Cosine
Also known as: Cosine Genie, Genie, Lumen
Autonomous AI software engineer (Genie) that takes a ticket from GitHub, Jira, or Linear and ships tested pull requests, powered by Cosine's own Lumen models with air gapped options.
Cosine is a fully agentic AI software engineering company whose product, Genie, works like a member of the team rather than an autocomplete. You assign Genie a ticket from GitHub, Jira, or Linear and it plans the change, edits the code, runs the tests, and opens a pull request for review, working either fully autonomously or paired with a developer in the loop, and it can run several tasks in parallel. Founded in 2022 and backed by Y Combinator, Cosine drew wide attention in 2024 when Genie posted the highest SWE-bench score of any company at the time, and it has since been named in the UK government's Sovereign AI initiative with design partners reported among regulated firms like HSBC, NatWest, and BAE Systems.
Genie's edge comes from two things. The first is deep codebase understanding: rather than being a plain model wrapper, Cosine indexes a repository on multiple levels, combining a graph of relationships between files and functions with semantic search and static analysis, so it can answer natural language questions about the code and ground its changes in how the project actually fits together. The second is training: Genie was trained on data that models how real engineers reason through problems, not just on finished diffs. It works through a terminal native CLI, a VS Code extension, and Cosine Cloud, with Slack in the loop and external tools reachable over MCP.
Cosine also builds its own model family, Lumen, tuned exclusively for production software and post trained for enterprise and niche languages including COBOL, Fortran, Verilog, Rust, and complex SQL. Lumen spans variants for different layers of the workflow, including an on device model called Scout and sovereign, air gapped versions for regulated industries, and Genie can alternatively run on your own OpenAI, Copilot, or Claude subscription. Deployment scales from managed Cosine Cloud to a dedicated private tenant to a fully air gapped install inside your own security perimeter. Cosine states it does not train on customer code and holds no training use agreements with its model vendors.
Pricing is seat based with a monthly Cosine Credit pool per seat. The Hobby plan is twenty dollars per seat a month with five million credits, and the Professional plan is two hundred dollars per seat a month with sixty million credits, where credits are spent whenever the agent reads, plans, or writes code. Additional seats add both users and credits to the team pool, included credits do not roll over, and top ups are available mid month. Notably, when Genie runs on your own Claude, OpenAI, or Copilot subscription through the CLI, that model usage is billed to your subscription rather than to Cosine. Dedicated and air gapped deployments are handled through enterprise agreements.
Vendor details
Canonical URL
https://cosine.sh
Category
Coding agent
Subcategory
Autonomous AI software engineer (SWE agent) with proprietary models
Funding status
Independent. A Y Combinator backed company founded in 2022 (San Francisco and London), Cosine has raised seed funding from investors including Soma Capital. Its Genie agent topped the SWE-bench benchmark in 2024, and it was named in the UK government's Sovereign AI initiative with reported design partners including HSBC, NatWest, and BAE Systems.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
First class integrations with GitHub, Jira, Linear, and Slack for ticket intake and delivery, plus a VS Code extension and a terminal CLI, with external tools reachable over MCP. It acts by editing code, running tests, and opening pull requests, and can run on Cosine's own Lumen models or on your OpenAI, Copilot, or Claude subscription.
In practice
You have a backlog of well scoped bug tickets. Assign them to Genie in Jira or GitHub and it plans, writes, and tests each fix, then opens a pull request for your team to review.
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 indexes it as a graph plus semantic search and answers natural language questions about how files and functions actually connect.
Sources & related URLs
Related / legacy domains
Capability coverage
9.0 / 14 capabilities · 64%
| Integrations & Tool CallingFirst class integrations across GitHub, Jira, Linear, Slack, and VS Code for ticket intake and delivery, external tool access via MCP, and it acts by editing code, running tests, and opening pull requests, broad integration with real action. | Full |
|---|---|
| Workflow OrchestrationOrchestrates the full engineering workflow end to end (ticket intake, planning, code edits, test runs, and pull request) autonomously, with multi agent orchestration and the ability to run multiple coding tasks in parallel across threads, a genuine autonomous orchestration engine. | Full |
| Knowledge Grounding & RAGIndexes and understands the codebase on multiple levels, combining a graph of relationships between files and functions with deep semantic search and static analysis to ground its work and answer natural language questions about the code, a comprehensive headline grounding capability. | Full |
| Human Oversight & GuardrailsDelivers work as a pull request for the developer to review and approve before merge and can run collaboratively with a developer in the loop, a real human oversight surface, though it is a review gate rather than a runtime guardrail enforcement engine. | Partial |
| Security, Identity & GovernanceStrong governance posture: air gapped and sovereign deployment, a stated policy of not training on customer code, and no training use agreements with model vendors, serving regulated design partners, though explicit certifications and an identity matrix like SSO, RBAC, and audit are not enumerated in sources. | Partial |
