Augment Code vs Claude Code: retrieval platform or agent

Claude Code is a mature coding agent built around Anthropic models and a strong terminal harness. Augment adds a proprietary context service, multi-model routing, review, and enterprise orchestration. They can compete, and they can run together.

By the Continuum team. We build a workbench that runs Claude Code, Codex, and their peers, so the model rates quoted here are the ones our own cost analytics ship with.

The short version

Choose Claude Code for a focused Anthropic agent with an excellent terminal loop, mature permissions, checkpoints, CLAUDE.md, skills, hooks, MCP, subagents, agent teams, local and cloud execution, and a $20 individual entry route through Claude Pro. Choose Augment when the hard problem is repository retrieval across services, model flexibility, shared code review, enterprise deployment controls, or organization-wide agent workflows. Augment Business costs $100 monthly for up to 50 seats plus variable usage. The strongest combined setup often uses Claude Code as the agent and Augment Context Engine as an MCP retrieval service.

What you need to know
  • Claude Code is the stronger focused agent product. Its terminal harness and Anthropic ecosystem are deep and cohesive.
  • Augment is the stronger dedicated retrieval layer. Context Engine can serve several agents and repositories.
  • Model strategy differs. Claude Code uses Anthropic models; Augment offers Claude, GPT, Gemini, Kimi, and Prism routing.
  • Claude Code wins individual entry price. Claude Pro is $20 monthly or $200 annually.
  • Augment wins fixed team seat math. Business is $100 monthly for up to 50 seats before overage.
  • Combination is first-class. Augment documents Claude Code as a Context Engine MCP client.

Pick by job

JobPickWhy
One developer delegating repository tasksClaude CodeLower entry price and a focused terminal workflow
Semantic retrieval across several large repositoriesAugmentContext Engine is a dedicated cross-agent service
Use only Anthropic models and featuresClaude CodeDirect access to Claude models and Anthropic-native releases
Switch among several model labsAugmentClaude, GPT, Gemini, Kimi, and Prism options
Hooks and local workflow customizationClaude CodeMature hooks, skills, plugins, subagents, and settings hierarchy
Organization code review and repeatable SDLC workflowsAugmentCosmos experts, shared memory, risk analysis, and automation triggers
Run Claude with stronger semantic searchBothClaude Code can call Augment Context Engine through MCP
Operate Claude beside other vendor agentsContinuumOne workbench, separate worktrees, and cross-device control

The products overlap at the agent layer because Auggie can perform the same basic loop as Claude Code: inspect, edit, execute, observe, and continue. Their underlying bets differ. Claude Code makes the model and harness the center. Augment makes retrieval and organization workflow the center, then allows several models to use it.

How each agent builds context

Claude Code starts from the working directory, git state, CLAUDE.md, auto memory, skill descriptions, MCP tool names, and the user's prompt. During the agent loop it searches filenames and content, reads files, calls code intelligence plugins when installed, runs commands, and carries results in the session context. Tool Search defers MCP schemas until needed, subagents keep verbose investigation outside the main context, and compaction summarizes older history when the window fills.

Augment maintains a semantic index ahead of the task. The Context Engine uses code-oriented embeddings, branch-aware personal indexes, relationship and activity signals, and connected sources to retrieve a ranked slice. Its published design updates within seconds and requires proof of local possession before returning indexed content. That slice enters Auggie, an IDE agent, review, Cosmos, or an external MCP client.

Context mechanismAugment CodeClaude Code
Repository discoveryPre-indexed semantic retrieval plus agent toolsAgent search, file reads, git, and optional code intelligence
Persistent project instructionRules, memories, skills, and integrationsCLAUDE.md, rules, auto memory, skills, and plugins
Cross-repository searchHosted Context Engine and connectorsAdd directories, MCP services, or separate tools; no built-in global semantic index claim
Branch freshnessPersonal real-time index per developer and branchReads the current working tree directly
Long-session controlRetrieved slices, agent memory, subagents, and model contextCompaction, context inspection, deferred MCP tools, skills, and subagents
External reuseMCP, SDK, connectors, and HTTP servicesClaude Code consumes MCP and can expose SDK capabilities

Claude Code's direct reads have a valuable property: the content comes from the current filesystem at the time of the call. Its weakness is exploration cost. A vague task can trigger broad search, several file reads, and repeated context replay. Anthropic's own cost guide recommends specific prompts and code intelligence plugins to reduce that work.

