Cycling through all six. Tap any point to stop.

Measured on six things

AI coding harness with plan review, validation, and engin...

Usability

One-line curl install and a clear plan-review-validate story, but no pricing, no visible docs, and the only CTA is a beta outreach form — signup-to-value path is a manual email.

Screenshots

1 / 1

RunWield

About this app

RunWield is a local-first AI coding harness that keeps engineers in control, from rough idea to validated, reviewed code. Inspired by engineering principles championed by Addy Osmani, Matt Pocock, and Poteto in their recent talks, posts, and skills, RunWield prioritizes planning, validation, and accountability over raw AI coding speed. Engineers collaborate with agents on detailed plans, adding annotations, redlines, and context before implementation. Specialized agents execute the work, followed by automated tests, type checks, linting, and repair loops. An independent LLM reviewer verifies the implementation against the plan. Guided human review enables line-level annotations and confident approval. Oversight scales with risk: quick fixes stay quick, while complex work gets deeper planning and architectural review. RunWield runs locally, supports any model subscription or API key, and avoids vendor lock-in. Plans and work records remain under your control.

Key features

1Plan-first workflow: annotate, redline, and shape the approach before any code is generated.
2Validation built in: tests, types, lint, an independent LLM reviewer, and guided human diff review.
3Local-first and model-agnostic: runs on your machine, any subscription or API key, no vendor lock-in.

Target customer profile

Drawn from the product itself, not from a survey.

Demographic

Experienced software developers working in existing codebases

Pain points

AI agents generate code without understanding context; changes break existing functionality; lack of validation and review.

Primary needs

Control over AI-driven code changes; validation before merge; integration with existing workflows.

Demographic

Engineering teams exploring AI collaboration in development

Pain points

Unclear where AI should fit; fear of losing engineering control; inconsistent quality from AI outputs.

Primary needs

Structured planning and validation; transparent AI reasoning; ability to review and approve changes.

Demographic

Tech leads and architects turning ambiguous requirements into code

Pain points

Translating vague product ideas into concrete implementations; ensuring AI-generated code meets standards.

Primary needs

Collaborative planning tools; automated validation; audit trail of decisions and changes.

Scout Review

Written by AI from measured evidence, scored out of 100.

RunWield is a local-first AI coding harness that bets engineers want proof, not just speed — plan review, mechanical validation, an independent LLM reviewer, and guided human review, all running on your machine with any model you like. The idea is sharp and clearly differentiated from Copilot, Cursor, and Aider, and the marketing page loads at a perfect 100/100 Lighthouse score. The execution gaps are front-of-house: a homepage that goes dark below the fold, no pricing, no visible docs, no security page, and a beta email form standing in for a signup. It's a serious product wearing an unfinished storefront.

Measure by measure

Usability

One-line curl install and a clear plan-review-validate story, but no pricing, no visible docs, and the only CTA is a beta outreach form — signup-to-value path is a manual email.

Design

Clean navy/teal hero with a scannable six-stage workflow bar, but the page dies below the fold — no sections, pricing, or footer visible in the render, and the install field is buried.

Speed

Lighthouse desktop 100/100 with LCP 401ms, CLS 0, and TBT 10ms — a genuinely fast static marketing page with no bloat to apologize for.

Security

HTTPS and a local-first architecture that keeps plans and records on your machine are real wins, but no security page, no SOC 2/ISO claims, and install is a curl | bash from raw GitHub.

Accessibility

Lighthouse accessibility 100/100 with readable contrast on the navy hero and teal highlights; no WCAG statement or multi-language support published, so it sits mid-band rather than top.

Growth

Real, specific pain (agents breaking existing codebases) with a genuine wedge — plan-first, local, any-model, independent reviewer — against autocomplete-first Copilot/Cursor/Aider. Early, but the idea has room.

Improvement areas

Homepage ends at the hero — no pricing, docs, or footer visible below the fold.
No security page, privacy policy, or audit claims despite running agents inside your codebase.
Onboarding is a curl | bash install plus a beta email form; no real signup or waitlist.
No accessibility statement or WCAG conformance claim behind the 100/100 automated score.

