SST Put Itself in Maintenance Mode to Build an AI Agent — 650K Devs in Five Months Shows Where Infra Builders Are Betting

The team that taught you to write your entire AWS stack as TypeScript stopped writing infrastructure tooling to write an AI coding agent — and 650,000 developers followed in five months.

SST, the TypeScript-native framework whose Ion engine deploys via Pulumi and Terraform, has been in maintenance mode since 2025: critical fixes still ship, but active development moved to OpenCode, a terminal-native, MIT-licensed, provider-agnostic agent built in Go. By April 2026 OpenCode reported 120,000+ GitHub stars, 75+ model providers, and 650,000+ developers; by June, tallies had it at 160,000+ stars in under a year — faster than Claude Code — with 900+ contributors and no marketing team.

This is not a one-off. Gitpod rebranded as Ona in September 2025 and sunset Gitpod Classic on October 15, 2025; Coder repositioned as agent infrastructure between April 2025 and February 2026. When infra teams see bigger distribution on the agent side, they follow it. The question for you is narrower: what does staying on SST cost once its maintainers' attention — and revenue motive — plainly sit elsewhere?

This post answers that. The box below is the whole decision on one screen. The rest shows receipts.

TL;DR — If you are on SST today

Tables and checklists below put numbers behind that summary — early, so you can find your row before reading the why.

What SST Is — and Why Teams Picked It​

SST is not "Terraform in TypeScript." It is a TypeScript program that defines infrastructure, frontends, and functions in one sst.config.ts, with real code — loops, conditionals, maps over tables — instead of HCL templating, plus sub-10ms live Lambda reloads via sst dev.

SST v3 (Ion) is the version most teams run today. Earlier SST sat on AWS CDK and CloudFormation; Ion replaced that engine with Pulumi and Terraform. Deploys got faster, CDK npm conflicts disappeared, and the model became "your app is a TypeScript program that provisions AWS."

Teams picked it for three reasons that still hold: one language everywhere if your app is TypeScript; one config file for functions, buckets, queues, and frontends; and sst dev for live development without redeploys.

How that compares to the paths you would actually take if you leave:

*Same stack rebuilt in each tool, 2026 benchmark. HCL is shortest raw; TypeScript is shortest once schema boilerplate is counted.

The point is not that SST is shortest. It traded raw provider breadth for app-level ergonomics — a bet that pays only while high-level components keep up with AWS.

The Pivot, Plainly Stated​

"Maintenance mode" is a sequence, not a single commit:

• 2024 — Ion ships. SST replaces CDK/CloudFormation with Pulumi/Terraform. Faster deploys, new component surface. This is the SST most teams use.
• 2025 — Maintenance begins. SST stays open-source and ships critical fixes, but active feature work stops. No new sst.aws.* components, no Ion iteration beyond stability. The repo goes quiet, not archived.
• June 2025 — OpenCode launches. Terminal-native, Go, TUI + CLI + desktop + IDE extension. Provider-agnostic (Anthropic, OpenAI, Google, local models via BYO key), MIT. Explicit thesis: as models commoditize, provider lock-in is the wrong bet.
• April 2026 — The 650K mark. Third-party tallies: 120K+ stars, 75+ providers, 650K+ developers. By June, 160K+ stars, 900+ contributors, and one widely-shared analysis claiming a 7.5M active ceiling. Treat 650K as the conservative floor the authors themselves used and 7.5M as a ceiling, not a baseline.
• August 2026 — Still diverging. OpenCode's changelog and opencode.ai/docs show weekly provider, TUI, and LSP work; SST's changelog is security bumps.

The maintainers proved TypeScript-as-infra, shipped Ion, then concluded the distribution curve for a TypeScript infra framework is flatter than for a terminal agent that works with any model. The next section quantifies why.

Why 650K in Five Months Rewrites the Economics​

One agent breakout is anecdote; four numbers are market:

1. The market is $4–5B today, growing mid-teens. SNS Insider (Gartner-referenced) values AI code assistants at $4.70B in 2025, projected to $14.62B by 2033 at 15.31% CAGR. VentureBeat's 2025 departmental survey had coding tools at $4B — 55% of all AI spend, the largest category. The order of magnitude, not the second decimal, is the signal.

2. Adoption is already mainstream. By end of 2025, ~85% of developers regularly used AI coding tools; in YC Winter 2025, 25% of startups had 95%+ AI-generated codebases. Copilot, Claude Code, and Cursor each crossed $1B ARR and together hold >70% of that $4B. They sell seats and tokens, not framework downloads.

3. The unit economics flip. An infra framework monetizes indirectly (consulting, sponsorships, hoped-for hosting). An agent monetizes directly per seat or per token. Gartner Peer Insights, June 2026: 23% of orgs already spend $200–$500 per developer per month on tokens for Claude Code/Cursor/Codex — hence Gartner's headline that agents "will cost more than real developers." That price is the vendor's margin and the builder's motive. Cognition AI (Devin) closed a $1B+ Series D on May 27, 2026 at $26B — more than double its $10.2B eight months earlier. No IaC framework raised at that velocity, because no IaC TAM told that story.

