Augment Code

The main question is not "is Augment good," but whether its new scale matches your work.

Who it is for. A team with a large multi-repo codebase, where the same context is needed by agents in tasks, reviews, CI and incidents. Cosmos turns a successful manual run into a repeatable cycle with human checkpoints, and a flat $20 or $100 for up to 50 seats keeps the entry predictable for a small engineering organization.

Who it is not for. For an individual developer looking for an assistant next to the cursor. Since July 1, the old plugins have been retained only for active enterprise contracts, while Cosmos requires adopting separate orchestration, cloud context, and variable spending.

Quick router. If you need a personal agent, start with Claude Code and OpenCode. If you are picking a fleet for a team, look at how Cosmos is built below and compare the full cost, not just the $20 or $100 on the storefront.

What vibecoding isStarting from scratch: how to give AI tasks with no coding experience

The product has three layers, and mixing them up means comparing different tools. Context Engine continuously indexes repositories and retrieves the needed relationships for a task. Through MCP, the same context can be provided to Claude Code, Codex, Cursor and another compatible agent - Augment then works not as an executor, but as memory over a large codebase.

Cosmos. Brings experts together in repeatable cycles. Official templates take a task from ticket to PR or launch an initial review; the result passes between stages, while a person approves branches and the final code.

Auggie CLI. Augment's terminal agent with MCP support. The new v2 shell is built on a fork of the open-source Pi project, while the Context Engine stays closed. The vendor promised v2 for September-October, but as of Sep 26, 2026 the stable version is still 0.36.0.

Models. There is no lock-in to a single provider: the list holds more than twenty models from Anthropic, OpenAI, Google and others, topped by Claude Fable 5.1 and GPT-6 Astra. By default, since Jul 29, 2026, Cosmos picks GPT-5.6 Sol, and the model is changed in the expert or session settings. GPT-6 Sol and Claude Opus 5.5, released on September 22, are not on the list as of Sep 26, 2026.

Start with the surface that fits the task, not with an old review of the extension. For teams Cosmos is open from the $20 Standard plan onward, with up to 50 seats and 50 parallel sessions. You connect repositories and integrations, then the team describes its experts and puts human checks between the stages of the cycle.

For the terminal. Install Auggie through npm and sign in to your workspace:

npm install -g @augmentcode/auggie

auggie login

For another agent. If Claude Code or Codex already works for you as an executor, connect only Context Engine MCP. This lets you test the value of the index without moving the whole team to Cosmos.

IDE. This is not a door for a new user. On Jul 1, 2026 Augment announced that the VS Code and JetBrains extensions are supported only under existing enterprise contracts. The Standard and Business plans include Cosmos and Auggie CLI; the extensions are not listed there.

Commands and CLI version verified Sep 26, 2026 against Augment documentation · current version 0.36.0

All recent figures below were published by the vendor itself. They show direction, but without independent replication they do not become a final rating.

Auggie v2. On SWE-bench Pro, at the same task-solving rate, the new shell cost $1.27 per task versus $2.70 for Claude Code - 53% less; average time was 330 seconds versus 409. On Terminal-Bench 2.0, the task-solving rate rose from 70.8% to 74.2%. There is no external replication, and v2 itself is still awaiting release.

Context Engine MCP. Across 900 attempts on 300 Elasticsearch PRs, the vendor reports a 30-80% quality gain for different agents. The range is broad and the test was prepared by an interested party - the figure supports a hypothesis but does not guarantee the same effect on your code.

The $20 and $100 price tags look like a subscription, but in fact they are an entry ticket with usage included. Up to 50 seats the fee does not grow with each member; what gets expensive is a long agent run.

Standard and Business. One amount for the whole team: $20 or $100 of usage a month, up to 50 seats and 50 parallel sessions. By the comparison table the plans' capabilities are identical, the difference is the included usage. After that you top up the shared balance, and both heavy and occasional users spend from it.

How usage is calculated. A model is billed at the provider's public API price, with Augment's 40% service fee on top. Compute carries no such markup, but Cosmos time at $0.19 an hour is also drawn from the pool. That is why there is no “so many tasks a month” formula.

Two agreements. On paid plans, training on customer data is prohibited by the commercial terms. Community terms allow anonymized customer code, usage information, and output to be used for training and improvement. Check the workspace agreement, not just the wording on the security page.

The cloud remains the cloud. By default, data is stored and processed in the US. Certifications validate the process, but do not turn the closed Context Engine into a local tool; your own encryption keys and region selection are Enterprise features.

Humans keep the gate. Cosmos has checkpoints between experts and loops. They need to be designed: another agent's automated review is not a human decision on architecture, data, and money.

Our experience. In 2025, Augment handled a large codebase in the editorial team's day-to-day work. We did not retest this scenario in 2026.

Practical risk. Users praise its context quality on large codebases, but call the engine a cloud "black box" and complain about unpredictable usage on long tasks. These are signals from several threads, not market frequency.

Augment should be compared with Claude Code. The models overlap, but the product tradeoff is the opposite.

Claude Code. This is a personal tool: the agent runs in the terminal, learns the project as the task goes, and is paid for with a Claude subscription or through the API. Before you choose, also see the Amp card: it covers how to run it and what the models and cloud machines cost.

Augment. This is a shared layer: the Context Engine outlives a single session, and Cosmos runs parallel cycles, handing them one organizational context. Up to 50 seats the $20 or $100 fee does not grow with each member, but the usage on top of it stays variable.

You do not have to choose everything at once. Context Engine MCP can be connected toClaude Code. This is an intermediate test: keep your familiar executor and pay Augment only for the difference you chose it for.

Over two years, Augment went from a smart plugin to an agent factory. What matters to the reader is not release speed, but the July agreement change: the old product remains, but its door is closed to new personal users.