Last updated 08-12-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
pre.dev
pre.dev is a project management and autonomous coding agent built for long-horizon software builds. You describe an app or import a repo, and pre.dev researches the goal, drafts milestones, user stories, architecture, and a tech stack before writing code. Focused agents then tackle each task in isolated context, verify outputs, and keep running across sessions without losing state.
Unlike single-prompt coding assistants, pre.dev treats planning as a first-class step. Its Deep Spec flow decomposes work into milestones, acceptance criteria, and hour estimates, then executes the roadmap in one run. On Terminal-Bench 2.0 (89 tasks, pass@1), pre.dev paired with Haiku 4.5 scored 41.6% while Claude Code on Sonnet 4.5 scored 38.2%, and pre.dev on Sonnet 4.6 beat Claude Code on Opus 4.5 (56.2% vs 53.9%).
Startup founders, indie hackers, and engineering leads use pre.dev to ship full-stack apps from a plain-language brief. The platform also offers Browser Agents for scripted web automation via MCP or API, plus an RL Environments catalog mined from private production repos for frontier model training. Install the CLI with curl, connect GitHub, and kick off builds from the web UI or terminal.
Deep Spec decomposes projects into milestones, user stories, and hour estimates before any code ships
Terminal-Bench 2.0 pass@1: pre.dev + Haiku 4.5 scored 41.6% across 89 benchmark tasks
Browser Agents run web tasks at $0.0159 average cost with 4.9s median wall time per task
Free plan includes 100 credits and 1 Browser Agent task; Pro adds 2,500 monthly credits for $199
CLI installs via curl -fsSL https://pre.dev/install | bash with MCP server at api.pre.dev/mcp
RL Environments catalog draws from 1,000+ private production repos with 15-30 graded checks each
Plans milestones, stories, and architecture before agents touch code, which cuts context loss on long builds.
Published Terminal-Bench 2.0 scores beat Claude Code while using a cheaper model tier.
Browser Agents and coding agents share one credit system with CLI and MCP access.
Open trajectory logs let you audit every agent decision on benchmark runs.
Free tier includes architecture planning, GitHub automation, and one browser task to test the workflow.
Deep Spec and higher credit limits require the $199 Pro plan or Enterprise contract.
Browser Agent benchmarks compare default tiers, so costs vary if you switch models.
Heavy compiler debugging and multi-branch git tasks still trail brute-force reasoning in pre.dev's own benchmark notes.
What does pre.dev do?
pre.dev is an autonomous coding and project planning agent. You describe software in plain language, pre.dev builds a roadmap with milestones and acceptance criteria, then focused agents write and verify code across isolated tasks until the build finishes.
How much does pre.dev cost?
pre.dev offers a Free plan at $0 per month with 100 credits and 1 Browser Agent task. Paid tiers include Plus at $25, Premium at $49, and Pro at $199 per month with up to 2,500 credits and roughly 25,000 Browser Agent tasks on Pro.
How do I install pre.dev?
pre.dev installs from the terminal with curl -fsSL https://pre.dev/install | bash. You can also add the MCP server in Claude Code using claude mcp add --transport http predev https://api.pre.dev/mcp and authorize in your browser.
What is Deep Spec on pre.dev?
Deep Spec is pre.dev's planning mode that researches your goal and outputs milestones, user flows, tech stack choices, user stories, architecture notes, and a roadmap before coding starts. Pro and Enterprise plans include Deep Spec access.
Does pre.dev include browser automation?
Yes. pre.dev Browser Agents navigate websites, interact with pages, and extract structured data from a URL or CSV upload. The product page reports 2.6x lower average cost and 5.1x faster median runtime than Browser Use Cloud on 100 real-world tasks.
How does pre.dev perform on benchmarks?
On Terminal-Bench 2.0 with 89 tasks at pass@1, pre.dev with Sonnet 4.6 scored 56.2%, beating Claude Code on Opus 4.5 at 53.9%. pre.dev publishes open trajectory data so each agent decision can be replayed from public harbor job logs.