Cosine
Also known as: Cosine AI, Genie, Cosine Genie, Lumen, Lumen Scout, Lumen Outpost, Lumen Sovereign, Cosine CLI, Cosine Cloud, Red Team
UK sovereign AI lab whose Lumen coding models power a terminal CLI and cloud workspace, with Swarm mode multi-agent orchestration, a menu of more than 25 models, cross-session memory, and fully air-gapped deployment for regulated industries.
Cosine describes itself as The Sovereign AI Lab, which trains AI models and specialized coding agents for organizations that need frontier capability inside secure environments. Founded in 2022 and backed by Y Combinator, it is now based in London and positions around regulated and sovereign deployment instead of the open benchmark race it first became known for. Its Genie agent drew wide attention in 2024 for topping SWE-bench at roughly 30 percent, the highest by any company at the time. The Genie branding has since been retired, and the company now leads with the Lumen model family.
Lumen is a family of coding models, not a single one. Scout is post trained from Devstral 123B and runs cheaply and on device for the work around the coding loop. Outpost is post trained from Kimi K2.6 and handles everyday production implementation, trained with behavioral reinforcement learning against the failure modes developers complain about: messy patches, duplicated logic, unnecessary abstraction, and agents that claim a task is finished before it is. Sovereign, aimed at reasoning at frontier scale, is announced but not yet shipped.
It is targeted for late 2026 and will be trained on UK soil using the Isambard-AI supercomputer under the UK Government's sovereign AI program. The models are proprietary and post trained for enterprise and niche languages including COBOL, Fortran, Verilog, Rust and complex SQL. Cosine evaluates them on three internal benchmarks of its own design: Niche-Bench for robustness in legacy languages, Slop-Bench for maintainability and Vibe-Bench for collaboration.
The agent reaches developers through two surfaces. The CLI installs via Homebrew and runs in the terminal with a local to remote execution model, moving between planning, implementation and review modes. Every agent turn is a lightweight git commit, so any change can be undone instantly. A live todo list tracks complex jobs, Memory persists project conventions and architecture decisions across sessions, and language servers supply symbol level understanding for references, definitions, renames and diagnostics.
Swarm mode spawns specialized child agents that break a task down and execute in parallel, and browser tools drive a real Chrome, navigating, clicking and filling forms, with each action approved by default. Cosine Cloud runs multiple tasks at once in a shared project across engineers, product managers and stakeholders, and remote agents keep the terminal local while long running work executes in the cloud. MCP connections install automatically with no configuration, covering GitHub, Jira, Postgres, Figma, Stripe, Slack and Linear.
The company is explicit about not locking customers into its own models. The CLI ships a menu of more than twenty five models from providers including OpenAI, Anthropic, Google, Kimi and Qwen alongside the Lumen models, with credit multipliers per model ranging from 0.1x for Lumen Scout to about 5.5x for Fable 5.
Deployment is the commercial center of gravity. Cosine offers fully managed public cloud, a managed single tenant environment with enterprise controls, and fully air gapped installation inside the customer's own perimeter, with Scout available on device. That posture underpins an industry coalition co designing Lumen Sovereign whose members include BAE Systems, Babcock, BT, HSBC, Lloyds, LSEG, NatWest, PwC, Thales UK, Leonardo UK, QinetiQ, Fujitsu, Deloitte, Vodafone, EY and The Alan Turing Institute.
Pricing is credit based across three published tiers: Starter at $19 a month with 4M credits, Team at $199 with 47M, and Enterprise at $999 with 240M, with add on credits priced per million and enterprise or private deployment scoped through sales. Credits are consumed by agent work, model calls and cloud execution, so cost varies with task size and model choice. Cosine's security FAQ states SSO, role based access and audit logging, but the company only says it aligns with SOC 2 and ISO 27001 and publishes no attestation or trust center. There is no public API or SDK for embedding Cosine, though the CLI has a one shot mode for CI pipelines.
