The estimating platform that runs the whole bid — intake, knowledge, takeoff, pricing, scheduling, review, and quote delivery — on top of an AI agent that actually has hands.

Drop the bid package. Index the spec. Take off the drawings and the model. Parameterize the assemblies. Price it with real burdens. Let the agent review it. Send the branded quote.

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.

Multi-provider out of the box: Anthropic, OpenAI, OpenRouter, Gemini, LM Studio. Local embeddings via Ollama. Hybrid retrieval over pgvector.

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.

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.

• 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.

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)
• Nested sub-assemblies (AssemblyComponent.subAssemblyId)
• Per-component overrides for cost, markup, and UOM
• Snapshotting so worksheet items remember the source assembly state at the time of insertion

This is the layer most takeoff tools punt on. It's shipped here.

Bidwright's EstimateStrategy walks a quote through Scope → Execution → Packaging → Benchmark → Reconcile → Complete, with:

• A scope graph of items, constraints, and alternates with confidence levels
• An execution plan (self-perform vs. sub, crew strategy, procurement risk)
• A package plan splitting work into detailed / allowance / historical / subcontract by phase
• Benchmark comparables against prior projects
• Assumptions with evidence and explicit user confirmation flags

You're not just bidding — you're keeping receipts.

Hit review and Bidwright spawns an isolated agent session that ingests the project documents and your estimate, then produces:

• Coverage findings (YES / VERIFY / NO) tied to spec evidence
• Gaps, risks, and cost anomalies
• A competitiveness score
• Recommendations with status tracking (resolve, dismiss, defer)
• An EstimateCalibrationFeedback record that closes the loop on systematic over/under-pricing

You see exactly which document chunk the finding came from. No hand-waving.

• Rate schedules with tiered multipliers (regular / overtime / double / custom tiers)
• Labour cost tables by trade and role
• Burden periods with date-ranged percentages
• Travel policies with per diem, mileage, fuel surcharge, accommodation, and embed modes
• Catalogs for equipment and material lookups
• Reusable conditions library

Built for the estimator who knows that "rate × hours" is the easy part.

Three tiers, all org- or project-scoped:

• Knowledge Books — ingested PDFs (estimating books, specs, manuals) chunked, embedded, and searchable
• 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.

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.

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.

Windows (PowerShell):

iwr -useb https://raw.githubusercontent.com/braedonsaunders/bidwright/main/scripts/launcher/install.ps1 | iex

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/braedonsaunders/bidwright/main/scripts/launcher/install.sh | bash

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

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.

• Frontend: Next.js 16, React 19, Tailwind CSS, Radix UI
• API: Fastify 5
• Worker orchestration: BullMQ
• Database: PostgreSQL, Prisma, pgvector
• AI: Anthropic, OpenAI, OpenRouter, Gemini, LM Studio, Ollama embeddings
• 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

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 licensed under the GNU Affero General Public License v3.0 only (AGPL-3.0-only).

Copyright (C) 2026 Braedon Saunders.