Augment can find semantically related code before the model explores, especially across naming and service boundaries. Its weakness is retrieval error. A ranking system can surface a plausible neighbor while missing a small decisive invariant. The current working tree still has to be read and the result still has to be tested.

The combined path gives Claude Code a codebase-retrieval MCP tool backed by Augment. Claude decides when to call it, receives relevant spans, and continues with its native file, shell, permission, and verification loop. A team keeps Anthropic's agent while testing whether Augment reduces turns and repairs.

Models and reasoning

Claude Code is vertically integrated with Anthropic's models. The current Claude product includes Sonnet, Opus, Haiku, and newer Anthropic releases according to plan and availability. The harness can choose a model, adjust effort, use extended thinking, route simple subagent work to a cheaper model, and switch within the Anthropic family during a session.

Augment supports models from several labs. Its August 2026 documentation lists multiple Claude, GPT, Gemini, and Kimi versions plus two Prism routers. Prism selects within a curated family according to task and operating conditions. The Context Engine, rules, tools, and agent workflow remain stable while the chosen model changes.

QuestionAugmentClaude Code
Model providersAnthropic, OpenAI, Google, Moonshot, and routing groupsAnthropic
Model selectionIDE picker, /model, CLI flag, or organization default/model, settings, model flags, and subagent configuration
Automatic routingPrism routes within a documented familyAnthropic may choose defaults; user generally selects the Claude model or alias
Latest Anthropic accessRollout depends on Augment supportDirect Anthropic product path
Cross-model comparisonPossible inside one Augment accountRequires another agent or external provider for non-Claude models
Billing basisAugment plan usage and provider list price plus service feeSubscription allowance, usage credits, API, or cloud provider

Multi-model support reduces dependence on one lab's quality curve. It also places Augment between the developer and each provider. Model rollout, prompt adaptation, service availability, and accounting depend on Augment. Claude Code gives up provider choice and receives the shortest route to Anthropic-specific capabilities.

Benchmarks should keep model and effort constant when comparing harnesses. Running Auggie on one premium model and Claude Code on a cheaper model produces a purchasing anecdote rather than evidence. Use the same model where both products expose it, set comparable permissions, clear state, and run multiple repetitions because agent outcomes vary.

Tools, permissions, and extensibility

Claude Code has a mature customization stack. CLAUDE.md supplies always-on repository guidance. Skills load reusable knowledge and workflows on demand. Hooks run scripts, prompts, HTTP calls, or subagents on lifecycle events. MCP connects outside services. Plugins package commands, skills, hooks, agents, and servers. Custom subagents get separate contexts and tool policies. Experimental agent teams coordinate independent sessions with a shared task list and direct messages.

Permissions are integrated into the daily loop. Manual asks before edits and commands. Accept Edits permits file changes and common filesystem work. Plan explores without editing. Auto evaluates actions with background safety checks. Checkpoints snapshot file contents before edits and can rewind changes independently of git, though external actions cannot be reversed that way.

Auggie supports project rules, custom commands, skills, plugins, integrations, MCP, tool permissions, and custom subagents. Its subagent files can select a model and restrict tools, and several subagents can run in parallel. Augment also offers native integrations for GitHub, Linear, Jira, Confluence, Notion, Sentry, Stripe, and other systems.

CapabilityAugmentClaude Code
Project instructionsRules and memoriesCLAUDE.md, rules, and auto memory
Reusable workflowsSkills, custom commands, and pluginsSkills, commands, hooks, and plugins
External toolsNative integrations and MCPMCP plus shell CLIs
SpecialistsCustom subagents and Cosmos expertsSubagents and experimental agent teams
File rollbackIDE checkpoints and git workflowAutomatic file checkpoints plus git
Permission scopeGranular CLI permissions; IDE control is separateSettings hierarchy and interactive modes across the Claude Code harness
Headless budget controlsScriptable print mode and organization usage controlsPrint mode, turn limits, budget flags for API use, and structured output

Claude Code has the more cohesive local customization story because the same vendor controls model, loop, permissions, and documentation. Augment has the broader organization story because Cosmos can orchestrate experts, triggers, review, compute, and shared memory beyond one terminal session.

