Coding Agents

Cursor is a coding agent that runs in the terminal, collaborates in Slack and reviews pull requests in GitHub, distributed as a desktop application with a macOS download. Users can choose between models from OpenAI, Anthropic, Gemini, SpaceXAI and Cursor's own — named as written on the page — or bring their own model; tools and MCP servers are invoked with @, and skills and commands with /. Cursor offers a targeted-edit mode and what it calls a full autonomy agentic version, and always-on agents can be set to run on schedules or triggers to build, maintain and fix software. Agents use their own computers to build, test and demo features end to end for the user to review, which is the only oversight point the page describes and one that comes after the work. Cursor says it is trusted by over half of the Fortune 500. Against that deployment claim, the page checked publishes no certification, no regulatory posture, no retained record of agent actions, no pricing at all, no AI liability cover or indemnity, and no outcome-linked payment term.

An autonomous software-engineering agent. On the page checked it ships pull requests the way a team does, picking up review feedback and CI results to get each PR approved and merged, with a human kept in the loop to manage the project and approve its changes. The vendor describes migration and large refactoring as the work it suits, states it can spin up a team of Devins for large tasks, and lists Cloud, Desktop, CLI, Review and Windows VM surfaces alongside integrations with GitHub, Slack, Snowflake, Databricks, Stripe and others. Nubank is the one named customer; the vendor reports that business units there completed migrations in weeks rather than months, and that it observed a doubling of task-completion scores against its own evaluation set after fine-tuning on examples that customer supplied. That evaluation is the vendor's own, not an independent one. No compliance certification, audit-trail mechanism, insurance, indemnity or outcome-linked pricing appears on the page checked.

Maige is an open-source AI copilot that labels, assigns, comments on, and reviews GitHub issues and pull requests, used by a reported 4,300+ repositories.

As of our capture of refact.ai on 1 August 2026, the homepage carries a banner stating that Refact Cloud - the hosted service - is shutting down soon; the open-source, self-hosted build is not said on that page to be affected, and the pricing and hosted-tier facts below should be read with that in mind. Refact.ai is a coding agent that runs inside the developer's IDE, working against their codebase and stack; the vendor says it plans, executes and deploys from a described task, completing work step by step with reasoning the user can preview and control. It is powered by the Qwen2.5-Coder model with retrieval-augmented generation, and customers can use Claude 4, GPT-4o or GPT-4o mini, or connect their own API key to run Gemini, Grok, OpenAI or Deepseek models, and fine-tune on their own stack. Plugins are published for VS Code, JetBrains IDEs (PyCharm, WebStorm, GoLand, IntelliJ, CLion), Visual Studio, Neovim and Sublime Text, several marked beta. The Pro tier starts at $10 per month, and on-premise deployment is offered as the data-control option. The homepage headline calls the product open source but names no licence on the page checked. Beyond that, Refact.ai publishes little a buyer could rely on: no certification, no regulatory posture, no retained record of agent actions, no AI liability cover or indemnity, and no outcome-linked payment term. It scores 0 on four of the five trust-gap dimensions and 1 on the fifth.

Currency caveat: the only capture AI Dispatch holds of usetusk.ai is from 1 August 2026. That page markets the product normally and also lists, among the company's own resources, a post titled "Tusk Is Sunsetting" dated June 10, 2026. AI Dispatch has not verified from a fresh capture whether the product, or which part of it, is still offered, and the fields below should be read as a record of that 1 August capture rather than as current terms.

Agentic software development platform. Agents are triggered by an issue, a chat message or a schedule and coordinated from triage through review to a mergeable pull request, with a screenshot or video captured per run so work can be verified before a PR ships. Runs on Warp's cloud or fully self-hosted in the customer's own VPC, and is model-agnostic: customers bring their own model or harness. Exposed as a platform via API, CLI, SDK and MCP. Pricing is usage-based per agent run. The vendor claims over 800,000 developers and states that most organisations start with 20-30% of pull requests fully automated. Its terminal product is open source. No certification, regulatory alignment, insurance or indemnity is published on the page checked.