Skip to content
Build Perch

Features

Everything in one local-first workspace.

Build Perch brings provider-flexible models, a deep tool surface, native file and media generation, and first-class automation into a single Windows desktop app.

built-in tools
95
models via OpenRouter
400+
providers, cloud to offline
3
local media-editing tools
10
workflow trigger types
6
on-device memory & search
100%

A model layer that never locks you in

Connect one OpenRouter key and the model picker fills with the live catalog (GPT, Claude, Gemini, Llama, Mistral, and hundreds more), organized by what each model can do, with price-tier badges and a per-model rate link. Install Ollama and your local models appear alongside, fully offline. Add an ollama.com subscription and Ollama Cloud’s hosted models join too.

  • New models appear the day OpenRouter lists them. No app update needed.
  • Each chat is pinned to a model but switchable any time, and every reply records the exact model that produced it.
  • Per-model capability detection adapts the UI to what each model can actually do: vision, tools, audio, image generation, context size.
  • Ollama Cloud runs bigger open models without a local GPU.

95 built-in tools, scoped by modes & skills

When the AI needs to actually do something (read a contract, fix a spreadsheet, search the web, build a slide deck) it calls one of 95 built-in tools, and each call shows up as an expandable activity row: what it’s doing, with what arguments, and how it went.

  • 76 tools work with zero account connections; 19 more switch on when you connect Gmail, Discord, or Notion.
  • Anything that would leave your machine stops at an approval banner showing exactly what is being sent and where. Those sends can never auto-approve.
  • 5 modes and 23 skills keep the model carrying only what the task needs, and it can visibly pull in an out-of-bundle capability mid-turn.
  • Research the web, read and search across 75+ file types, and recall past conversations.
  • Create Word, Excel, PowerPoint, PDF, HTML, and SVG files, fill your own branded templates, and edit everything but PDF in place: Office edits preserve themes, charts, and masters, and HTML, SVG, and code get precise anchored edits, not full rewrites.

A task board you share with the AI

The Lists panel is a Kanban board: cards moving through stages you define (To Do, In Progress, Done by default, all renamable). Cards carry notes, deadlines, a priority scale, colored categories, flags, attached files, and links. The AI works the same board through its card tools, so it can turn a meeting transcript into cards or update stages, with each change surfaced for your approval.

  • Send your files, a list, or saved work straight into a chat as context, handed to whichever model you have connected.
  • Explorer-style file browser with drag-move, previews, and an in-app viewer for documents, spreadsheets, code, and media.
  • Multi-workspace: each workspace is just a folder you switch between freely, with a Manage Workspaces window for sizes, archiving, export, and cleanup.

Local semantic memory (100% on device)

Two layers, both inspectable. Workspace memory is an opt-in, visible markdown file the AI curates about that workspace; open it, edit it, or revert a bad rewrite. Conversation memory is measured, not guessed: a live context meter shows real progress toward summarization, and a cheap summarizer model condenses older turns in the background while your main model keeps chatting.

  • On-device embeddings index your workspace files across 49 file types, including Word, Excel, PowerPoint, and PDF, plus prior chats, task cards, and fetched pages. The model is bundled and runs on your CPU; nothing ships to a cloud.
  • Hybrid retrieval fuses by-meaning and keyword search in every query, so exact codes and identifiers land even when meaning alone would miss them.
  • Delta-scan indexing picks up external edits and cloud-sync changes automatically.
  • Context is never silently dropped, and the AI can search your other chats when asked.

Media generation & on-device editing

Generate images, speech, music, and video right in chat through OpenRouter media models, including image-to-image and image-to-video, with inline players and drag-out to your files. Editing is the opposite of cloud: 10 FFmpeg-backed tools run entirely on your machine.

  • Text-to-image with image-to-image from attached or sketched references; speech with per-chat voices; music from a description and lyrics; asynchronous video.
  • Trim, join, split, reframe to vertical, clean audio, extract audio, and burn captions, all processed locally.
  • Transcribe audio and video on-device with local Whisper. Your footage never uploads anywhere.
  • Dictate anywhere: the mic button transcribes your speech on-device with local Whisper, and no audio ever leaves your machine.

Agents, workflows & triggers

Agents are AI workers you save and reuse: a standing prompt, a model, a tool allow-list, optional scoped folders, and an optional validator model that can veto bad tool calls. Workflows chain steps in order, pass named outputs forward, and run on triggers.

  • Six trigger types: manual, schedule, incoming email, Discord message, file changes, and chaining on another workflow.
  • Every run gets a History card with one-click Revert that restores every file the run touched.
  • Runs can be confined to a single input folder and a single output folder, and unattended sends only go to destinations you pre-authorized.

Curated integrations, not plug-ins

Discord and Notion connect through their official APIs, Gmail through standard IMAP/SMTP with an app password, and GitHub over MCP, running GitHub's own first-party server. There is no plug-in marketplace, by design.

  • Every secret is written once into the Windows credential vault and can never be read back by the interface. Only the native backend ever uses it.
  • No arbitrary third-party plug-in touches your data. That is a security posture, not a gap.
  • Every integration stays off until you connect it, and any outbound action needs your approval.