Amp
by Sourcegraph
An agentic coding tool that delegates to parallel subagents and charges LLM tokens at cost.
Amp is an agentic coding tool from Sourcegraph that runs in the terminal and inside VS Code, Cursor, Windsurf, Neovim and Zed. It routes work across Claude, Gemini and GPT models, spawns independent subagents to run multi-file changes in parallel, and bills frontier-model tokens at provider cost with zero markup for individuals and teams.
Amp is an agentic coding tool from Sourcegraph that runs as an interactive terminal TUI and as editor extensions for VS Code, Cursor and Windsurf, with plugins for Neovim and Zed. Rather than exposing a model picker, it routes work across frontier models from Anthropic, Google and OpenAI through four effort modes — Low for small well-defined tasks, Medium as the default balance of quality, speed and cost, High for complex work, and Ultra for the hardest open-ended problems. Its differentiating mechanic is delegation: Amp spawns independent subagents with their own context windows so multi-file refactors execute in parallel while the main thread stays clean, and an 'Oracle' tool provides a second-opinion reasoning model for architecture review, routing to Claude Fable 5 or GPT-5.6 depending on the user's subscriptions. Context is steered with AGENTS.md files, @-file mentions and thread references; the tool layer spans shell and file operations, custom MCP servers, a Skills system and a plugin API, with an execute mode and streaming JSON output for scripting. Amp launched publicly as ad-supported 'Amp Free' on 21 October 2025, replacing Sourcegraph's Cody, whose Free and Pro plans were retired that July; Sourcegraph announced in December 2025 that Amp would be spun out as a separate company, and a rebuilt CLI, Neo by Amp, launched in May 2026. Commercially it is unusual in charging LLM tokens at provider cost with zero markup for individuals and teams, with Enterprise paying a 50% premium. Team features — shared billing pools, workspace thread visibility, multiplayer threads, SSO, SCIM directory sync and per-user entitlements — target engineering organisations, and Amp is SOC 2 Type II certified with model training disabled by default and no self-hosted option.
An engineering leader running a team on a large codebase who wants agentic coding without per-seat lock-in — Amp bills tokens at provider cost, pools spend across a workspace, and gives thread visibility and SSO/SCIM for governance.
Large multi-file changes finish in parallel across subagents instead of sequentially, on a bill that matches what the underlying model APIs would have charged directly.
At a Glance
- Category
- Developer Tools
- Pricing
- Freemium, Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, VPs of Engineering, Engineering Managers, Enterprise Developers, Platform Engineers
- Deployment
- Cloud-only
- Founded
- 2013
- Headquarters
- San Francisco, United States
Key Features
- ✓Parallel subagents
Spawns independent agents with their own context windows so separate areas of a codebase are changed concurrently without polluting the main thread.
- ✓Four effort modes (Low/Medium/High/Ultra)
Explicitly trades cost and latency against reasoning depth per task, instead of forcing one model choice across every request.
- ✓Oracle second-opinion model
Routes hard reasoning and architecture review to a separate high-reasoning model, so plans are critiqued before an agent executes them.
- ✓Zero-markup token billing
Charges the same provider rates as a direct Anthropic, Google or OpenAI contract for individuals and teams; Enterprise pays a 50% premium.
- ✓Workspaces, multiplayer and thread sharing
Shared billing pools plus private, group or workspace-visible threads let teammates reuse an agent session as code-review context.
- ✓MCP servers, Skills and plugin API
Extends the built-in shell and file tools with custom MCP servers, bundled skill instructions and event handlers for org-specific workflows.
- ✓Enterprise identity and audit controls
SSO via Okta and other IdPs, SCIM directory sync, passkeys for sensitive actions, per-user entitlements and 30-day minimum audit logging.
Capabilities
Use Cases
- •Cross-cutting interface refactor
Parallel subagents update interface definitions across a large Go codebase in minutes rather than an hour of sequential edits.
- •Pre-merge architecture review
Oracle mode critiques a proposed change at full reasoning depth before the team spends review cycles on the pull request.
- •Scripted agent runs in CI
Execute mode with streaming JSON output drives Amp non-interactively from build pipelines and internal automation tooling.
- •Onboarding to an unfamiliar service
AGENTS.md conventions plus thread references let a new engineer inherit prior agent context instead of re-explaining the codebase.
- •Governed team rollout of agentic coding
Workspace billing pools, per-user entitlements, SSO and SCIM give a platform team spend and access control across developers.
