Augment Code is an AI software engineering platform whose defining component is the Context Engine, a live semantic index designed to retrieve the relevant code, history, and connected knowledge from large repositories. Developers can use Augment through VS Code and JetBrains extensions, the Auggie terminal agent, Context Engine MCP, code review, and a Slack bot. The current company offer also includes Cosmos for orchestrating agent workflows across the software development lifecycle. It is strongest when repository scale, cross-service context, governance, and shared team workflows matter more than owning a particular editor.
- The Context Engine is the product center. It indexes code and retrieves a small relevant slice for each task.
- Auggie is the terminal agent. It can read, edit, run commands, call integrations, and expose retrieval to other agents through MCP.
- VS Code and JetBrains remain supported. Inline completions ended for most non-enterprise plans in March 2026, while agent workflows remain.
- Slack is a codebase-aware question surface. It uses linked repository context and excludes that context in channels with external members.
- Cosmos is the current enterprise layer. It coordinates coding, review, testing, incident, and automation experts.
- Augment and Continuum solve different layers. Augment sells retrieval and enterprise workflows; Continuum operates several vendor agents in isolated worktrees across devices.
The short definition
Augment Code is a hosted AI coding and agent platform for teams with large, long-lived codebases. Its main technical claim is that coding models fail less often when they receive a carefully retrieved slice of the repository instead of repeatedly searching broad file spans. Augment calls the retrieval layer its Context Engine. The company then puts that engine behind an IDE agent, a terminal agent named Auggie, pull request review, Slack, MCP access, and the Cosmos orchestration platform.
The distinction matters because the name covers several products built at different times. A 2024 description of Augment as a VS Code and JetBrains assistant was accurate then. A 2025 review may focus on Agent, completions, or Remote Agents. The August 2026 home page leads with Cosmos, Context Engine, code review, and organization-scale agent workflows. Search results mix all three eras.
Current and historical Augment surfaces, checked in August 2026.
| Surface | Job | Current status |
|---|---|---|
| Context Engine | Semantic retrieval across code, services, history, and connected sources | Current core capability |
| Auggie CLI | Interactive or headless terminal agent, plus MCP and ACP server modes | Current |
| VS Code and JetBrains | Agent inside an existing IDE | Current |
| Code completions and Next Edit | Inline prediction while typing | Enterprise only after the March 2026 sunset for other plans |
| Augment for Slack | Codebase-aware questions in channels and direct messages | Current |
| Augment Code Review | GitHub pull request analysis using repository context and custom rules | Current |
| Remote Agents | Independent cloud workers started from VS Code | Earlier product; analytics documentation now marks this feature deprecated |
| Intent | Mac workspace for specs, specialists, diffs, terminals, and isolated work | Current public product at intentapp.dev |
| Cosmos | Organization-level orchestration for agent workflows across the SDLC | Current flagship |
How the Context Engine changes an agent task
A coding model has a finite context window. Even a large window cannot carry a mature monorepo, its commit history, every service contract, and the output from a long agent loop. A useful system therefore has to choose what enters the prompt. Augment treats that selection problem as a search product rather than leaving every model turn to rediscover the repository with file listings and text grep.
Index the working codebase
Augment computes code-oriented embeddings and maintains an index that reflects each developer's branch. Its engineering write-up says updates arrive within seconds and each developer receives a personal view, which matters when two branches contain different symbols or implementations.
Interpret the task as a retrieval query
A prompt about payment request logging is related to routes, middleware, services, client code, database records, and webhook handling even when those files do not share the exact prompt words. Semantic retrieval looks for those relationships and ranks the likely evidence.
Send a selected slice to the model
The model receives relevant code spans and supporting material instead of a repository dump. Augment says ranking also considers what is active or deprecated, how services connect, and which part of the workspace is currently active in the IDE.
Refresh as code changes
Edits, branch changes, and new commits alter the index. A retrieval system that lags behind the working tree can confidently return the wrong definition, so freshness is part of correctness rather than a background convenience.
The company publishes concrete engineering numbers for its large-index implementation. On a 100 million line codebase, it reports reducing search memory from 2 GB to 250 MB and typical query latency from more than two seconds to under 200 milliseconds while maintaining 99.9 percent parity with exact search results. Those measurements describe Augment's own system and test conditions. They are useful architectural evidence, not an independent guarantee for every repository.
Augment also exposes the retrieval layer through Model Context Protocol. Auggie can run as a local MCP server against the current workspace, discover workspaces on demand, or connect through Augment's hosted MCP for cross-repository context. That means Claude Code, Codex, Cursor, Gemini CLI, and other MCP clients can use Augment retrieval without adopting Auggie as the coding agent. This modular path is important for teams that like their current agent and only want stronger code search.
