Delv
Codingby Magic★ 3.2

Magic

Frontier code model lab building a superhuman coding agent with ultra-long context (100M tokens) aimed at end-to-end software delivery.

Try it
C
Safety & Trust

Delv Safety Grade: C

Score 54/100 · assessed 2026-04-18

Maintainer65
Permissions40
Supply chain30
Transparency35
Incidents100

Magic is a well-funded frontier AI lab (backed by notable investors including Eric Schmidt and Nat Friedman) building autonomous coding agents with extreme context windows. However, the offering presents significant transparency concerns: no public repository, no open-source code, waitlist-only enterprise access with undisclosed pricing, and minimal technical documentation about capabilities or safety measures. The autonomous agent nature implies broad permissions including filesystem write, shell execution, and repository modification. Without verifiable supply chain artifacts (no npm package, no GitHub releases, API-only delivery), users must trust Magic's internal security entirely. The company appears legitimate with real backing and a credible team, but the closed nature and powerful autonomous capabilities warrant caution until more operational details emerge.

Green flags

  • Well-funded lab with credible backing (Eric Schmidt, Nat Friedman)
  • No known security incidents or breaches to date
  • Focused on enterprise use cases with presumed vetting
  • Frontier research team with published technical work

Red flags

  • No public repository or open-source code available
  • Waitlist-only with no transparent pricing or access terms
  • Autonomous agent with likely broad filesystem and shell access
  • No verifiable supply chain artifacts or package distribution
  • Minimal documentation on safety measures or capability limits

Permissions requested

Read filesWrite filesShell executeRepo readRepo writeOutbound networkExternal LLM call
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Pricing

ENTERPRISE — Waitlist

Platforms

api

Review

Magic positions itself as the endgame for coding agents: a 100-million-token context window that can supposedly hold an entire codebase in working memory, then reason across it to ship features end-to-end. The pitch is seductive. Most agents choke on anything larger than a microservice. Magic claims it can refactor a monolith, trace dependencies across dozens of files, and deliver pull requests that actually pass CI. I haven't used it. Magic is invite-only, enterprise-tier, and the company has shared almost no public demos beyond glossy marketing collateral. The waitlist has been open for months. What we know: they've raised serious capital, hired research talent from OpenAI and Anthropic, and are training frontier models in-house rather than wrapping GPT-4. The 100M context claim is technically plausible but unverified in practice. The autonomy promise is ambitious. Where GitHub Copilot suggests lines and Cursor suggests blocks, Magic allegedly takes a Jira ticket and returns a branch. That would be transformative for teams drowning in grunt work: dependency upgrades, API migrations, refactoring legacy modules. But autonomy at that scale introduces new failure modes. A model that can touch 500 files can also break 500 files. Without tight guardrails, you're trading velocity for systemic risk. The competitive set is thin because nobody else is really playing this game yet. Devin (Cognition) is the closest analogue, also targeting end-to-end task completion, but it's more focused on sandboxed environments and junior-level work. Magic seems aimed at senior-level refactors and architectural changes. If it works, it's a different category. If it doesn't, it's vaporware with a nine-figure valuation. The lack of transparency is frustrating. No public benchmarks, no case studies, no pricing. Enterprise-only usually means six figures minimum. For teams already spending that on engineering headcount, the ROI could be obvious. For everyone else, it's a wait-and-see. I'd love to test it on a gnarly Rails monolith migration, but until Magic opens access or publishes real-world results, this is speculative.
Verdict

If you're an enterprise engineering org with budget and patience, the waitlist might be worth joining. For individual developers or smaller teams, there's nothing here yet. Wait for public access or credible third-party reviews.

Good at

  • 100M token context window could handle entire codebases in-memory
  • Built on proprietary frontier models, not reliant on OpenAI API limits
  • Targets end-to-end feature delivery, not just autocomplete
  • Serious research pedigree and funding behind the project

Watch out

  • Invite-only with no public access timeline
  • No published benchmarks, demos, or case studies
  • Enterprise pricing likely prohibitive for most teams
  • Autonomy at scale introduces high risk without proven guardrails
  • Unproven in real-world production environments

Use cases

  • long-context coding
  • research-grade models
  • enterprise engineering