Vendor details
Canonical URL
https://cosine.sh
Category
Coding agent
Subcategory
Sovereign coding agent with proprietary models and air-gapped deployment
Funding status
Independent. Founded 2022, Y Combinator backed, and now describing itself as the UK's sovereign AI research lab, based in London. Its Genie agent topped SWE-bench in 2024 at roughly 30 percent, the highest by any company at the time. Cosine is backed by the UK Government's sovereign AI program and leads an industry coalition co-designing Lumen Sovereign, with MoUs reported from BAE Systems, Babcock International, BT, HSBC, Lloyds Banking Group, LSEG, NatWest, PwC, Thales UK, Leonardo UK, Telefonica Tech UK&I, QinetiQ, Fujitsu, Deloitte, Vodafone, EY and The Alan Turing Institute. Lumen Sovereign is to be trained on UK soil using the Isambard-AI supercomputer.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
MCP connections are auto-installed with zero configuration, covering GitHub, Jira, Postgres, Figma, Stripe, Slack and Linear across source control, ticketing, database, design, payments and chat. Language servers give the agent symbol-level understanding for references, definitions, renames and diagnostics. The CLI installs via a Homebrew tap and works alongside the editor, tests and package manager with a local-to-remote execution model; Cosine Cloud runs tasks in parallel in a shared project. The model menu spans more than twenty five models from OpenAI, Anthropic, Google, Kimi, Qwen and Cosine's own Lumen family, with credit multipliers per model.
In practice
You have a backlog of well scoped bug tickets. A developer hands them to Cosine from the CLI or Cosine Cloud, and the agent plans, writes and tests each fix, with every turn a git commit the team can review or revert.
Your code cannot leave your perimeter for compliance reasons. Cosine offers air gapped deployment and on device Lumen models, so an autonomous coding agent can run entirely inside your own security boundary.
A new engineer needs to understand a sprawling legacy codebase. Cosine gathers context on demand, loading the relevant files and using language server lookups such as go to definition and find references, to answer questions about how files and functions connect.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
10.0 / 14 capabilities · 71%
| Integrations & Tool Calling | Full |
|---|---|
|
Auto installed MCP connections cover GitHub, Jira, Postgres, Figma, Stripe, Slack and Linear, language servers supply references, definitions, renames and diagnostics, and the agent runs terminal commands in the developer's shell. Sourcecosine.sh/cli and cosine.sh/docs/customizing/modesread 2026-09-29 |
|
| Workflow Orchestration | Full |
|
Swarm mode makes the primary agent an orchestrator that delegates work to subagents running in parallel, and Cosine Cloud runs long tasks remotely while the terminal stays local. Sourcecosine.sh/docs/customizing/modes and cosine.sh/cliread 2026-09-29 |
|
| Knowledge Grounding & RAG | Partial |
|
Context is gathered on demand, loading relevant files and using language server operations such as go to definition and find references, with MCP connections for outside systems. There is no persistent index or embeddings layer over the codebase. Sourcecosine.sh/docs/concepts/memory-and-context and cosine.sh/cliread 2026-09-29 |
|
| Human Oversight & Guardrails | Full |
|
Manual mode, the default, asks for confirmation before every action that changes something (edits, file operations, terminal commands and MCP tool calls), and Plan mode is read only until the user chooses how the plan is carried out. Auto mode is opt in, and every turn is a git commit that can be reverted. Sourcecosine.sh/docs/customizing/modesread 2026-09-29 |
|
| Security, Identity & Governance | Partial |
|
The security FAQ states SSO, role based access and audit logging. Cosine says only that it aligns with SOC 2 and ISO 27001, and it publishes no attestation or trust center. Sourcecosine.sh security FAQread 2026-10-25 |
|
| Observability & Auditability | Partial |
|
A live task list shows what the agent intends and has finished, and every turn is recorded as a git commit in the customer's repository. Cosine keeps no run history, trace of tool calls or audit log on its side. Sourcecosine.sh/cliread 2026-09-29 |