| Observability & AuditabilitySurfaces what the agent is doing during task execution and delivers a reviewable pull request containing the full change, giving visibility into the agent's work, real observability, though not a dedicated tracing or audit analytics system. | Partial |
| Memory & State PersistenceProjects persist codebase context and a multi level index for the agent, but no distinct runtime agent memory layer that accumulates state across tasks is documented. | Unable to verify |
| Deployment & Data ResidencyOffers a full deployment spectrum from managed Cosine Cloud to a dedicated private tenant to fully air gapped deployment inside the customer's own security perimeter, plus on device and sovereign air gapped Lumen model variants, a comprehensive self host and data residency capability. | Full |
| Prebuilt Agents, Templates & PacksGenie and the Lumen model variants are the core product; no library of prebuilt agents, templates, or workflow packs shipped to users is documented. | Unable to verify |
| Triggers & Channel CoverageBroad channel coverage across a terminal CLI, a VS Code extension, and Cosine Cloud, with tickets from GitHub, Jira, and Linear triggering the agent and Slack in the loop, plus background asynchronous execution, strong channel and trigger coverage. | Full |
| Model Flexibility & RoutingOffers strong model flexibility: its own Lumen model family with variants routed across layers of the workflow (including on device Scout and air gapped versions), plus the ability to run on your own OpenAI, Copilot, or Claude subscription through the CLI, genuine multi model flexibility and routing. | Full |
| APIs, SDKs & MCP ExtensibilitySupports the Model Context Protocol for external tool access and provides a cross platform CLI, real extensibility, though a public API or SDK for embedding Cosine into other applications is not clearly documented. | Partial |
| Testing, Debugging & OptimizationWrites and runs tests as part of delivering each pull request, verifying its own work before review, and the team maintains coding benchmarks for its models, real testing, though the user facing capability is test inclusive delivery rather than a dedicated agent evaluation suite for users. | Partial |
| Browser & Computer UseExecutes code and tests through a terminal native, local to remote execution model, edits files, and accesses external tools via MCP, genuine computer use, though not general browser automation. | Partial |
Pricing
Hobby $20/seat/mo (5M credits) · Professional $200/seat/mo (60M credits) · credits per task; top-ups available · BYO Claude/OpenAI/Copilot uses your own subscription · enterprise air-gapped custom
Seat-based subscription with a monthly Cosine Credit pool per seat (credits consumed when the agent reads, plans, or writes code); additional seats add users and credits; top-ups available; enterprise/air-gapped via custom agreements. BYO model subscription path bills model usage to your own provider.
Included quota
Seat-based + credit pool. Hobby $20/seat/mo: 5M Cosine Credits/seat/mo. Professional $200/seat/mo: 60M Cosine Credits/seat/mo. Additional seats add 5M (Hobby) or 60M (Pro) credits to the team pool. Credits spent when Cosine reads code, plans, or writes. Included credits do not roll over; top-ups available mid-month; work pauses when credits exhausted until renewal or top-up. BYO path: run Genie on your own Claude (Pro/Max)/OpenAI/Copilot subscription via CLI, using your subscription quota with no extra Cosine model billing. Own Lumen models (incl on-device Scout + air-gapped). Enterprise: dedicated tenant + air-gapped (custom).
What is public
Hobby ($20/seat) and Professional ($200/seat) prices and their per-seat credit pools (5M / 60M) are public. Enterprise/dedicated/air-gapped pricing is custom.
Billing mechanics
Per-seat subscription + monthly credit pool (no rollover); credits spent per agent action; top-ups available; seats add users + credits. BYO-model path bills model usage to your own provider subscription.
Cost watchouts
Credit pool exhaustion pauses work (top-ups add cost); no rollover; Professional seats + top-ups scale quickly; BYO path shifts cost to your own model subscription.
Variable cost rationale
Cost combines a per-seat fee 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 gap between the Hobby (5M) and Professional (60M) credit allotments is large, and the $200 Professional seat plus top-ups can add up for active teams. The BYO-subscription path shifts model cost to your own Claude/OpenAI/Copilot plan, which can lower or change the exposure.
Additional watchouts
Credits do not roll over and can pause work when exhausted (top-ups needed). Professional at $200/seat plus top-ups can be costly for active teams. No free tier. Enterprise/air-gapped pricing 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 surfaced; entry is the $20/seat Hobby plan. Note: running Genie on your own Claude/OpenAI/Copilot subscription via the CLI incurs no extra Cosine model billing.
Lowest paid plan
Hobby at $20/seat/month: includes a 5M Cosine Credit pool per seat per month; additional seats each add 5M credits to the team pool. Credit top-ups available; included credits do not roll over.
Commercial notes
Task-outcome-oriented (credits track work done). Big jump Hobby->Professional (5M->60M credits) reflects heavier autonomous workloads. The BYO Claude/OpenAI/Copilot path is unusual and can remove Cosine model billing. Enterprise value is in air-gapped/sovereign deployment and Lumen models for regulated industries.
Key ambiguities
Exact credit cost per task and enterprise/air-gapped pricing are not published; real cost depends on usage intensity and whether you use Cosine credits or your own model subscription.
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
- Aider — Open-source, model-agnostic terminal coding agent that works…
- Anthropic Claude Code — Anthropic's terminal and IDE-native coding agent with sub-agent…
- Augment Code — AI coding assistant evolving into an AI-native engineering platform
- Blackbox AI — AI coding assistant for code generation and developer workflows
- Blink — Agentic AI app builder that turns a natural language prompt into a…