Building a Visual Agentic Tool for AI-First Workflow Transformation

craft · Tech · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: AI-native work transformation

Craft needed an AI workflow tool that was genuinely useful for engineers and non-engineers after years of AI features that created initial interest but little repeat usage.

Constraints:

• Previous AI assistant, RAG, summarization, and retrieval efforts lacked stickiness.
• Non-technical users struggled with terminal-based Claude Code style workflows.
• Support, marketing, HR, finance, and engineering needed different data sources, write permissions, and workflows.
• Agentic workflows had to avoid exposing sensitive credentials while allowing APIs, databases, MCP servers, and Craft Docs as sources.

Architecture

Craft Agents is a visual, open-source wrapper around Claude Code and the Claude Agent SDK. It connects Craft Docs, APIs, databases, MCP servers, Zendesk, Linear, and backend APIs as sources; supports parallel agents, workflows, labels, themes, and skills; and provides read-only, ask-to-edit, and auto modes so non-engineers can run agent workflows without a terminal.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  craft-user[Craft User] --> craft-agents[Craft Agents]
  craft-agents --> claude-agent-sdk[Claude Agent SDK]
  claude-agent-sdk --> claude-code[Claude Code]
  craft-agents --> sources[Sources]
  sources --> credential-injection[Credential Injection]
  craft-agents --> permission-modes[Permission Modes]
  craft-agents --> workflow-skills[Workflow Skills]
  workflow-skills --> zendesk[Zendesk]
  workflow-skills --> linear[Linear]
  workflow-skills --> sources

Components

Data flow

A user starts a Craft Agents session, selects sources and a permission mode, invokes a workflow or skill, and the agent reads Craft Docs plus external sources such as Zendesk, Linear, backend APIs, databases, or MCP servers. For support cases, it enriches a ticket with user details, searches code/Linear for root cause, drafts a customer response, and creates engineering tickets when needed.

Why this architecture

Craft moved from product-embedded AI features to a visual agentic workflow tool because the useful breakthrough was not another text assistant but a source-connected, permissioned, non-terminal interface that made agents usable across the company.

Models used: OpenAI GPT-3.5 · OpenAI GPT-4o · Opus 4.5 · Claude Code · Claude Agent SDK

Scale

• Craft has more than 1 million active users and over 50,000 paying customers.
  “Craft has more than 1 million active users, over 50,000 paying customers, and an engineering team of 20 with a median tenure of nearly four years.”article, stated
• Craft has a 20-person engineering team with median tenure near four years.
  “background. Craft has more than 1 million active users, over 50,000 paying customers, and an engineering team of 20 with a median tenure of nearly four years. They care deeply about engineering excellence and product quality: the startup won Apple’s Mac App of the Year award in 2021, and built”article, stated
• About 80% of staff participated in the 2023 AI hackathon.
  “with around 80% of staff taking part, including designers, product managers, and customer support representatives”article, stated
• Craft Agents was built in two weeks and completed by January 5.
  “In the end, Balint hit his self-imposed milestone: by January 5 of this year, he had completed a polished Craft Agents app that ran on Mac, Windows, and Linux, thanks to it being built on top of Electron.”article, stated
• Support tickets dropped from 20-30 minutes to 2-3 minutes in the support team's account.
  “> We no longer need to look up subscription details in a separate tool because the agent fills in this detail. > > I feel that tickets which used to take 20-30 minutes to process are down to 2-3 minutes. >”article, stated
• A migration of 60 backend services finished in one week instead of the estimated 3-4 months.
  “They moved all 60 services to Serverless 4, migrated to Bun as the package manager, and deploy time dropped from 2.5 minutes to 1 minute.”article, stated
• Balint was spending about $3,000/month on top of a $200 Claude Max subscription.
  “The cost of models is, surprisingly, not an issue – at least not yet. Balint himself spends $3,000/month on top of the $200 Claude Max subscription, and is the heaviest user of tokens at Craft.”article, stated

Alternatives considered (roads not taken)

Ship AI features before a true wow moment

Rejected because: Craft wanted to avoid users associating AI with gimmicks or uselessness.