|
| Memory & State Persistence | Partial |
|
The agent saves reusable facts with a save_memory tool into .cosine/agents.md, a file scoped to the project that persists across sessions and loads at the start of each, beside the team's own AGENTS.md. No lifetime, expiry or purge path is published. Sourcecosine.sh/docs/customizing/memory and concepts/memory-and-contextread 2026-09-29 |
|
| Deployment & Data Residency | Full |
|
Customers choose among fully managed public cloud, a managed single tenant cloud with enterprise controls, and a fully air gapped deployment inside the organization's own security perimeter, and Lumen Scout is sized to run on a device. Sourcecosine.shread 2026-09-29 |
|
| Prebuilt Agents, Templates & Packs | Partial |
|
The orchestrator picks Swarm mode's specialized subagents, not the customer, so they are Cosine's own machinery. There is no catalog of prebuilt agents or templates a customer adopts. Sourcecosine.sh/docs/customizing/modes and cosine.sh/cliread 2026-09-29 |
|
| Triggers & Channel Coverage | Partial |
|
Work starts from the three surfaces, the terminal CLI, Cosine Cloud and Desktop, each started by a person. No schedule, webhook or ticket event wakes an agent on its own. Sourcecosine.sh/docs/ways-to-use-cosineread 2026-09-29 |
|
| Model Flexibility & Routing | Full |
|
The CLI offers a model menu of more than 25 models across providers including OpenAI, Anthropic, Google, Kimi and Qwen alongside Cosine's own Lumen models, described as model agnostic by design, with the user choosing the model. Sourcecosine.sh/cliread 2026-09-29 |
|
| APIs, SDKs & MCP Extensibility | Partial |
|
The CLI has a one shot automation mode (--prompt or --prompt-file, with a JSON schema response format) documented for CI pipelines, but no published API, SDK or Cosine MCP server lets an outside caller drive the platform. MCP connections let Cosine reach other systems, which runs the other way. Sourcecosine.sh/docs/cli/config/flagsread 2026-10-25 |
|
| Testing, Debugging & Optimization | Partial |
|
Every agent turn is a git commit that can be reverted, and Cosine publishes its own Niche-Bench, Vibe-Bench and Slop-Bench results for its Lumen models. These measure Cosine's models and let a change be undone, and there is no harness a customer runs against agent behavior on its own work. Sourcecosine.sh and cosine.sh/cliread 2026-09-29 |
|
| Browser & Computer Use | Full |
|
Browser tools let the agent navigate, click, fill forms and run JavaScript in a real Chrome over the DevTools protocol, with every action gated by approval by default. CLI 2.0.29, released 3 September 2026, turned the browser and computer use tools on by default. Sourcecosine.sh/docs/customizing/browser and cosine.sh/docs/cli/changelogread 2026-10-25 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Computer use tasks in Cosine Desktop 0.0.26 now run in the user's own browser windows by default, with each task kept isolated from the others. Browser actions show up as readable summaries instead of raw automation code and output. The release also adds OAuth sign in for connectors, starting with Linear.
Bears on: Browser/computer use
View sourcePricing
Starter $19/mo (4M credits)
Flat monthly subscription with a bundled credit pool per tier, not per seat. Credits are consumed across agent work, model calls and cloud execution, so burn varies with task size, model choice and runtime. Credit multipliers per model range from about 0.1x (Lumen Scout, DeepSeek v4 Flash) to about 5.5x (Fable 5), with GPT 5.5 and Claude Opus at 2.75x. Add on credits can be bought at any time, and enterprise and private deployment pricing is scoped with sales.
Included quota
Credit-based across three published tiers, not seat-based. Starter $19/mo: 4M credits/month, add-on credits $6.50 per 1M. Team $199/mo: 47M credits/month, add-on credits $5.00 per 1M. Enterprise $999/mo: 240M credits/month, add-on credits $4.50 per 1M. Credits represent usage across agent work, model calls and cloud execution, with actual consumption depending on task size, model choice and runtime. All tiers can buy top-ups at any time. Enterprise, private tenant and air-gapped deployments are scoped through sales.
