Amp
Also known as: ampcode.com, Amp Code, Amp Frontier Corporation, Amp CLI, Amp Orbs, Amp Puck
Amp is a coding agent from Amp Frontier Corporation that runs each thread in an orb, a cloud machine holding a clone of the customer's repository, or locally through its CLI. It edits, runs tests, previews apps and ships to the base branch, picks up GitHub, GitLab and Linear issues on its own, and runs on schedules and webhooks. Users choose a mode from low to ultra, each mapped to a model, and plans run from a free tier to $200 a month.
Amp is a coding agent sold by Amp Frontier Corporation of San Francisco, which spun out of Sourcegraph in December 2025 as an independent company and calls itself an agent research lab of about 20 people. Its about page lists Intel, Westpac, Indeed, CircleCI, Algolia, Buildkite and ngrok among the companies using it, and a separate Amp Labs team embeds engineers with clients, as it has at Westpac in Sydney. The editor extensions were retired in March 2026, so the product now runs as a CLI, a web app, desktop and mobile apps, and an API.
Each thread runs in an orb, a cloud machine created for the thread with the repository cloned, a shared terminal and portals that show live previews of the app, where reviewers leave comments the agent picks up. The agent edits code, runs commands and tests, splits work across parallel subagents and can start further orbs to fan out a task, and the Ship button commits, rebases, runs the full test suite and pushes to the base branch, or opens a pull request instead.
Work reaches it from schedules written in plain language, from GitHub and GitLab issues, pull requests and CI results, from Linear assignments and Slack mentions, from plugin webhooks and from CI pipelines through service accounts. Puck, a personal assistant reached by voice, Slack or any MCP client, starts and steers agents, and Spaces gives a team one chat where an agent coordinates threads. Orbs can also run on the customer's own Docker, Kubernetes, Modal or Daytona infrastructure, and the CLI can turn a developer's machine into a runner.
The Hobby tier is free with nothing included, so usage is bought as credits at list price. Megawatt costs $20 a month with $20 of agent usage and 750 hours of small orbs, Gigawatt costs $200 with $200 of usage and 1,000 hours of xxlarge orbs, Teams adds no charge and pools credits, and Enterprise is custom. Customers can also bring their own keys or link ChatGPT, Claude, SuperGrok or GitHub Copilot subscriptions with no Amp fee. Amp holds SOC 2 Type II, supports SAML and OIDC sign in with SCIM on Enterprise, never trains on Enterprise data, and runs its server on Google Cloud in the United States with no self hosted edition.
It fits engineering teams that want agents working unattended in the cloud on issues and CI failures, with a way to keep execution inside their own cloud account and control over which model each mode uses. It fits less well for buyers who want approvals on by default, since tools run without asking until a policy plugin adds gates, or who need the server itself in a region of their choosing, and teams on GitHub Enterprise Server will find it unsupported.
Vendor details
Canonical URL
https://ampcode.com
Category
Coding agent
Subcategory
Coding agent that runs threads in cloud machines called orbs, from a CLI, a web app and desktop and mobile apps, and ships changes to the base branch
Funding status
Independent company, spun out of Sourcegraph in December 2025. Amp said at the spin out that it was profitable, and its about page lists a16z, Sequoia, Felicis, Redpoint, Craft Ventures and Goldcrest as investors, without amounts.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
GitHub App that clones, pushes, opens pull requests and reads CI as the user or as the Amp bot, and GitLab on gitlab.com and self managed with merge requests and pipelines. Linear issues by assignment or mention, Slack read and write tools, and Gmail, Calendar, Docs, Sheets and Drive through the workspace's own OAuth client. MCP client for local and remote servers with OAuth, plugins that register custom tools, Bitbucket and other Git hosts over Git credentials, and TypeScript and Python SDKs.
In practice
A team turns on an orb spawner for its main GitHub repository, so every new issue gets its own orb. The agent reproduces the bug, writes the fix, runs the tests and opens a pull request, and a developer checks the change in the portal preview before merging.
A platform engineer schedules a nightly thread in plain language to update dependencies. When CI fails on that commit, the agent, which follows CI for it, reads the failure, pushes a fix and ships to the base branch once the full suite passes.
A bank's security team runs orbs on its own Kubernetes cluster with outbound connections only, limits the workspace to approved models and MCP servers, and pushes a policy plugin that rejects shell commands outside an approved list.