Interfaces, remote execution, and teams

Claude Code runs in the terminal, VS Code, JetBrains, desktop, web, Slack, CI, cloud environments, and Remote Control. Local execution uses the developer's machine. Cloud execution uses Anthropic-managed or organization-hosted environments. Remote Control lets a browser steer a process whose files and commands remain on the local machine. Sessions are tied to directories, and Anthropic recommends git worktrees for parallel write-capable sessions.

Augment runs through VS Code and JetBrains extensions, Auggie CLI, Slack, GitHub review, Context Engine MCP, Intent on macOS, and Cosmos. Intent attaches a spec, code, diffs, terminals, and specialists to a workspace. Cosmos schedules agents and experts across code authoring, review, verification, incidents, migration, and automation with enterprise controls.

Both offer Slack, with different emphasis. Claude Code's Slack integration can delegate coding work. Augment for Slack is documented as a codebase-aware conversation surface using linked default branches, repository selection, external-channel protection, and optional channel allowlists. Confirm execution behavior separately from question answering when a Slack workflow is part of the purchase.

WorkflowAugmentClaude Code
Terminal foregroundAuggieClaude Code CLI
VS CodeExtensionExtension
JetBrainsExtensionExtension
Browser controlling local hostCosmos or product-specific web paths; verify exact deploymentRemote Control documented directly
Cloud delegated taskCosmos; earlier Remote Agents are deprecatedClaude Code on the web and cloud environments
Mac agent workspaceIntentClaude desktop and Code surfaces
Organization orchestrationCosmos experts and triggersEnterprise controls, agent teams, SDK, GitHub Actions, and CI building blocks

Claude Code is simpler to reason about as one agent available in several environments. Augment covers more layers, which can serve a platform team and create product-boundary questions. Ask which session record is canonical, how state migrates among IDE, CLI, Intent, and Cosmos, and which metrics remain supported when a surface is retired.

Security and deployment

Claude Code local mode runs with access to the current project, terminal, git credentials, and configured services. Permission modes determine when actions require approval, while managed settings can enforce organization policy. Cloud tasks run in remote environments. Enterprise customers can use identity, access, analytics, network controls, custom retention, and cloud or provider billing paths documented by Anthropic.

Augment processes repositories for semantic indexing and model inference. Its paid pages promise no training on customer code and describe proof of possession, non-extractable retrieval, SOC 2 Type II, ISO 42001, customer-managed encryption keys, data residency, audit, SIEM, VPC, single-tenant, and on-prem options. Community terms permit anonymized community data use, so commercial enrollment matters.

Security decisionAugment questionClaude Code question
Data pathWhich index, model provider, region, and retention apply?Does work run local, Anthropic cloud, API, or a cloud provider?
Command authorityWhich IDE, CLI, Intent, and Cosmos policies apply?Which permission mode and managed settings apply?
Repository isolationWhich sandbox and workspace boundary applies to each expert?Is each parallel session in a separate worktree or cloud environment?
IdentityHow do SSO, SCIM, service accounts, and GitHub App scopes map?How do Claude org identity, console keys, cloud identity, and MCP OAuth map?
AuditWhich actions and retrieved context are exported?Which local logs, enterprise analytics, API metrics, hooks, and cloud logs exist?
DeletionHow are indexes, memories, sessions, and connected sources removed?How are local JSONL sessions, cloud sessions, provider logs, and organization data removed?

Augment has the stronger published menu for customer-perimeter deployment and context authorization. Claude Code has the clearer local execution story and a detailed permission model. The safer result depends on configuration. A local agent with broad shell and credentials can carry more practical authority than a tightly scoped hosted worker.

Pricing and total cost

Public pricing checked in August 2026.