Ideal For
Best For
- ✓Large multi-file refactors where parallel subagents beat a single sequential agent thread
- ✓Teams that want frontier-model access at provider API cost rather than a marked-up per-seat subscription
- ✓Terminal-first developers who script agents into CI or automation via execute mode and streaming JSON output
- ✓Engineering orgs needing shared billing pools, thread sharing for code review context, SSO and SCIM provisioning
- ✓Architecture review and second-opinion reasoning on a plan before committing to a large change, via Oracle mode
Not Ideal For
- ✗Organisations with self-hosted, air-gapped or on-premise requirements — Amp's own security documentation states it offers no self-hosted deployments
- ✗Developers who want the polished native IDE experience of Cursor; hands-on reviewers consistently rate Amp's editor integration behind it
- ✗Finance teams that need a flat, predictable per-seat line item — usage is metered against actual model spend, so cost varies with how hard the agent works
- ✗Tightly coupled legacy codebases where autonomous multi-file edits are risky; reviewers note Amp's dependency tracking 'is not infallible' and plans need review before large runs
Deployment
Market Analysis
Pros
- ✓Parallel subagent execution is a real throughput win on large changes — one hands-on review measured seven concurrent subagents finishing a Go interface refactor in eleven minutes versus roughly an hour sequentially
- ✓Zero-markup token billing removes the usual agentic-coding margin, and the free ad-supported tier is described by reviewers as genuinely capable rather than a crippled demo
- ✓Multi-model routing means the tool picks Claude, Gemini or GPT per workload instead of locking the team to one vendor's model roadmap
- ✓SOC 2 Type II with model training disabled by default, secret redaction before data leaves the client, SSO, SCIM and audit logging — a stronger default posture than most coding agents
- ✓Hacker News commenters praise the CLI's deliberate design, one noting it is 'very well thought out to make you do things the correct way'
Cons
- ✗IDE polish lags Cursor's native experience, which is the most consistent complaint in hands-on reviews and the main reason developers keep both tools
- ✗Dependency tracking on tightly coupled codebases 'is not infallible' per one detailed review, so execution plans need human review before large autonomous runs
- ✗Usage-based billing makes spend unpredictable relative to a flat per-seat licence, which is a procurement problem even when the unit rate is lower
- ✗No self-hosted or on-premise deployment at all, ruling it out for air-gapped and some regulated environments
- ✗Product maturity risk: it replaced Cody, was announced for spin-out from Sourcegraph in December 2025 and had its CLI rebuilt as Neo in May 2026 — three platform shifts in under two years, and Cody users already lost their Free and Pro plans once
- ✗Almost no independent review volume — Product Hunt shows zero user reviews and G2 carries no Amp-specific listing, so there is little aggregated evidence beyond individual blog posts
Pricing
Amp Free
$0
- ✓Daily credit grant underwritten by sponsor ads
- ✓No credit card required
- ✓Available in VS Code, Cursor, Windsurf and the CLI
- ✓Ads shown in non-coding contexts
Pay-as-you-go
Provider API cost, zero markup
- ✓Frontier model tokens at Anthropic/Google/OpenAI rates
- ✓Purchased credits expire after 12 months
- ✓No subscription commitment
Subscription
Contact for pricing
- ✓Included monthly agent and orbs usage
- ✓Credits reset monthly
- ✓Priority routing and higher subagent throughput
Enterprise
Contact for pricing
- ✓SSO and SCIM directory sync
- ✓Per-user cost entitlements
- ✓Model training cannot be enabled
- ✓50% premium on model pricing
Amp's commercial angle is that it does not mark up model tokens: individuals and teams pay the same rate as a direct Anthropic, Google or OpenAI API contract, and Enterprise pays a 50% premium in exchange for identity, audit and data-handling controls. The entry point is a free, ad-supported daily credit grant requiring no credit card, after which usage is metered against real model spend rather than a per-seat licence — so cost tracks how hard the agent works, not headcount. Subscriptions bundle a monthly allowance of agent and orbs usage that resets each month, while separately purchased credits expire after twelve months. No public list price is published for the paid tiers.
Security & Compliance
Sources
This page was written from 6 sources, 4 on domains other than ampcode.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Momentic
Agentic QA platform that writes, runs and self-heals end-to-end tests for web and mobile apps
Raindrop
AI agent monitoring that catches silent production failures: Sentry for AI agents
GitLab Duo Agent Platform
Agentic AI across the whole GitLab DevSecOps lifecycle: planning, coding, code review, CI/CD and security agents under one governance model
CodeRabbit
AI code review and agentic change management for teams shipping human- and machine-written code