AI Agent Index
ByHeather MacAvelia·Last verified Jun 6, 2026
Cosine logo

Cosine

4.1/ 5

by Cosine

MCP✓ Verified Review
Visit cosine.sh

AI coding platform built on Lumen production-first models with UK Sovereign AI programme backing. Hobby $20/seat/mo; Professional $200/seat/mo; Enterprise custom. Air-gapped deployment.

From

$20

subscription

GitHub

Stars

G2

View on G2 ↗

MCP

⚡ Yes

Compatible

Cosine is the AI coding platform built around Lumen, its family of production-first coding models designed for real engineering work: maintainable software, long-lived systems, and the complex codebases most models struggle to handle. Founded in 2022 by Alistair Pullen and Yang Li (who previously scaled Mobike to 220 million users), Cosine is a Y Combinator graduate that has raised $8 million from Lakestar, SOMA Capital, and Gaingels. The UK government selected Cosine as one of the first partners in its newly launched 500 million pound Sovereign AI programme, and the company is already deployed across UK defence primes and critical national infrastructure operators including organisations involved in the UK nuclear deterrent programmes. Lumen models are trained to eliminate slop: duplication, dead code, and unnecessary complexity. They are optimized for minimal diffs that match codebase style and stay maintainable over time. Lumen is post-trained for enterprise and niche languages including C, R, Matlab, Fortran, Verilog, and Rust, using an 8-step data pipeline that turns real production code into verifiable training trajectories. The current model lineup includes Lumen Scout (lightweight), Lumen Outpost (frontier), and Lumen Sovereign (UK sovereign deployment). On Niche-Bench, Lumen Outpost scored 59.3%, outperforming GPT-5.4 (42.4%), Gemini 3.1 Pro (44.9%), GPT-5.5 (47.4%), and Kimi K2.6 (48.3%). Integration depth spans multiple product surfaces: native desktop apps for Mac, Windows, and Linux, a CLI distributed via Homebrew, and cloud-based execution. The platform integrates with Git, GitHub, GitLab, Slack, Jira, and Linear. MCP support enables external tool connections. Cosine Swarm enables parallel AI agents for long-horizon, complex software tasks. A separate Red Team product provides AI-driven continuous penetration testing. Enterprise deployment options include public cloud, managed single-tenant cloud, and fully air-gapped configurations with custom model weights deployed to customer GPUs, addressing environments where sending code to foreign-managed servers is legally prohibited. Pricing is fully public across three tiers. Hobby at $20/seat/month provides 5M Cosine Credits per seat per month with credit top-ups at $20 for 5M additional credits and full access to all platform features. Professional at $200/seat/month provides 60M Cosine Credits per seat per month. Enterprise is custom-priced with cloud, VPC, or fully air-gapped deployment, custom model weights, zero data egress, and dedicated support. A training layer for turning private engineering knowledge into specialized models is listed as coming soon. Cosine is not the right fit for teams needing the broadest autonomous agent capabilities at the lowest price point. Cursor at $20/month and Claude Code at $20/month both provide broader feature sets for general-purpose coding. The Professional tier at $200/seat/month is significantly more expensive than mainstream alternatives, with the premium justified by production code quality, niche language support, and sovereign deployment capabilities. G2 reviews are not yet available, limiting independent peer sentiment data for procurement teams that rely on review platforms. Current state Q2 2026: Cosine was selected for the UK government's 500 million pound Sovereign AI programme. The Lumen model family expanded to three tiers: Scout, Outpost, and Sovereign. Cosine Swarm launched for parallel agent execution on complex tasks. Red Team launched as a separate product for AI-driven continuous penetration testing. Lumen Outpost leads Niche-Bench at 59.3%. The company has 32 employees and operates from London with deployment across UK defence and critical national infrastructure.

Pricing

subscription · $20

Segment

b2b

Setup

easy

Verified

Jun 6, 2026

Transparency

Public

Contract

Month-to-month

Data training

Not Disclosed

Autonomy

Human Optional

Capabilities

agentic-codingmulti-file-editingcode-generationautonomousterminal-agent

Pros & Limitations

Editorial assessment

Pros

  • Lumen models purpose-built for production code quality with benchmark-leading niche language support: trained to eliminate duplication, dead code, and unnecessary complexity with post-training for C, R, Matlab, Fortran, Verilog, and Rust. Lumen Outpost scores 59.3% on Niche-Bench, outperforming GPT-5.5 and Gemini 3.1 Pro.
  • UK Sovereign AI programme partner with air-gapped deployment: selected by the UK government for its 500 million pound programme, already deployed across UK defence primes and nuclear deterrent programmes. Fully air-gapped deployment with zero data egress for environments where foreign-managed servers are prohibited.
  • Fully public self-serve pricing with multiple deployment options: Hobby at $20/seat/month and Professional at $200/seat/month are transparent and immediately purchasable, with Enterprise adding VPC, air-gapped, and custom model weight deployment for regulated industries.

Limitations

  • Professional tier at $200/seat/month is significantly more expensive than Cursor ($20/month) or Claude Code ($20/month): the premium is justified by production code quality and sovereign deployment, but requires validation before committing at scale for general-purpose coding.
  • Early-stage community and review presence: 6 Product Hunt reviews at 4.7/5 and no G2 reviews limit independent peer sentiment data for procurement teams that rely on review platform evidence before committing to a coding tool.
  • Small team and early funding relative to competitors: 32 employees and $8M raised versus Cursor ($2B+ revenue), Augment Code ($252M raised), or Cognition ($1B+ raised), creating questions about long-term product velocity and support capacity for enterprise customers.

Technical Details

Avg setup timeUnder 30 minutes (download desktop app for Mac/Windows/Linux or install CLI via Homebrew, sign in, first task)
Autonomous rateConfigurable: Cosine agents execute engineering tasks end-to-end without human supervision, including branching, implementing, testing, and submitting pull requests. Cosine Swarm enables parallel agents for long-horizon complex tasks. Developers review and merge completed PRs.
MCP compatibleYes
Integrations
GitHubGitLabSlackJiraLinearMCPMacWindowsLinuxCLI

Similar agents

Rating

4.1/ 5

Editorial score

How we score this →

Score breakdown

AutCap 4 · IntDepth 4 · PriceTrans 5 · IndEvid 4 · SetupAcc 4 = 4.05

Industries

B2BEnterpriseSaaSDevTools

Leave a review

Never displayed publicly.

Agent Stacks

See workflow stacks that feature Cosine.

Compare

Related Content

Is this your tool?

Claim this listing to update your details and get a Verified badge.

Claim listing →