Sources & related URLs
Related / legacy domains
Research sources
Agentic Index coverage score
12.0 / 14 capabilities · 86%
| Integrations & Tool Calling | Full |
|---|---|
|
The GitHub App clones private repositories, pushes commits, opens pull requests and reads CI as the user, or as the Amp bot when a project chooses, with signed commits, and GitLab on gitlab.com or self managed gets merge requests and pipeline reads through glab. Linear issues are read and updated through Linear's MCP server, Slack messages are read and posted with built in tools, and Gmail, Calendar, Docs, Sheets and Drive connect through the workspace's own OAuth client with access set per service. Any local or remote MCP server can be added with OAuth or bearer headers, and plugins register custom tools. Sourceampcode.com/docs/integrationsread 2026-11-08 |
|
| Workflow Orchestration | Full |
|
Inside a thread, Task subagents take parts of the work in parallel, each with its own context window, and report a summary back, with roles for code search, reading outside repositories and a second opinion from a reasoning model. Agents can also start new orb threads, fan a job out across them, exchange files and collect the results, Puck can split one request across several agents, and Spaces gives a team one chat where an agent starts and coordinates threads. Plugins can define custom subagents for a workspace. Sourceampcode.com/docs/models-and-subagentsread 2026-11-08 |
|
| Knowledge Grounding & RAG | Partial |
|
The agent searches the repository on each run through its search role and the finder plugin, the librarian clones outside repositories to read them, and attachments such as PDFs, logs and datasets join a thread. Project snapshots prepare orb images that are reused for up to 72 hours, but they hold the working environment, not a search index, and the server keeps no index or embeddings of the whole codebase between runs. The External API adds regex and path search over repositories hosted in the workspace. Sourceampcode.com/securityread 2026-11-08 |
|
| Human Oversight & Guardrails | Full |
|
Tools run without asking by default, and gates come from policy plugins that approve or reject each tool call, which admins can push to a whole workspace, along with Ask First or Disabled settings for each remote MCP tool and allow and reject rules in settings. Before shipping, changes sit in a Changes pane for review, and the Ship button can push straight to the base branch, push to a branch for a pull request, or follow a custom prompt. Linear sessions ask before opening a pull request, and posting to Slack asks first. Sourceampcode.com/docs/customize/pluginsread 2026-11-08 |
|
| Security, Identity & Governance | Full |
|
Workspaces sign in through SAML or OIDC once one member pays, and Enterprise adds SCIM and sign in without a paid member, with Member and Admin roles plus Auditor, Billing Manager, Group Manager, Usage Viewer, Global Plugin Manager and Model Provider Key Manager on Enterprise. Admins can enforce passkeys, limit MCP servers to an allowlist and restrict models, and IP allowlisting comes on request at extra charge. Amp holds SOC 2 Type II with an annual penetration test, encrypts with TLS 1.2 and AES 256, redacts secrets, and never trains on Enterprise data. Sourceampcode.com/docs/collaborate/workspacesread 2026-11-08 |
|
| Observability & Auditability | Full |
|
Every thread keeps the prompts, responses, tool calls and results as a record of what the agent did, shareable as private, workspace, group or unlisted and exportable as Markdown or JSON, and commits carry a thread ID trailer that links code back to the thread that wrote it. A usage dashboard breaks spending down by model, mode, member, orb size and provider with CSV export, each thread shows its cost, Enterprise admins see authentication audit logs, and admin access to private threads is logged. Application audit logs go to customers on request. Sourceampcode.com/docs/threadsread 2026-11-08 |
|
| Memory & State Persistence | Partial |
|
Past threads can be found, searched and referenced by ID from a new thread, Puck keeps a separate history for each of its conversations, and artifacts stay attached to the thread that made them. AGENTS.md files at repository, folder, home and system level, plus personal and workspace global versions edited in settings, carry standing instructions into every session and can be viewed and edited. There is no memory store the agent writes and recalls across threads on its own, and a deleted thread is removed within 30 days. Sourceampcode.com/docs/threadsread 2026-11-08 |
|
| Deployment & Data Residency | Full |
|
The server runs multitenant on Google Cloud in the United States, with inference on US servers of the model providers and hosted orbs on e2b, and the server has no self hosted edition. Orbs themselves can run on infrastructure the customer controls, on Docker, a Kubernetes Helm chart for GKE, EKS or AKS, Modal, Daytona, Vercel and other providers, with outbound connections only, so code checkout and execution stay in the customer's cloud, region or account while the agent loop and inference stay with Amp. The CLI can also turn a developer's machine into a runner, and Enterprise adds regional endpoints for bring your own key providers. Sourceampcode.com/docs/orbs/self-hostedread 2026-11-08 |
|
| Prebuilt Agents / Templates / Packs | Full |
|
