AI platform for research, creation, and exploration.

A conversation-first interface that expands into a multi-step research agent, generative UI for inline data and maps, and a canvas workspace that compiles single-file React artifacts live.

_{Polymorph end to end — canvas artifacts, generative UI, conversational workflows, and the evals dashboard. View the high-resolution video on GitHub.}

Features

• Canvas artifacts — generates and previews single-file React apps with live editing
• Multi-step research agent — searches the web, reasons across sources, and synthesizes answers
• Generative UI — tables, charts, geo maps, timelines, citations, callouts, and link previews render inline
• Three conversation modes — Search, Research, and Build share one chat surface with intent-aware prompting
• Geo intelligence — interactive maps, real directions, reachability polygons, and static map images
• Multi-provider AI — GLM-5.3 Flash/GLM-5.3 via OpenRouter for text, Meta Muse Image via Vercel AI Gateway for inline generation, plus direct OpenAI, Anthropic, Google, OpenAI-compatible, and Ollama providers
• Voice mode — speech input and text-to-speech playback
• Configurable guest access — optional instant search without sign-up, rate-limited per IP in cloud deployments

Technical highlights

• RLS-keyed multi-tenant Postgres — every user-scoped table is protected by Row-Level Security policies keyed on a per-request session GUC, set through withRLS/withOptionalRLS (lib/db/with-rls.ts)
• Generative UI as typed message parts — ~15 display, canvas, and image tools stream structured output that renders as tables, charts, maps, timelines, and interactive widgets (lib/tools/*, components/tool-ui/*)
• Single-artifact canvas — one React app compiled server-side to a persisted, versioned single-file HTML document and served client-side via iframe.srcdoc, with HMAC-SHA256-signed guest edit tokens that rotate on every write (lib/canvas/*)
• 9-evaluator LLM-judge pipeline — 3 deterministic + 6 LLM-judge evaluators run against Phoenix experiments, including a weekly regression canary on Railway cron (services/evals)
• End-to-end OpenTelemetry tracing — every LLM call and tool invocation is traced into Arize Phoenix, with production span-content masking (instrumentation.ts, lib/utils/telemetry.ts)

Documentation

Browse all documentation →

• Quickstart Guide — local setup, Supabase, auth, and a guided first search
• Architecture Overview — system design, data flow, tech stack
• Geo & Spatial Tools — geocoding, directions, isochrones, static maps, and displayGeoMap
• Environment Reference — all environment variables
• Deployment Guide — Vercel deployment and production config
• Contributing Guide — development workflow and quality gates
• API Reference — chat API endpoint and schemas

Architecture

_{Three-agent chat system with generative UI · See architecture overview for the full breakdown.}

Evaluation

_{Offline LLM-judge pipeline · Railway cron → Phoenix experiments → Postgres dashboard.}

Status & roadmap

• Canvas is one artifact per chat — createCanvasArtifact/updateCanvasArtifact/readCanvasArtifact are only registered when a chat has an active canvas context; there's no multi-artifact workspace yet
• OpenRouter/Z.ai is the default text provider — GLM-5.3 Flash/GLM-5.3 via OpenRouter, with direct OpenAI, Anthropic, Google, OpenAI-compatible, and Ollama providers supported for self-hosted or BYO-key setups
• Deferred dependency upgrades — Vitest 5, @vitejs/plugin-react 6, and ESLint 10 were attempted and rolled back (ESLint 10 is blocked on eslint-plugin-react compatibility); TypeScript 7 and the Drizzle ORM v1 beta are also not yet adopted
• Evals cron runs a low-cost weekly canary by default — the scheduled Railway cron replays a single synthetic regression case; full capability/traffic-monitor suites are triggered on demand from the Railway dashboard

Attribution

Polymorph is derived from miurla/morphic under the Apache-2.0 license. See LICENSE for details.