
Coding agent for CLI, desktop and cloud, with air-gapped Lumen models.
Cosine suits engineering teams that need an agent for legacy languages or a closed network and want to swap models. It is a poor pick for hobbyists who want autocomplete, and its Lumen Sovereign model is not due before 2027, so judge it on the agent it ships today, not the roadmap.
A coding agent backed by the company's own models and a deployment mode that keeps code, prompts and outputs inside the customer's boundary.
Main trade-off
The flagship sovereign model is not due before 2027 and has no published results.
Credits cover agent work, model calls and cloud execution, so a task's cost depends on its size, the model chosen and the runtime rather than on tokens. Every tier can buy top-up credits at any time. Enterprise and private deployments, including air-gapped ones, are scoped and priced with Cosine's sales team.
Cosine is a London AI lab that sells a coding agent for the command line, desktop and cloud, built on its own Lumen models and a menu of third-party ones. The agent plans, edits and tests code with checkpoints and approval-gated actions, and Cosine offers air-gapped deployment for regulated buyers.
Checkpointing
Every agent turn in the CLI is a lightweight git commit, so any change can be undone instantly.
Swarm subagents
In Swarm mode an orchestrator spawns specialised child agents that work on parts of a task at the same time.
Parallel cloud agents
Twenty-plus agents work at once on isolated worktrees, each in a MicroVM that Cosine says boots in about 125 ms.
Approval-gated actions
Reads run automatically, while writes such as pushing code, sending a Slack message or filing a pull request comment wait for one confirmation.
One model menu
Lumen Outpost and Lumen Scout sit beside OpenAI, Anthropic, Google, Kimi, MiniMax and GLM models, with a cost multiplier shown for each in Desktop.
Legacy-language focus
Its niche-languages page names COBOL, ABAP, Fortran and Verilog, with repository mapping, test scaffolding and documentation for old code.
Air-gapped deployment
Runs where source code, prompts and outputs cannot leave the boundary, with audit visibility into agent actions, tool use and review status.
Cosine Desktop
A Windows app that applies the agent to spreadsheets, vendor research and customer-feedback review for non-engineers.
| Agent | Where it runs | Price | Maturity | Rating |
|---|---|---|---|---|
| Cloud · Your machine | $19/mo | Available | 3.5 | |
| -- | -- | -- | -- | |
| -- | -- | -- | -- | |
| Cloud · Your machine · API | Free plan | Available | -- |
Cosine lists Starter at $19 a month with 4M credits, Team at $199 with 47M credits and Enterprise at $999 with 240M credits. Add-on credits cost $6.50 per 1M on Starter, $5.00 on Team and $4.50 on Enterprise. Enterprise and private deployments are priced with sales (pricing page, read 15 November 2026).
No free plan or trial appears on the pricing page, so the Starter tier is the cheapest way in. Some third-party directories still list an older free allowance, which the current page does not show.
Lumen Sovereign is Cosine's planned frontier model, trained on the Isambard-AI supercomputer with UK government backing, in 1T and 200B parameter variants. Its homepage gives Q1 2027, which is later than the late-2026 target in the 8 June 2026 coalition post, and no results are published, so nothing can be tested yet.
Claude Code is Anthropic's own agent for the terminal, IDEs, desktop, web and mobile, tied to Anthropic's models. Cosine ships a menu that mixes its Lumen models with models from OpenAI, Anthropic, Google and others, and offers air-gapped deployment. Hokai found no independent head-to-head test of the two.
Yes. The CLI installs with winget add Cosine.CLI, and Cosine Desktop has a Windows download aimed at tasks such as spreadsheet reports and vendor research. Cosine Cloud runs each agent session in a MicroVM on Cosine's workers or on your own SSH targets, so heavy work need not run on the laptop.
Cosine is a London AI lab founded in 2022 by Alistair Pullen (CEO) with co-founders Yang Li and Sam Stenner, a Y Combinator graduate that tech.eu said in April 2026 had raised $8 million from Lakestar, SOMA Capital and Gaingels. In April 2026 the UK government chose it as one of the first partners in its £500 million Sovereign AI programme, with 500,000 GPU hours on the Isambard-AI supercomputer, and the FT, relayed by AI Weekly in August, described a team of roughly 30. Its homepage now leads with sovereign models, "We train AI models for organisations that need frontier capability inside secure environments", while its docs describe Cosine as an AI coding agent for software engineering work, from quick fixes to complex migrations.
The agent works across three surfaces. The CLI makes every agent turn a lightweight git commit, so a change can be undone at once, and it keeps a live to-do list, project memory and, in Swarm mode, an orchestrator that spawns child agents. Cosine Cloud runs 20 or more agents in parallel on isolated worktrees, each session in a hardware-isolated MicroVM; reads run on their own and writes wait for a confirmation. Cosine Desktop applies the same agent to spreadsheets and research for non-engineers. Its model menu mixes the company's Lumen Outpost and Lumen Scout with models from OpenAI, Anthropic, Google, Kimi, MiniMax and GLM.
It fits teams that maintain systems in COBOL, ABAP, Fortran or Verilog, which its niche-languages page names as a focus, and regulated buyers that cannot send code out: its air-gapped page promises no data egress, a choice of Lumen, open-weight or customer-trained models, and a record of agent actions. Lumen Sovereign is the planned frontier model for that market, in 1T and 200B parameter variants; a coalition of UK banks, defense and telecom firms signed memoranda to co-design it in June 2026, and the homepage now gives Q1 2027 for the first release, against a late-2026 target in that announcement.
Skip it if you want in-editor autocomplete, for which Cursor is the usual choice, or if you want one vendor's agent across every surface: Claude Code is Anthropic's, Devin is Cognition's hosted engineer and OpenHands is an open-source, model-agnostic option you run yourself. The AI coding agents list shows the rest. Lumen Sovereign has no published results to test, so judge Cosine on the agent it ships today. Pick it only if legacy languages or a closed network decide the purchase; for everything else, start with Claude Code or OpenHands.
Autonomy, guardrails, integrations and pricing on this page are read from Cosine's own documentation. The verdict is Hokai's reading of them, not the vendor's claim.