4. The pattern already held next door. Gitpod did not kill environments; it rebundled them as Ona's agent sandboxes (Sept 2025, gitpod.io sunset Oct 15 2025). Coder did not kill workspaces; it reframed them as agent boundaries (Apr 2025 → Feb 2026). SST did not kill components; it froze them as the stable substrate you already run but should stop expecting extensions from. The common move is "stop extending, rebundle as substrate" — which is why "still ships security patches" is simultaneously true and insufficient.

Sensitivity note: the conclusion does not need the 7.5M ceiling to hold. Whether OpenCode is 650K or 7.5M, 120K or 160K stars, the ranking holds: a terminal agent built by an infra team found more developers in five months than the infra framework found in years. How wide the gap is can be debated; that a gap exists cannot.

What "Maintenance Mode" Actually Costs You If You Stay​

If you are unsure where you sit, grep pulumi. inside sst.config.ts. One or two hits is normal. Five-plus in a medium stack is a quiet rewrite off SST without a decision memo.

Three typical profiles — find yours​

Stay or Go: Your Migration Timing Checklist​

Use this as a change-control gate, not a vibe.

Stay if all three hold — and re-check quarterly:

• Stack fits "small, stable" and you pin sst + provider versions with a tested upgrade path.
• New AWS services via raw Pulumi resources inside sst.config.ts do not break your team's review story.
• Losing one sst.config.ts owner does not block deploys.

Start planning an exit if any one is true:

• 2+ raw pulumi/aws or awscc resources in the last 90 days because SST had no component.
• A needed provider feature exists but SST has no wrapper and no merged PR in flight.
• Hiring "SST experience" has collapsed toward "Pulumi or Terraform with opinions."
• >20 resources or cross-stack outputs, and onboarding a new hire to sst.config.ts takes more than a day.

If you go, where to?​

There is no successor — SST was a layer on Pulumi/Terraform, not a third engine. Migration is layer removal.

Do not fork SST to "own the components" unless you will own the provider chase too. Forking is not free — you inherit tracking every AWS release and re-abstracting each resource. If that sounds like "build a second Pulumi component library," it is.

A Different Bet on the Same Thesis​

SST's read — distribution is in agents, not frameworks — is probably right for a vendor. It is the wrong bet to copy as a platform operator. A PaaS that abandons the platform to build an agent now competes with 900-contributor TUI repos and three $1B-ARR incumbents for seat share, instead of competing where the fleet is the moat.

The opposite wager is to make the platform agent-operable from day one, so the same 650K developers can deploy without leaving their terminal.

That is the wager behind Bex. Bex is a self-hosted, open-source Render alternative — push a git repo, get a running HTTPS service on machines you own — with a Render-compatible API and an open MCP server so any agent, including OpenCode, can call the platform like a human does. No managed databases, no multi-cloud abstraction, no closed SaaS — the fleet is yours. The agent is the operator, not the replacement for the thing being operated.

Whatever you do about SST, keep this migration constraint: pick a path whose control plane an agent can drive. The next infra moat is not a shinier sst.config.ts; it is a fleet whose API and machine-readable state an agent already understands.

SST still deploys real apps well today. The question is no longer whether it works, but whether you want to bet it will be the one extended. For a platform that is both agent-operable today and ownable tomorrow, that bet has already been placed — on your fleet, not someone else's token meter.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.

Sources

• SST / Ion (moving away from CDK announcement, sst.dev, Pulumi+Terraform engine, sst.config.ts model)
• Lushbinary Apr 2026 (120K+ stars, 75+ providers, 650K+ devs, MIT, Go, TUI/CLI/desktop/IDE) · Medium Jun 2026 (160K+ stars under a year, 900+ contributors, 7.5M ceiling)
• GitHub sst/opencode & opencode.ai docs/changelog Aug 2026
• Gitpod → Ona Sept 2025 / gitpod.io sunset Oct 15 2025 (The Register, Ona blog) · Coder Apr 2025 → Feb 2026 rebrand
• SNS Insider / Gartner 2025–33 AI code-assistant market ($4.70B→$14.62B, 15.31% CAGR) · VentureBeat 2025 dept spend ($4B, 55%)
• Gartner Peer Insights Jun 2026 (23% at $200–500/dev/mo, "cost more than developers") · Faros/vertika 70% share, three >$1B ARR · 85% dev adoption end-2025 · YC W25 25% with 95% AI-generated · Cognition $26B Series D May 27 2026 · Forrester/McKinsey 40% apps with agents by 2026, $150B infra spend by 2027
• Pulumi vs Terraform vs OpenTofu LOC benchmarks, 2026 three-tier VPC studies