Cosine — Summary

Cosine is a UK-based AI coding company (cosine.sh) whose primary product is Lumen — a family of production-first coding models optimized for maintainability, minimal diffs, and real engineering work in enterprise and niche languages (C, R, Matlab, Fortran, Verilog, Rust). The product has three surfaces: Cosine Desktop (native app for active engineering with multi-window workflows and visible agent state), Cosine Cloud (parallel cloud execution for long-running tasks), and Cosine CLI (cos binary, installable via Homebrew). An enterprise air-gapped deployment option exists. Cosine received UK government backing as part of a £675M sovereign AI strategy in April 2026.

Cosine's distinctive architectural bet is "production-first" models: trained with an 8-step data pipeline that produces verifiable training trajectories, with behavioral RL on "scope discipline, honesty, evidence, and plan calibration." The benchmarks (Niche-Bench, Vibe-Bench, Slop-Bench) are proprietary custom benchmarks testing characteristics mainstream benchmarks miss.

Closest seed: kiro — both are closed-source commercial IDE/agent products with proprietary model-model integration. Cosine differs by having its own first-party models (Lumen family) rather than routing to 3rd-party LLMs, and by being cloud-hybrid rather than local-only.

Cosine — Overview

Origin

UK-based AI company. Tagline: "The copilot era is over." Received UK Government backing in April 2026 as part of a £675M sovereign AI fund. The company has a news presence including coverage in Wired and Tech.eu.

Philosophy

From homepage: "Hand off complex coding tasks without sacrificing maintainability or visibility."

Cosine positions itself against the "copilot era" — interactive line-by-line AI completion — and toward autonomous task delegation. The Lumen model family is explicitly described as "production-first" with three training goals:

1. Eliminating Slop: "Trained to avoid duplication, dead code, and unnecessary complexity. Optimized for minimal diffs that match your codebase's style."
2. Beyond Mainstream Languages: "Post-trained for enterprise and niche languages including C, R, Matlab, Fortran, Verilog, and Rust."
3. Perfecting Vibe: "Behavioral RL on scope discipline, honesty, evidence, and plan calibration. Less filler, more clarity."

Model Benchmarks

Cosine publishes three custom benchmarks:

• Niche-Bench: Performance on rare/enterprise languages
• Vibe-Bench: Behavioral quality (scope, honesty, calibration)
• Slop-Bench: Code quality / anti-slop metric
• Cost per successful task: Lumen Outpost shows competitive cost efficiency

Lumen Outpost benchmark results (from homepage):

• Niche-Bench: 53.9% (vs GPT-5.4: 44.9%, Gemini 3.1 Pro: 42.4%)
• Vibe-Bench: 31.9% (vs GPT-5.5: 29.4%, GPT-5.4: 27.9%)

Enterprise Deployment Options

1. Public cloud (standard): Fully managed, fastest onboarding
2. Dedicated tenant (managed single-tenant): Private environment, enterprise controls
3. Air-gapped (fully isolated): Runs inside customer's security perimeter, no external dependencies

Cosine — Architecture

Distribution

• Type: SaaS + desktop app + CLI (proprietary)
• Models: Lumen family (proprietary, first-party)
• Closed-source — no public GitHub repository

Surfaces (3)

1. Cosine Desktop — native desktop app (macOS, Windows, Linux)

   • Multi-window workflows
   • Visible agent state
   • Reviewable changes
   • Bridge to cloud execution

2. Cosine Cloud — cloud execution backend

   • Run work in parallel
   • Long-running tasks
   • Shared workspace for teams (engineers, PMs, stakeholders)
   • Plan/review/ship interface

3. Cosine CLI — cos binary

   • Install: brew install CosineAI/tap/cos (macOS)
   • Also: Linux and Windows installers
   • Terminal-native workflow
   • Local-to-remote execution handoff
   • MCP tool access from terminal
   • Multi-agent orchestration from CLI

Enterprise Architecture

• Public cloud: Standard managed service
• Dedicated tenant: Private cloud environment with enterprise controls
• Air-gapped: Customer's own infrastructure, no external dependencies, option to fine-tune/post-train on internal codebases

Lumen Model Family

• "Production-first coding models built for real engineering work"
• 8-step data pipeline generating verifiable training trajectories
• Behavioral RL training
• "Lumen Outpost" is the model variant benchmarked publicly

Config Files

Unknown — closed-source, no public documentation of file conventions

Cosine — Components

Note on Limited Public Documentation

Cosine is closed-source with no public SDK or API documentation (as of analysis date). The following is derived from the public marketing site.

CLI Tool (cos)

• Binary: cos
• Install: brew install CosineAI/tap/cos
• Capabilities (from marketing): local-to-remote execution, MCP tool access, project context, multi-agent orchestration
• Subcommands: unknown

Desktop Application

• Available for macOS (download from cosine.sh), Windows, Linux
• Features (from marketing):
  • Multi-window workflows for active engineering
  • Visible agent state (transparency during execution)
  • Reviewable changes (diff view)
  • Bridge to cloud execution (hand off local work to cloud)

Cloud Platform

• Session types: Parallel sessions for independent tasks
• Collaboration: Shared workspace for engineers, PMs, stakeholders
• Long-running tasks: Cloud keeps going when local session ends

Lumen Models

• Lumen Outpost — the benchmarked variant (may be the production model name)
• Training: 8-step data pipeline, behavioral RL
• Specializations: niche languages (C, R, Matlab, Fortran, Verilog, Rust)

MCP Integration

• CLI supports MCP tool access (mentioned in marketing copy)
• Unknown which MCP servers are bundled vs. user-provided

VS Code Extension

• A VS Code extension is mentioned in the marketing copy under "Work where you already are"
• No public extension listing found

Cosine — Uniqueness

Differs From Seeds

Closest to kiro (both closed-source, commercial IDE/agent products), but Cosine's distinguishing architectural bet is having first-party proprietary models (Lumen family) rather than routing to OpenAI/Anthropic/Google. No seed framework builds its own models. The "production-first" training philosophy (minimal diffs, slop elimination, niche language post-training) is a model-level differentiation that no framework in the corpus — open or closed — attempts. The three-surface strategy (CLI + Desktop + Cloud with local→cloud handoff) is similar to Factory Droid but Cosine adds the proprietary model layer.

Distinctive Opinion

The copilot era is over. Production-grade AI coding requires models trained specifically on real engineering outcomes (maintainability, niche languages, scope discipline) rather than benchmark performance on mainstream tasks.

Positioning

• UK sovereign AI strategy backing — unique government relationship in the corpus
• Primary competitors: Devin, Factory Droid (cloud coding agents), Cursor (IDE)
• Key differentiator vs. all others: first-party Lumen models trained on production-code trajectories

Observable Failure Modes (implied from positioning)

• Closed-source model dependency — customers can't audit model behavior
• "Air-gapped" deployment requires Cosine to maintain separate model infrastructure per customer
• Niche language benchmarks may not generalize to all enterprise codebases

Cross-References

• Factory Droid — similar three-surface strategy (CLI + desktop app + cloud)
• Devin — similar cloud execution model for long-running tasks
• UK Government sovereign AI strategy — analogous government backing to US DoD/DARPA programs