What is public
All three tier prices and their credit allocations are published exactly, along with per-million add-on credit rates: Starter $19 with 4M credits, Team $199 with 47M, Enterprise $999 with 240M. Per-model credit multipliers are published on the CLI page. Not public: credit cost per task, and pricing for enterprise, single-tenant or air-gapped deployment, which is scoped with sales.
Billing mechanics
Monthly subscription tiers, each with a credit pool that does not roll over. Credits are spent per agent action, model call and cloud execution, with per-model multipliers, and add-on credits are priced per million.
Cost watchouts
Running out of credits pauses work, and top ups add cost. Credits do not roll over, and Team at $199 plus top ups scales quickly. Premium models burn credits several times faster, and the bring your own path shifts cost to the customer's own model subscription.
Variable cost rationale
Cost combines a flat monthly tier with a credit pool consumed by every read, plan and write, and included credits do not roll over, so heavy or exploratory use can exhaust the pool and require top-ups. The jump from Starter (4M credits) to Team (47M) is large, and model multipliers up to about 5.5x speed consumption. The bring your own subscription path shifts model cost to the customer's own Claude, OpenAI or Copilot plan.
Additional watchouts
Credits do not roll over and can pause work when exhausted (top-ups needed). Team at $199 a month plus top-ups can be costly for active teams. No free tier. Enterprise and air-gapped pricing is opaque. Proprietary (not open source).
Overage / add-ons
Credits are consumed per agent action (read/plan/write); when the pool is exhausted, inference pauses until a top-up or the next cycle. Top-ups purchasable anytime. Unused included credits do not roll over.
Sales call required
No, self serve available
Free / trial
No free tier published; entry is the $19/month Starter plan
Lowest paid plan
Starter $19/month, including 4M credits
Commercial notes
Task-outcome oriented: credits track work done. The step from Starter (4M credits) to Team (47M) reflects heavier autonomous workloads. Enterprise value is in air-gapped and sovereign deployment and Lumen models for regulated industries.
Key ambiguities
Credit cost per task is not published and depends on task size, model choice and runtime, though the multipliers per model are. The docs billing page lists six tiers with different names and prices (Starter $19, Builder $49, Team $99, Scale $199, Professional $999, Enterprise custom), while the pricing page lists three (Starter $19, Team $199, Enterprise $999). The prices quoted here follow the pricing page. Enterprise, single tenant and air gapped deployment pricing is quoted rather than listed.
Cancellation / refund
Monthly seat-based subscription; included credits do not roll over. Specific cancellation/refund terms not detailed.
Support SLA / resale
Enterprise agreements for dedicated tenant and air-gapped deployments; standard support on Hobby/Professional. UK Sovereign AI participant. No reseller/whitelabel program surfaced.
Missing data
Exact credit consumption per task type and enterprise/air-gapped pricing are not public; no free tier surfaced.
Related vendors
- Cognition — Maker of Devin, an autonomous AI software engineer
- 10Web — Agentic website platform whose AI agents build, host and manage…
- AgentUI — Managed AI app builder for operations teams that generates and hosts…
- Aider — Open source, model agnostic terminal coding agent that edits your…
- Amp — Amp is a coding agent from Amp Frontier Corporation that runs each…
- Anthropic Claude Code — Anthropic's agentic coding system across terminal, desktop, IDE, web…
Alternatives to Cosine
The closest documented capability profiles to Cosine among coding agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Tabby11.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and APIs, SDKs & MCP Extensibility
- Playcode10.5 / 14Fuller documented coverage on APIs, SDKs & MCP Extensibility and Testing, Debugging & Optimization
- Cline12.0 / 14Fuller documented coverage on Observability & Auditability and Prebuilt Agents, Templates & Packs
- CodeGPT10.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and APIs, SDKs & MCP Extensibility
- JetBrains AI12.0 / 14Fuller documented coverage on Observability & Auditability and Prebuilt Agents, Templates & Packs
- v0 by Vercel11.0 / 14Fuller documented coverage on Security, Identity & Governance and Prebuilt Agents, Templates & Packs
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded