Kamai
AI takeoff for construction: turn PDF blueprints into structured, traceable quantities.
Installation
Ajoutez ceci à la configuration MCP de votre client (Claude Desktop : claude_desktop_config.json ; Cursor : .cursor/mcp.json ; VS Code et les autres utilisent la même forme « mcpServers »).
claude mcp add --transport http kamai 'https://mcp.kamai.io/mcp'
{
"mcpServers": {
"kamai": {
"type": "http",
"url": "https://mcp.kamai.io/mcp"
}
}
}
Documentation
Tirée du README du projet sur GitHub, actualisée chaque semaine. Les liens ouvrent GitHub.
Sommaire
kamai-mcp
Kamai's MCP server: tools and MCP Apps widgets for Kamai construction-blueprint projects, blueprints and takeoffs. It runs against Kamai's MCP API and validates access tokens issued by Kamai's OAuth authorization server. It holds no Kamai secrets.
Tools
| Tool | What it does |
|---|---|
view_projects | Interactive panel of the user's projects and blueprints |
list_blueprints | Blueprints grouped by project, with their ids and whether they are ready |
view_blueprint | A blueprint page with its take-off shapes drawn on top |
view_takeoff | Measured take-off quantities as a sortable table |
list_elements | Per-element rows (rooms, walls, doors, windows, objects) with outlines |
list_folders | A blueprint's legend folders |
download_plan | A PDF of a processed blueprint with its detected elements highlighted, as a temporary download link |
update_elements, move_elements, create_folder | Rename, recolour and refile elements; create a folder |
set_opening_height | Store a user-stated height, or a door/window tag, on wall-surface opening pieces |
create_project, update_project | Create or rename a project |
list_jobs, get_job, cancel_job | Processing jobs of a project |
view_upload | Upload panel for a blueprint PDF |
ingest_blueprint_from_chat | Import a PDF attached to a ChatGPT message (ChatGPT only) |
open_kamai | Open the Kamai app in a panel |
With KAMAI_QUERY_TOOLS=on (see below):
| Tool | What it does |
|---|---|
count_elements | Counts and totals by Kamai category (doors, windows, walls, wet rooms…), grouped, summed in the database |
find_elements | Individual elements with type, folder, tag, handing and measurements |
calculate_wall_surface_area | Net wall (paint or plaster) area of rooms |
render_table | A table panel whose rows highlight their elements on the plan |
set_scale | Set a blueprint's drawing scale |
set_rotation | Turn how a blueprint is shown (display only) |
With KAMAI_HOST_PROBE on (the default on dev and debug, off in prod; see below):
| Tool | What it does |
|---|---|
probe_host | A diagnostic panel that tests what widgets can use in this chat (WebGPU, WebGL2, workers, WebAssembly, the GCS tile host) and logs the result as a widget_probe event |
list_blueprints also takes include_inventory (with a project_id): what each
blueprint contains.
list_projects, get_project, get_blueprint, request_blueprint_upload,
finalize_blueprint_upload, and with the query tools get_table_row_elements and
get_element_outlines, and with the probe report_host_probe, are app-only: callable by the widgets, not advertised to the
model.
Run locally with your API key
The quickest way to run this server against your own Kamai data is local mode: the host spawns the server, talks to it over stdio, and the credential is a personal API key from Kamai (API Keys in the Kamai app) instead of OAuth.
bun install bun run build:widgets KAMAI_API_KEY=... MCP_API_BASE_URL=https://mcp-api.kamai.io bun run local
Claude Code:
claude mcp add --env KAMAI_API_KEY=... --env MCP_API_BASE_URL=https://mcp-api.kamai.io \ --transport stdio kamai -- bun /path/to/kamai-mcp/src/local.ts
Claude Desktop (claude_desktop_config.json):
{ "mcpServers": { "kamai": { "command": "bun", "args": ["/path/to/kamai-mcp/src/local.ts"], "env": { "KAMAI_API_KEY": "...", "MCP_API_BASE_URL": "https://mcp-api.kamai.io" } } } }
The key is used as the bearer for every API call; the tools run as the key's owner.
Add KAMAI_QUERY_TOOLS=on to the environment for the query tools.