Terminal-only Craft Terminal interface

Rejected because: External beta users liked the tool but found multitasking, plan review, and non-technical use awkward.

"there were complaints about the terminal environment:

• Multitasking was painful

• Reviewing complex plans was awkward"

— article, stated

Use Zendesk as the full customer-support workflow UI

Rejected because: Craft increasingly treated Zendesk as a backend and preferred API-first workflow control through agents.

Merge open-source pull requests directly

Rejected because: Craft used PRs as high-fidelity feature requests and rebuilt ideas to preserve product coherence.

Patterns

tool-use-orchestration (agent workflow)

Craft Agents connects APIs, databases, MCP servers, Craft Docs, Zendesk, and Linear as sources and routes workflows through skills.

multi-agent-orchestration (agent workflow)

The tool supports running parallel agents for bulk triage and other parallelizable workflows.

safety-policy-enforcement (governance)

Permission modes define read-only, ask-before-edit, and automatic-write behavior per session.

human-in-the-loop-gate (governance)

Ask to Edit mode keeps user confirmation in front of write operations when appropriate.

privacy-preserving-processing (security)

API sources inject credentials outside the agent so the agent can call tools without seeing sensitive credentials.

Tradeoffs

Agent interface

Chose Visual UI wrapper around Claude Code over Terminal-only agent interface.

Rationale: The visual interface unlocked non-engineer use, parallel review, and easier plan/code-change inspection.

Source permissions

Chose Explore, Ask to Edit, and Auto modes over Unrestricted agent writes or read-only-only agents.

Rationale: Different workflows need different risk levels; onboarding often uses discussion first and automation later.

Model/tooling cost posture

Chose Use the best models and tools first over Prematurely optimize AI spend.

Rationale: Craft wanted to unlock productivity first, then design a workflow around the new capability.

Vendor strategy

Chose API-first customer-support backend over UI-first SaaS with weak APIs.

Rationale: Once agents own the workflow, slow custom vendor UIs become a bottleneck and APIs become the real product surface.

Gotchas (what bit them)

AI features can drive initial adoption without becoming useful repeat workflows.

Root cause: Craft's early AI assistant and RAG/summarization experiments lacked sticky, high-value workflows.

Resolution: Wait for a genuine wow moment, then deploy source-connected agent workflows that solve recurring team problems.

Terminal UX blocked non-engineer adoption.

Root cause: Multitasking, reviewing plans, and the terminal environment felt inaccessible to non-technical users.

Resolution: Rebuild the experience as a visual Electron app with email/Slack-like affordances.

AI-native change can disrupt engineer identity and retention.

Root cause: The value of remembering codebase details and hand-writing code decreased suddenly for some engineers.

Resolution: Acknowledge different adoption stages and create impossible-goal exercises to produce personal wow moments.

Accepting many open-source PRs can erode product coherence.

Root cause: PRs often optimize for local needs rather than the whole product direction.

Resolution: Use PRs as signals and rebuild the underlying feature in a coherent design.

Evaluation

Craft judged AI usefulness by user stickiness, workflow adoption, support-time reduction, engineering migration outcomes, company-wide usage, and whether teams could achieve previously impossible weekly goals.

Metrics: Repeat usage / stickiness · Support ticket processing time · Ticket escalation rate · Number of workflows/skills built · Linear issues completed per week · Migration duration · Deploy-time reduction · Model spend and subscription usage

Outcomes

• Customer support reported tickets dropping from 20-30 minutes to 2-3 minutes and fewer engineering escalations.
• Non-engineering teams became heavy users, including support, marketing, HR, and finance.
• A 60-service migration finished in a week, with deploy time dropping from 2.5 minutes to 1 minute.
• Craft Agents was open sourced under Apache 2.0 and later noted as acquired by Polymarket in the article update.

The article is source-rich but has family-tie context and narrative framing; metrics are Craft-reported.