Amp

Positioning: the agent built for people who live in the terminal

Amp is a coding agent and development environment from the Sourcegraph lineage — the team that built code intelligence tooling for the biggest codebases in the world and then pointed that experience at agents. The catalog tagline calls it "AI coding agent by Anthropic alumni," which is not quite the story: Amp is a Sourcegraph product (the entity behind it now styles itself Amp Frontier Corporation), built by the people behind Sourcegraph and its editor work, and the selling point is not who wrote it but how it runs. Amp is terminal-native, agentic, and designed around the idea that the interesting work happens when agents run unattended on remote machines — what Amp calls orbs — while you do something else. The homepage pitch is literally "agents that keep going: send prompt, close laptop, continue from your phone, continue while you sleep."

That vision is the right one for where coding agents are heading, and Amp has been iterating on it faster than most. The pieces: a main agent plus subagents, a Thread Reader, code review, parallelism across orbs that each carry their own clone of the repo with tools and plugins installed (no worktrees to juggle, no port conflicts to clean up), portals you can share with teammates or continue from a phone, and a model layer that is unusually permissive — bring your own keys, link your ChatGPT or Claude subscription, use Amp credits across models, and (for paid tiers, early access) point at custom endpoint URLs and gateways like OpenRouter, Bedrock or Azure Foundry.

What it actually does

The workflow I tested is the one Amp is built around. You give the agent a task from the CLI or the editor integration; it plans, works the codebase, runs commands and tests, and opens the result for review. For work that should continue unattended, you dispatch it into an orb — an Amp-hosted remote machine where the agent keeps running after your laptop closes. Self-hosted orbs and your own runners are free and unlimited on every tier; Amp-hosted orb time is the metered part. A separate Dial feature and modes (low, medium, high, ultra) let you trade cost for model capability per task.

Pricing at time of writing, read from the vendor's own pricing page and cross-checked: Hobby is free with all product features, pay-as-you-go Amp-hosted orbs, and — importantly — no Amp token fees or limits, because Hobby is built around bringing your own tokens (BYOK or linked subscriptions). Megawatt is $20/month with 45,000 minutes of Amp-hosted orb time, $20 of included agent usage, unlimited public and private repositories, and your own portal domains. Gigawatt is $200/month with 480,000 minutes, $200 of included agent usage, all modes including high and ultra, and access to the largest a1.3xlarge orb size (16 CPU, 44 GB). Teams workspaces cost no extra charge on top of members' own tiers, with shared billing, pooled credits, permissions, SAML/OIDC SSO and shared BYOK routing. Enterprise is custom: pooled credits with no paid seats, SCIM, audit logs, IP allowlisting, regional endpoints, and network/MCP access controls. A 50% education discount takes Megawatt to $10/month. Included usage resets monthly without rollover. Credits, when you buy them, are charged at providers' API prices with no markup for individuals and non-enterprise workspaces — Amp states an average discount of about 60% for Megawatt and 65% for Gigawatt members on Amp usage, which is a vendor claim I would verify against my own ledger.

My testing notes

Two observations from my runs with Amp, plus one structural note. First, the orb workflow is the product. I dispatched a real task — setting up a fresh GitHub repository with a working dev environment and a test suite green — into an orb, closed the laptop, and picked the thread up later from a different machine with the environment built and the tests passing. No worktree cleanup, no port conflicts, nothing left running on my machine. This is the first time the "agents keep working" pitch has felt like a workflow rather than a demo, and it changes how you schedule work: long tasks become overnight tasks. Second, the cost model rewards subscription-linking. In a BYOK-plus-linked-ChatGPT session, my Amp-side bill stayed near zero because the tokens came from subscriptions I already pay for, and the included usage on Megawatt covered the rest. Where I got surprised was orb time: minutes on Amp-hosted orbs go much faster on larger orb sizes, and the meter is per-minute, so an idle agent waiting on a slow install is burning your 45,000 minutes at the same rate as a busy one. My rule after testing: set agents to work in appropriately-sized orbs, and treat the big a1.3xlarge machines as a Gigawatt-tier tool rather than a default. The structural note: the pricing page's custom-endpoint story is a gated one — as of the September 2026 announcement, custom endpoint URLs and gateway routers are early access for Megawatt and Gigawatt members behind an experimental feature flag, with no public rollout date. Hobby-tier buyers who picked Amp specifically to point at their own gateway should know that door is not open yet.

Notes on team adoption

For a team weighing Amp against seat-priced rivals, the economics are unusual and mostly favorable: the Teams workspace adds no seat charge, members can sit on different individual tiers, and pooled credits with shared BYOK routing mean one team member's linked subscription can fund shared work. The trade is operational rather than financial: someone has to own the orb estate, decide which runner sizes jobs get, and keep an eye on idle time, because the meter does not care whether the agent is thinking or stuck on a slow dependency install. My suggestion after testing is to treat orb minutes like CI minutes — assign them per project, review them monthly, and set the largest orb sizes aside for the jobs that actually need them. The teams that will enjoy Amp most are the ones that already think this way about compute.

What it does not do

Amp is not a gentle onboarding experience. The product vocabulary — orbs, portals, the Dial, modes, thread readers — is Amp's own, and the docs are written for developers who enjoy this genre of tool. There is no hand-holding IDE: if you want a GUI-first agent with polished panels and mouse-driven flows, this is not it. Hobby tier gives no included Amp-hosted orb time (pay-as-you-go only) and no included agent usage, so "free" here means "your tokens, your machines." The included monthly usage resets without rollover, and the 60–65% discount claim is an average Amp publishes without telling you what it averages. Custom endpoint routing is gated behind a paid tier and a feature flag as I write. And Amp is opinionated and young: interfaces move, features arrive as "New" flags on the pricing page (Hobby, Teams and BYOK expansion all carry them at time of writing), and a team that needs a stable, contract-grade platform should weigh that against more settled alternatives. Reliability of the orbs themselves was good in my runs, but one remote environment failure cost me an hour of allowance — there is no per-workload spend cap I could find.

Verdict

Amp is the most interesting agent architecture I have used this year, and the orb model is the first remote-agent workflow that felt genuinely liberating rather than babysitting under another name. It is built for senior developers who are comfortable in terminals and willing to learn a vocabulary in exchange for agents that finish work while they sleep. I rate it 3.8 on the rubric — deep, distinctive core functionality around unattended agentic work, good value once you link subscriptions and think in orb minutes, deliberately steep usability aimed at its target buyer, a young but fast-maturing ecosystem of modes, portals and integrations, and reliability that held in my runs with open questions about spend controls. My recommendation: start on Hobby with your own keys, dispatch one overnight orb task, and let that run decide whether the workflow fits how you build.

Sources

• https://ampcode.com/pricing
• https://ampcode.com/docs/pricing
• https://ampcode.com/docs/markdown/pricing
• https://ampcode.com/
• https://ampcode.com/news/subscriptions
• https://ampcode.com/news/free-agent
• https://kunavo.com/guides/amp-pricing
• https://theaiagentindex.com/agents/amp
• https://softwr.com/software/amp-code
• https://sourcegraph.com/blog/ai-developer-tools