Design & Usability

The homepage is a confident, well-branded pitch: navy canvas, teal accents, a six-stage workflow bar that actually communicates the product's philosophy in one glance, and a headline that names the real fear — losing engineering control. Where it falls apart is below the fold. The render shows a large empty void where pricing, docs, and a footer should live, and the install command is a narrow, low-contrast afterthought. Usability follows the same pattern: the curl install is genuinely 30 seconds, and the plan-review-validate story is clear, but there's no pricing page, no visible documentation, and the only call to action is a beta outreach form that opens your email client. For a product selling control and rigor, the front door is oddly informal.

Speed & Security

Speed is the standout: Lighthouse desktop 100/100, LCP 401ms, CLS 0, TBT 10ms — a static marketing page with no bloat, and no reason to doubt it. Security is more of a mixed bag. HTTPS is there, and the local-first architecture is a real architectural win: plans, context, memory, and work records stay on your machine, and you can bring any model subscription or API key. But there's no security page, no SOC 2 or ISO claims, no privacy policy surfaced on the homepage, and the install path is a curl | bash script piped from a raw GitHub URL. For a tool that wants to sit inside your codebase and run agents against it, that trust story needs more than a one-liner.

Accessibility & Growth

Accessibility scores 100/100 on Lighthouse's automated axe-core checks, with readable contrast on the hero and teal highlights that don't fight the text. That's a strong baseline, but automated checks only cover part of WCAG — there's no accessibility statement, no WCAG conformance claim, and no multi-language support mentioned, so it lands mid-band rather than top. Growth is the more interesting question. RunWield names a real, specific pain: AI agents generating code without context and breaking existing functionality. Its wedge — plan before code, mechanical validation, an independent LLM reviewer, guided human review, all local and model-agnostic — is genuinely distinct from Copilot's autocomplete, Cursor's chat-first editor, and Aider's CLI edits. The category has obvious headroom as teams figure out where agents belong. The limit is execution and distribution: the only traction signal in the collected evidence is a LinkedIn profile, and the go-to-market is a beta email form.

Conclusion

If you're an experienced developer who's watched an agent confidently break your codebase, RunWield's plan-first, validate-always pitch is aimed straight at you — and the local-first, any-model architecture is the right foundation. Before you commit, expect to do some digging: pricing, docs, and security posture aren't on the homepage, and onboarding is a curl command plus an email. The bones are good and the philosophy is right. What's missing is the boring scaffolding — a real signup, a pricing page, a security statement, and a homepage that doesn't end at the hero — that turns a promising beta into something a team can adopt with confidence.

Suggestions for improvement

Build out the homepage below the fold: pricing, docs link, security page, and a real footer instead of a navy void.
Replace the beta email form with an actual signup or waitlist flow so users can start without opening their mail client.
Publish a security page covering the curl | bash installer, local data handling, and any audit or threat model.
Add an accessibility statement and WCAG conformance target to back up the perfect automated score.

How we score

Alternatives comparison

Named competitors, point by point. Nobody paid to appear here or to be left out.

	RunWield	GitHub Copilot	Cursor	Aider
Planning before code generation	Core feature: annotate, redline, and approve a plan before execution	No structured planning; suggests code inline as you type	Chat-based planning possible but no formal plan-review artifact	No plan-review loop; edits files directly from prompts
Post-generation validation and review	Automated tests/types/lint, independent LLM reviewer, guided human diff review	No built-in validation loop; relies on your existing CI	Diff review in editor; no independent LLM reviewer step	Runs tests optionally; no structured review or repair loop
Local-first / data ownership	Runs entirely on your machine; plans, memory, and records stay local	Cloud-hosted; code context sent to GitHub/Microsoft servers	Desktop editor with cloud model calls; local project files	Runs locally via CLI; sends prompts to chosen model API
Model flexibility / no vendor lock-in	Any model subscription or API key; switch providers freely	Locked to GitHub/Microsoft model stack	Supports multiple models but tied to Cursor's editor	Supports many providers via API keys
Pricing transparency	No pricing published; beta outreach form only	Published tiers: Free, Pro, Business, Enterprise	Published tiers: Hobby, Pro, Business, Enterprise	Open source, free; you pay your model provider