InsForge

Verified

Track: launchedWebsite ↗𝕏 @InsForge_devLinkedIn ↗

potential 81 · scoop 58

InsForge offers an open-source Postgres backend for AI coding agents that uses Model Context Protocol servers. The project beat Supabase and Postgres MCP in its own MCPMark test run and reached 9,900 GitHub stars after admission to YC P26.

Scores run 0 to 100. Potential: how likely the firm is to become a real winner, judged across the six dimension scores and held consistent with them, whether or not anyone has heard of it. Scoop: how unlikely a well-read tech investor is to have heard of it yet. High means it has not reached the outlets that reader follows, whatever the firm's quality or age. How our top potential scores did.

ShareXLinkedIn
After six YC rejections, two engineers released an open-source Postgres backend that now beats Supabase on MCPMark tasks in its own tests, at 1.6 times the speed with 30 percent fewer tokens.

Snapshot

What they build
Postgres-based backend with built-in auth, storage, compute, branching, and agent tooling for AI developers.
Who buys
AI engineers and agentic development teams building production apps.
Comparables
Supabase for agents · Firebase for AI-native stacks
Sub-industry
AI-native backend infrastructure

Hang Huang is ex-Amazon PM, Yale MBA, and former top League of Legends pro; co-founder has Databricks and hackathon experience.

How we found them

Surfaced in issue
May 17, 2026
Announced
May 17, 2026 (demo day)
Founded
2025
HQ
San Francisco, US
Discovery track
launched

Originating post ↗

Dimension scores

  • Team: 78 out of 100
  • Product: 82 out of 100
  • Traction: 68 out of 100
  • Market: 87 out of 100
  • Moat: 62 out of 100
  • Capital: 72 out of 100

Analysis

Hang Huang worked as a senior product manager at Amazon after his Yale MBA. Tony Chang built networking infrastructure at Databricks and won an open-source AI hackathon. The six-person team shares no common prior employers from elite organizations.

The product auto-generates APIs from Postgres schemas and adds a diagnose CLI plus skill.md workflows. These pieces reduce token counts for agents versus standard developer tools. The GitHub repository grew to 9,900 stars. In its own run on MCPMark tasks, InsForge beat Supabase and Postgres MCP. YC admitted the company to batch P26 after a direct meeting with partner Andrew Miklas.

MindWorks Capital and Baidu Ventures supplied early capital. The approach targets agent-specific operations inside a market filled with general-purpose backends. Early mover status on MCP primitives holds for now but remains open to rapid addition by larger players.

What they do

InsForge gives AI coding agents a managed Postgres backend with auth, S3-compatible storage, edge functions, compute, branching, realtime, and an MCP/OpenRouter AI gateway so agents can autonomously configure, debug, and ship full-stack apps end-to-end.

Why now

The 2025, 2026 explosion of production-grade AI coding agents (Claude Code, Cursor, Cline, etc.) created immediate demand for backends that agents can natively read, reason about, and mutate via structured protocols rather than scraping docs or dashboards. MCP tooling and the public MCPMark benchmark (where InsForge now leads) only matured in the last 6, 12 months; prior to that, agents lacked reliable context for backend state. Open-source self-hosting + hosted option removes friction that previously blocked agent-driven prototypes from reaching production.

Founding story

Hang Huang and Tony Chang developed the InsForge idea across multiple YC submissions, persisting through six total applications (four focused on this concept) without pivoting. After a fifth-application rejection following user interviews that validated demand among AI-native small teams and startups, they shifted to aggressive public building. A March 2026 Launch Week (InsForge 2.0, #1 Product Hunt, #1 GitHub Trending) produced enough momentum for a direct meeting with YC partner Andrew Miklas, who admitted them to the P26 batch starting March 30 despite an initial post-interview rejection. Tony left Databricks explicitly to pursue the startup. Hang brought product experience from Amazon post-MBA. The personal driver is explicit founder-problem fit: “we wanted this product so badly, and we were stubborn enough to believe this is the future. Or maybe you can just call us lazy developers who want AI to handle everything.” No public account details exactly how the two met, but the Seattle/SF tech + hackathon overlap (Tony’s 10× judging/winning record) is the most plausible vector.

Competitive position

Obvious incumbents: Supabase, Firebase, Convex, Xano. Closest startup competitor: Supabase (same Postgres + auth/storage/functions market). InsForge occupies the same lane but with a sharply different wedge, agent-native MCP integration, semantic context layer, and benchmark-proven efficiency gains, positioning it as the “Supabase for coding agents” rather than a direct feature clone.

Defensibility

Genuine moat elements include OSS community ownership (Apache 2.0, active contributors shipping features), data lock-in via Postgres branching + schema APIs, and switching costs once agents are wired to the MCP semantic layer and CLI diagnostics. First-mover advantage in agent-native primitives is real but narrow; Supabase and others can copy MCP endpoints relatively quickly. Brand and exclusive talent (founders’ combined PM + infra + hackathon depth) provide short-term defense. Overall moderate defensibility.

Network density

0 of 6 early core employees from any single prior elite organization documented

Traction

  • YC P26 acceptance
  • GitHub repo growth from ~1.5k (early March 2026) to ~9.9k stars (mid-May 2026) with #1 Trending and Product Hunt wins
  • MCPMark leaderboard #1 (1.6× faster execution, 30% fewer tokens, 1.7× higher accuracy vs. Supabase on Claude Sonnet)
  • multiple Hacker News Show HN posts with developer testimonials for production CRM, nutrition, and agent apps
  • active community contributors and X usage anecdotes

Opportunities

  • MCP ecosystem maturation and new agent benchmarks favoring lightweight context layers
  • YC P26 batch network effects for intros to agent-tooling startups
  • continued open-source contributor velocity (recent BYOK, edge editor, custom domain PRs from community)
  • potential pricing/availability shifts in foundation-model inference that further reward token-efficient backends
  • regulatory tailwinds around auditable AI agent tooling in enterprise settings

Risks / red flags

  • Six YC rejections before P26 acceptance signals possible early narrative or traction weakness
  • Traction remains heavily OSS/GitHub-star driven with limited named paying enterprise logos or public ARR
  • Early revenue was cited at only ~$300/month in one rejection round
  • Pre-seed funding from a Pan-Asia regional firm (MindWorks) rather than top-tier U.S. lead investors
  • Crowded competitive lane against well-capitalized Supabase
  • Small 6-person team shipping rapidly but exposed to key-person risk on technical execution
  • No visible regulatory prep (e.g., data residency or AI governance) despite global agent usage

Hidden bets

  • Model Context Protocol (MCP) servers for structured agent interaction
  • OpenRouter as default gateway (with BYOK support)
  • Postgres schema-to-API auto-generation
  • database branching
  • dedicated insforge diagnose CLI plus skill.md canonical workflow
  • purpose-built “agent operating system” layer rather than general BaaS

Key relationships

  • YC primary partner Andrew Miklas
  • MindWorks Capital
  • Baidu Ventures

Founders

Hang Huang, Co-founder & CEO
Education: Yale School of Management MBA (2021–2024), BSBA in Finance and Data Science, Northeastern University (Magna Cum Laude)
Prior: Amazon / Senior Product Manager
𝕏 @hanghuang_
Tony (Yaowen) Chang, Co-founder & CTO
Key projects: Simplifier (winner, Open Source Software AI Hackathon, Oct 2024)
Prior: Databricks / Software Engineer, Networking Infra; Meta / Intern, Ads ML Infra; Amazon / Intern, Backend Infra
𝕏 @tonychang430GitHub

Citations