What it is
Amp is a coding agent built by the team behind Sourcegraph. It spun out of Sourcegraph as its own company, Amp Frontier Corporation, in December 2025, led by Sourcegraph co-founder Quinn Slack.
Amp runs on your own machine or on its cloud machines, called orbs, which are billed by the minute. Threads (agent conversations) can be shared, worked on together and reused across a team, and teams share plugins, skills and model routing. Since 19 October 2026 you can assign a Linear issue to Amp, which starts an orb and reports back in the issue; since 22 October 2026 companies can run orbs on their own infrastructure.
Since September 2026 Amp charges no token fee when you bring your own API keys or an existing model subscription such as ChatGPT or Claude. The product itself cannot be self-hosted.
Our review
Amp's argument is that a coding agent should be shared infrastructure, not a personal tool. Threads can be shared and worked on by several people, skills and plugins are pooled, and the Teams tier adds SAML or OIDC single sign-on at no extra charge. Assigning a Linear issue to Amp now spins up an orb and posts the result back into the issue, which is the workflow most teams actually want from a background agent.
The business model has settled on compute. Since September 2026 there is no Amp fee on tokens if you bring your own keys or a ChatGPT or Claude plan; you pay for orbs instead. The $20 Megawatt tier includes 45,000 orb minutes. The pricing page shows a Gigawatt toggle without a price, which Tech Insider puts at $200 a month, and it does not publish the per-minute rate for pay-as-you-go orbs. Tech Insider traces at least three earlier changes in 2026, from the paused ad-supported free tier to the July subscriptions, so budget on a short horizon.
The security documentation is unusually specific. Amp holds a SOC 2 Type II report and says neither it nor its subprocessors train on your data unless you opt in, and training can never be switched on for Enterprise workspaces. Threads sit in PostgreSQL on Google Cloud in the US; most inference runs on Anthropic, with OpenAI, Vertex, Bedrock and others behind it; orbs run on E2B and hold a clone of your repository. Deleted threads are purged within 30 days. One trap: linking a personal ChatGPT or Grok subscription puts those prompts under that provider's consumer terms, so training has to be switched off there.
Self-hosted orbs, launched on 22 October 2026, let regulated teams keep execution inside their own Docker, Kubernetes or Modal estate with no inbound ports. Amp itself says most people should not bother, and it is free for individuals and teams but billed at standard orb rates for Enterprise from 29 October. Results still depend on whichever model you bring, which makes Amp's value the harness and the team features, not raw model quality, and that is a thinner moat than Cursor's or Claude Code's.
Where it earns its place
A 10-engineer startup already paying for Claude Max seats
Hand routine Linear tickets to an agent and review the pull requests
Assign the issue to Amp; it starts an orb, works in a clone of the repo and reports in Linear. Model usage comes out of the existing Claude plans, so Amp's cost is orb time: Megawatt at $20 a head with 45,000 minutes, or $200 a month for ten.
A platform engineer at a bank that forbids code leaving its network
Run agent tasks on internal infrastructure with access to private services
Self-hosted orbs connect outbound from a provider process in GKE, EKS or AKS through Amp's Helm chart. Threads still live on Amp's US Google Cloud, so the security review has to cover that and the model provider route.
An engineering director standardising agent use across 40 developers
Share good prompts and skills, and control who can see what
Teams adds thread sharing, permissions, analytics and SAML or OIDC SSO with no seat fee. SCIM, audit logs, IP allowlisting and network and MCP access controls need Enterprise, at custom pricing.
Before you commit
- Setup
- Install the CLI, sign in, and connect an API key or a ChatGPT or Claude subscription; orbs and the Linear agent are configured in workspace settings.
- Integrations
- Linear agent sessions, GitHub repositories, MCP servers, plugins and skills; self-hosted orb examples for Docker, E2B, Modal, Vercel, Daytona and Kubernetes.
- Data and privacy
- SOC 2 Type II; threads in PostgreSQL on Google Cloud in the US; orbs on E2B; no training unless opted in, never on Enterprise; deleted threads removed within 30 days; not self-hostable.
- Cost at scale
- Ten developers on Megawatt pay $200 a month plus their own model plans; orb use beyond 45,000 minutes each is pay-as-you-go at a rate the pricing page does not publish.
- Limits
- Higher quotas and rate limits are Enterprise only; project snapshots persist up to 72 hours; secret redaction is best-effort.
What's good
- No token fee with your own keys or ChatGPT and Claude subscriptions
- Shared threads, multiplayer sessions, skills and SSO on Teams at no extra charge
- SOC 2 Type II; no training unless opted in, never on Enterprise
- Linear issues can be assigned straight to an orb
- Self-hosted orbs for teams that cannot run code in Amp's cloud
What's not
- Pricing has changed several times during 2026
- Gigawatt price and pay-as-you-go orb rate not on the pricing page
- Personal ChatGPT or Grok subscriptions fall under consumer training terms
- The Amp service itself cannot be self-hosted
- Results depend on the model plan you bring
Best for
Small teams that already pay for Claude or ChatGPT and want a shared, cloud-capable agent on top.
Alternatives
Amp: common questions
- Is Amp worth it?
- A team-first coding agent that charges for cloud machines rather than tokens, with a security page better than its pricing page. We rate it 3.8 out of 5.
- How much does Amp cost?
- Hobby free with pay-as-you-go orbs and your own model plan or keys; Individual (Megawatt) $20/month with 45,000 orb minutes; Gigawatt reported at $200/month; Teams no extra charge; Enterprise custom.
- What are the downsides of Amp?
- Pricing has changed several times during 2026; gigawatt price and pay-as-you-go orb rate not on the pricing page; personal ChatGPT or Grok subscriptions fall under consumer training terms; the Amp service itself cannot be self-hosted; results depend on the model plan you bring.
- Who is Amp best for?
- Small teams that already pay for Claude or ChatGPT and want a shared, cloud-capable agent on top.
- How much does Amp cost for a team or in production?
- Ten developers on Megawatt pay $200 a month plus their own model plans; orb use beyond 45,000 minutes each is pay-as-you-go at a rate the pricing page does not publish.
- Is Amp safe for company data?
- SOC 2 Type II; threads in PostgreSQL on Google Cloud in the US; orbs on E2B; no training unless opted in, never on Enterprise; deleted threads removed within 30 days; not self-hostable.
- What are the best Amp alternatives?
- The closest alternatives we have reviewed are Claude Code, OpenAI Codex, Warp and Factory (Droids).
Answers checked Oct 22, 2026.