tenant admin (organizations, users, super-admin setup and org switching), plugins, MCP, and CLI agent workflows are present in the codebase today. Areas like deeper third-party integrations and performance analytics dashboards are still filling in. The platform is moving fast — especially around agent autonomy, calibration feedback, and the model takeoff layer.
BidWright is estimating. OpenBooks is the books. BeaconHS is safety. Built to work together for project-based companies. Same design language. Same self-host posture. Each repo is still its own install today.
B
Bidwright
braedonsaunders·75 ดาว·
AI-native construction estimating platform for intake, takeoff, pricing, scheduling, knowledge, and quotes.