BuyerAugment CodeClaude Code
Individual$100/mo Business floorClaude Pro $20/mo or $200/yr; Max from $100/mo
Five people$100/mo plus overageTeam Standard $100/mo annual equivalent or $125 monthly
Twenty people$100/mo plus overageTeam Standard $400/mo annual equivalent or $500 monthly
Fifty people$100/mo plus overageTeam Standard $1,000/mo annual equivalent or $1,250 monthly
EnterpriseCustom plan and usage$20/seat/mo annually plus API-rate usage for self-serve Enterprise, or sales-assisted terms
Variable meterProvider API list price + 40% model fee, Context Engine, computeShared subscription windows and optional usage credits, or direct API and cloud token rates

Claude Code wins the individual price. Claude Pro includes Claude Code plus the broader Claude product, with a rolling five-hour window and weekly limits shared across Claude surfaces. Usage credits can continue work at API rates when enabled. Max raises the allowance for heavy individual use.

Augment's fixed seat math becomes attractive at team scale. The first $100 also includes $100 of pooled service usage, after which model, context, and compute overage apply. A busy group can spend far more than the base. A quiet 50-person workspace can have low effective seat cost.

A combined setup pays for Claude and Augment. Price the Augment retrieval queries against Claude turns, tokens, and engineering search time saved. If Context Engine improves outcomes without reducing total work, it may still be valuable; the benefit simply has to be valued as quality rather than token savings.

How to choose or combine them

01

Benchmark Claude Code alone

Choose cross-service and repository-archaeology tasks with known answers. Record files read, tool calls, tokens, time, patch quality, and reviewer corrections.

02

Add Augment Context Engine MCP

Keep Claude model, effort, prompt, permissions, and clean worktree fixed. Repeat each task several times and compare retrieval plus outcome.

03

Run Auggie as the agent

Use the same model where available. Compare tool reliability, edit quality, permissions, resumption, subagents, integrations, and structured automation.

04

Test organization workflows

Backtest code review, run a ticket-to-PR flow, inspect audit records, revoke access, and verify data boundaries. Claude Code alone and Cosmos are different scopes here.

05

Model the month

Include Claude seats or API usage, Augment base and overage, retained tools, compute, administration, and human review or search time.

Choose Claude Code alone when Anthropic's agent, local workflow, customization, and price solve the problem. Choose Augment alone when multi-model retrieval and organization automation are the main requirements. Combine them when Claude remains the preferred agent and Augment repeatedly improves repository evidence.

Questions people ask

Augment is stronger as a dedicated semantic retrieval and enterprise orchestration platform. Claude Code is stronger as a focused Anthropic agent with a mature terminal harness and cheaper individual access. The best result depends on which layer is missing.

Yes. Run Auggie in MCP server mode or connect to Augment hosted Context Engine MCP. Claude Code receives a codebase-retrieval tool while retaining its native model, permissions, files, shell, skills, and session.

Claude Code. Claude Pro costs $20 per month or $200 annually and includes Claude Code. Augment public Business pricing starts at $100 monthly, designed to cover up to 50 seats and include $100 of shared usage.

Augment. Its current model list spans Anthropic, OpenAI, Google, Moonshot, and Prism routers. Claude Code uses Anthropic models and gains earlier access to Anthropic-native product features.

Augment has the stronger dedicated retrieval claim and published architecture for real-time, branch-aware semantic indexing across very large codebases. Claude Code reads the current working tree directly and can add code intelligence or Augment MCP. Test both on your repositories.

Augment has the broader packaged offer through Context Engine, Augment Code Review, and the Cosmos review fleet. Claude Code can perform review through prompts, subagents, hooks, CI, and the SDK, giving teams more building blocks and less packaged workflow.

Claude Code has a more cohesive documented local permission system with Manual, Accept Edits, Plan, Auto, managed settings, and checkpoints. Auggie has granular CLI permissions, while the IDE and Cosmos surfaces require separate control evaluation.

Pay for both when a controlled test shows Augment retrieval materially improves Claude Code outcomes or cost on important repositories. Start with Context Engine MCP rather than migrating the entire agent and workflow at once.

Sources

Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.

  1. Augment Context Engine
  2. Augment Context Engine MCP
  3. Auggie CLI reference
  4. Augment available models
  5. Augment pricing
  6. Claude Code overview
  7. How Claude Code works
  8. Claude pricing
Try it

Claude stays the engine.
You keep the fleet.

Continuum runs the real Claude Code beside other provider agents, gives every managed task its own worktree, and keeps the host session controllable from desktop, web, and phone.

free app · your subscriptions · local-first