Run the HTTP server
bun install cp .env.example .env bun run dev
The values in .env.example are placeholders: set MCP_PUBLIC_URL and
MCP_AUTH_ISSUER to your tunnel and to a reachable authorization server before
starting, or every request will be refused with invalid_token.
| Variable | Meaning |
|---|---|
PORT | Listen port, default 8006 |
MCP_PUBLIC_URL | The URL hosts call and ask a token for (the resource in discovery); a tunnel, not localhost |
MCP_AUTH_ISSUER | Issuer of the access tokens, matched byte for byte against the token's iss |
MCP_AUTH_JWKS_URI | Optional. Defaults to {MCP_AUTH_ISSUER}/jwks.json |
MCP_API_BASE_URL | Kamai MCP API base URL, https://mcp-api.kamai.io. Also the audience an access token must carry |
MCP_ACCEPTED_AUDIENCES | Optional comma-separated audiences, for a cutover. Defaults to MCP_API_BASE_URL |
KAMAI_API_KEY | Local mode only: the personal API key the tools run as |
KAMAI_APP_ORIGINS | Optional comma-separated extra origins open_kamai may frame; the first becomes the default |
KAMAI_QUERY_TOOLS | on registers the query tools. Only once the API layer serves /v1/vocabulary (401, not 404) |
KAMAI_ROTATION | off drops set_rotation from the query tools (on by default; needs an API layer that serves PUT .../blueprints/{id}/rotation) |
KAMAI_HOST_PROBE | on or off registers probe_host and report_host_probe. Unset, it is on when Cloud Run's K_SERVICE starts with dev- or debug-, off otherwise. Never in local mode |
KAMAI_MCP_NO_INSTRUCTIONS | 1 leaves the server instructions out, as claude.ai does; for evals |
The MCP API and the authorization server are operated by Kamai, so running the
HTTP server end to end needs access to both. An MCP server is only really testable
against a real host, and a host can only reach a public HTTPS URL, so HTTP-mode
development means running behind a tunnel and pointing MCP_PUBLIC_URL at it.
bun run dev builds the widgets first: the server serves the built HTML from
dist/widgets/, not the React source, so rebuild after touching anything under
src/widgets/app/.
The widgets use React, TypeScript, Vite, Tailwind CSS, DaisyUI and Lucide. Vite produces one self-contained HTML resource per widget, so nothing needs hosting.
Local builds require Bun 1.3.3 and Node.js 22.12 or newer.
Layout
src/index.ts HTTP server, OAuth protected-resource metadata, MCP transport src/local.ts stdio entrypoint for local mode (personal API key) src/auth.ts Bearer-token verification (RFC 9728 challenges) src/server.ts Tool and resource registration src/api.ts Client for the Kamai MCP API src/elements.ts list_elements paging, filtering and result budgeting src/errors.ts The text a failed tool call shows the model src/query/ Query tools: descriptions, vocabulary, filters, budgets, contract src/widgets/ Widget resources, CSP metadata and the React widget sources scripts/ Widget build (escapes & so Open WebUI can embed the widgets); check-routes and check-vocabulary against a live API scripts/open-webui/ Open WebUI helpers: an open_kamai Tool, a display-mode host script, a local bridge demo
Privacy and support
- Privacy policy: https://kamai.io/privacy-policy
- Support: https://kamai.io/support or contact@kamai.io
Contributing
See CONTRIBUTING.md. Security issues go to the address in SECURITY.md, not to the issue tracker.
License
Apache-2.0. Copyright 2026 Kamai Enterprises LTD.
The Kamai name and logo are trademarks of Kamai Enterprises LTD and are not covered by the license.
Questions fréquentes
Qu'est-ce que Kamai ?
AI takeoff for construction: turn PDF blueprints into structured, traceable quantities.
Comment installer Kamai ?
Il fonctionne comme serveur distant. Dans Claude Code : claude mcp add --transport http kamai 'https://mcp.kamai.io/mcp'. Pour les autres clients, collez la configuration JSON de cette page dans leurs réglages MCP.
Faut-il une clé API pour Kamai ?
Son entrée au registre n'indique aucune clé ni secret obligatoire. Certaines fonctions peuvent tout de même demander un compte auprès du service auquel il se connecte.
Kamai est-il open source ?
Oui. Son code est sur GitHub sous licence Apache-2.0.
Qui maintient Kamai ?
Il est publié par KamaiEnterprises. AI.tn le référence sans le maintenir ni le recommander.
Issu du registre MCP officiel. Être listé n'est pas une recommandation : vérifiez le code et les permissions d'un serveur avant de lui donner accès à vos données.