What Auggie does
Auggie is Augment's terminal coding agent. In an interactive session it can inspect files, search through the Context Engine, edit the workspace, run shell commands, use native integrations, call MCP tools, and keep working through a multi-step plan. A print mode supports one-shot automation, while JSON output and session controls make it usable from scripts and CI.
| Capability | How it appears |
|---|---|
| Interactive agent | auggie in a repository |
| One-shot task | auggie --print "review the current diff" |
| Model selection | /model or the --model flag |
| Context Engine for another agent | auggie --mcp --mcp-auto-workspace |
| Automation authentication | Stored login, session JSON, or token commands |
| Rules and reusable workflows | Project rules, custom commands, skills, plugins, and subagents |
| Safety control | Granular CLI tool permissions |
| Billing check | auggie account status |
The current model catalogue is multi-provider. Augment documents Claude, GPT, Gemini, Kimi, and two Prism routing groups. A workspace can choose a model, while Prism chooses within a defined family based on the task and operating conditions. Model availability and consumption vary by organization. This makes Auggie different from a vendor-owned agent whose model family and billing are tied to one lab.
Auggie can also define subagents in Markdown files under a user or workspace directory. Each subagent gets its own context window, prompt, model, and optional tool restrictions. Several can run in parallel and return progress summaries to the main thread. That gives Augment a local delegation mechanism alongside Cosmos, which handles larger organization workflows.
IDE, Slack, review, Intent, and Cosmos
The IDE extensions remain the quickest path for developers who already use VS Code, WebStorm, PyCharm, IntelliJ, or another supported JetBrains product. The panel signs in, indexes the workspace, accepts natural-language tasks, shows terminal actions, exposes diffs, and saves checkpoints before changes. Agent Auto can continue through approved categories of action with less interruption.
The product is now agent-first. Augment ended code completions and Next Edit for Indie, Standard, Max, and legacy plans on March 31, 2026. Enterprise customers retained support at that date. Anyone evaluating an older review that praises Augment mainly for inline completion should account for that change. The current center is task execution, retrieval, review, and orchestration.
Augment for Slack answers codebase questions in channels and direct messages. It considers the Slack thread and the linked repository's default branch. Users can select among installed repositories per conversation. It does not expose repository context in channels with external members, and administrators can request a channel allowlist. The default-branch constraint makes Slack useful for shared architecture and support questions, while branch-specific debugging still belongs in the IDE or CLI.
Augment Code Review installs through the GitHub App and reviews pull requests with full-repository retrieval, custom rules, and adaptive feedback based on which findings developers address. Augment's current Cosmos review offer goes further: a risk analyzer, deep reviewer, pair reviewer, and PR author divide the work, with low-risk changes eligible for automatic approval and humans pulled into judgment calls. The performance figures on that page are Augment's measurements on its own codebase.
Intent is a macOS workspace where code, a living spec, tools, terminals, diffs, and specialist agents stay attached to the same task. Its coordinator decomposes work and keeps the agent lanes aligned with a shared goal. Cosmos moves that pattern to the organization: triggers, experts, shared memory, sandboxes, review, testing, incidents, migrations, and ticket-to-PR flows under centralized policy.
Security and data handling
Augment is aimed at customers who cannot treat source code as casual prompt material. Its paid offer says customer code is excluded from model training, carries SOC 2 Type II controls, and supports enterprise features such as SSO, OIDC, SCIM, customer-managed encryption keys, data residency, audit trails, SIEM integration, and ISO 42001 governance. Availability depends on plan and contract, so a procurement review should map each required control to the actual order form.
The unusual mechanism is Proof of Possession. The backend stores a code index, yet it should only return a code span when the requesting client demonstrates that it already has the file content. The client sends a cryptographic hash as evidence before retrieval. The design reduces the risk that a guessed identifier or cross-tenant mistake could extract source code from the index. It also lets the IDE use a branch-specific local view while sharing index infrastructure.
| Question | Published answer | What to verify |
|---|---|---|
| Is paid customer code used for training? | Current paid pricing and security pages say no | Contract language and the exact plan in use |
| Where does inference run? | Augment uses several model providers and its own routing | Region, retention, subprocessors, and BYOK terms |
| Can code leave the device? | Indexing and model requests process code through hosted services | Excluded paths, retention, encryption, and proof-of-possession behavior |
| Can Slack expose a repository? | External-member channels receive no repository context | Guest handling, channel allowlists, and audit logs |
| Can agents run commands? | Yes, subject to the surface and permission mode | IDE, CLI, Intent, and Cosmos controls separately |
| Can the product run in a customer perimeter? | The enterprise site advertises VPC, single-tenant, BYOK, and on-prem options | Architecture, service dependencies, support, and contract scope |
Community terms deserve separate attention because they permit use of anonymized community data for model improvement. That language differs from the paid-plan promise. A team handling proprietary code should confirm that every developer is enrolled under the commercial agreement and that trial or community use does not bypass the intended policy.
Who should choose it
| Choose Augment when | Choose a different route when |
|---|---|
| The hard problem is finding relevant code across a large monorepo or several services | The repository is small and ordinary search already gives the agent reliable context |
| The team wants one context layer usable from several MCP-compatible agents | A single vendor agent already meets the quality and governance bar |
| Code review, automation, audit, and shared expert workflows are part of the purchase | One developer wants the cheapest personal terminal agent |
| Existing VS Code or JetBrains adoption should remain in place | The desired product is an AI-first editor with deep predictive typing |
| Procurement needs enterprise identity, security, and deployment controls | A self-hosted or fully open-source stack is mandatory |
| The team can measure retrieval quality and agent outcomes on its own repositories | The decision must be made from public benchmarks alone |
A useful trial starts with code that defeats generic search: a cross-service bug, an internal API migration, an old feature with sparse documentation, and a pull request whose defect sits outside the diff. Run the same task with and without Context Engine MCP while holding the agent and model fixed. Record relevant files found, tool calls, tokens, wall time, test pass rate, and reviewer corrections. That isolates the retrieval value from model variance.
Then test operational recovery. Switch branches, edit a symbol that exists in both branches, disconnect an integration, resume a session, revoke a repository, and inspect what Slack can answer. Enterprise tooling earns its price in the failure and administration paths as much as in the first successful prompt.
Define the adoption boundary before the trial
A platform this broad can appear successful while solving a different problem from the one that triggered the evaluation. Write down which component is in scope: Context Engine retrieval, Auggie, the IDE extension, Code Review, Intent, or Cosmos. Name the existing tool it may replace, the data sources it may index, and the people who will operate it. A retrieval pilot does not prove that Cosmos should own incident response, and a good review backtest does not prove the IDE should become a team standard.
Set exit criteria for every layer. Context Engine should beat native search on a defined share of difficult tasks. Auggie should complete a defined task set within a cost and correction budget. Code Review should meet precision and recall thresholds on historical pull requests. Cosmos should shorten a measured workflow without weakening a human gate. Clear layer-specific thresholds prevent a strong result in one product from becoming a blanket approval for all of them.
Finally, assign ownership for indexes, connectors, rules, memories, experts, permissions, spend, and vendor changes. These assets become production infrastructure once engineers depend on them. A team that cannot name the owner will discover stale context and overbroad access only after the initial champion has moved on.
Questions people ask
Teams use Augment to retrieve relevant code from large repositories, ask architecture questions, implement and test changes through an agent, review pull requests, expose code search to other agents through MCP, and automate engineering workflows through Cosmos.
Augment has extensions for VS Code and JetBrains rather than requiring a separate editor. Intent is a separate macOS workspace for specs and coordinated agents. Cosmos is the organization orchestration layer, and Auggie is the terminal client.
Auggie is Augment Code's CLI agent. It can search with the Context Engine, edit files, run commands, call integrations, run headlessly, define subagents, and serve codebase retrieval to other MCP-compatible tools.
Yes. The official quickstart supports VS Code and JetBrains products including IntelliJ, WebStorm, and PyCharm. The current non-enterprise focus is Agent rather than inline completions, which were sunset for most individual legacy plans in March 2026.
Yes. Auggie can expose Context Engine retrieval as an MCP server, and Augment documents Claude Code, Cursor, Codex, Gemini CLI, and other MCP clients as supported consumers. The external agent keeps its own model, tools, and billing.
The current paid pricing and security pages state that paid customer data is excluded from AI training. Community terms allow anonymized community data to be used for improvement, so proprietary teams should verify that every user and repository is covered by the commercial agreement.
The product can work on smaller repositories, and the current Business plan covers teams up to 50 seats. Its clearest differentiation and current marketing focus are enterprise codebase retrieval, governance, review, and agent orchestration.
The live August 2026 pricing page lists a $100 monthly Business plan and custom Enterprise pricing. Trials and beta access may exist, while older $20 Indie and $60 Standard articles describe earlier packaging. Check the live pricing page and your billing dashboard before budgeting.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.