A shared plugin catalog holds installable subagents, the oracle for a second opinion from a reasoning model, the librarian for reading outside repositories and the finder for code search, plus a mode built on Grok 4.6, and admins can push plugins and skills to a whole workspace. Built in skills cover writing new skills and plugins, built in prompts ship, update from the base branch, refactor commits and review, and Painter generates images. Skills can sit at personal, workspace or repository level, and skills written for Claude Code load as they are. Sourceampcode.com/docs/customize/global-plugins-and-skillsread 2026-11-08 |
|
| Triggers & Channel Coverage | Full |
|
Threads take schedules written in plain language, one time or repeating, and orb spawners start a fresh orb for every new GitHub or GitLab issue or pull request, while subscriptions follow comments and CI results for a commit. Assigning a Linear issue or mentioning @Amp in Linear or Slack starts work, plugin webhooks accept public POST requests with at least once delivery, and service account keys let CI pipelines and scripts start orb threads. A website feedback widget that sends reports to the agent sits under an experimental tab. Sourceampcode.com/docs/orbs/automationsread 2026-11-08 |
|
| Model Flexibility & Routing | Full |
|
Each thread starts in a mode, low, medium, high or ultra, mapped by default to GLM 5.3 Flash, Claude Opus 5.5, OpenAI's GPT 6 Astra and Claude Fable 5.1, and Tune Modes lets a user or workspace admin pin a different model to each mode, while Allowed Models limits the choice across a workspace. Customers can bring API keys or Bedrock, Vertex and Azure AI Foundry accounts, route through gateways such as OpenRouter, Vercel, Cloudflare or a custom URL, or link ChatGPT, Claude, SuperGrok or GitHub Copilot subscriptions, with a ChatGPT preset sending low, medium and high to OpenAI models. Sourceampcode.com/docs/the-dialread 2026-11-08 |
|
| APIs / SDKs / MCP Extensibility | Full |
|
TypeScript and Python SDKs drive Amp with an API key, the CLI runs headless with streaming JSON in and out, and External API version 2 covers threads, messages, usage and daily analytics under OAuth scopes with an OpenAPI specification, for machine to machine applications. Puck's MCP server exposes three tools that start Puck conversations and work with orbs, so assistants such as Claude or ChatGPT can hand work to Amp. Sign in with Amp for outside apps is not generally available. Sourceampcode.com/docs/sdkread 2026-11-08 |
|
| Testing, Debugging & Optimization | Partial |
|
Shipping runs the project's full test suite and lets a failing test through only when it also fails on the base branch, a review prompt and a review subagent read the diff before it lands, and fanning out fresh orbs can show how well a repository suits agents. Amp has no harness, scored test set or experiment loop for measuring how the agent itself behaves when prompts, models, skills or plugins change, so quality checks run on the customer's code with the customer's own tests. Sourceampcode.com/docs/orbs/shippingread 2026-11-08 |
|
| Browser / Computer-use | Partial |
|
Each orb has a Desktop tab with an interactive Linux desktop, and on website tasks the agent drives the same Chrome window, records the run and reports back with screenshots, though computer use in the orb desktop is labeled early and unsupported. Chrome Profiles keep sign ins across orbs only with Experimental Features turned on, and macOS Satellites, cloud Macs the agent builds on and taps through, are in beta behind a waitlist. Portals show live previews where people leave comments for the agent. Sourceampcode.com/docs/orbs/portalsread 2026-11-08 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Free Hobby tier with credits bought at list price, or Megawatt at $20 a month with $20 of agent usage and 750 hours of small orbs.
A monthly plan fee that converts into the same amount of agent usage, plus orb compute by the hour beyond the included hours and prepaid credits for further usage.
Included quota
Megawatt includes $20 of agent usage and 750 hours of a1.small orbs a month. Gigawatt includes $200 of usage and 1,000 hours of a1.xxlarge orbs. Included usage does not roll over.
What is public
Every plan price, the usage each plan includes, hourly orb and macOS Satellite rates, credit expiry, and what Enterprise adds.
Billing mechanics
A monthly plan fee converted into agent usage, prepaid credits that expire after 12 months, and hourly orb charges. Teams pool purchased credits while each member keeps a personal tier, and Enterprise bills pooled credits with no paid seats plus a platform fee set in the order form.
Cost watchouts
Hobby includes no usage, so every request draws on bought credits. Higher modes and larger orbs cost more per hour of work, a1.3xlarge orbs need Gigawatt or Enterprise, macOS Satellites cost $2 to $4 an hour, and self hosted orbs on Enterprise bill at the standard orb price.
Variable cost rationale
The plan fee becomes the same amount of usage, so a light month costs the plan price. Heavy months draw on prepaid credits at list model prices and hourly orb charges, which grow with the mode chosen, the orb size and how long threads run. Bringing the customer's own keys moves model costs to its provider.
Additional watchouts
Plan naming differs between the pricing page and the docs, and the Gigawatt price appears only in the docs.
Overage / add-ons
Beyond the included amount, usage draws on purchased credits at model list price with no markup for individuals and non enterprise workspaces, and orbs bill by the hour from $0.08 for a1.tiny to $2.13 for a1.3xlarge. Purchased credits expire after 12 months.
Sales call required
Mixed (some tiers require a call)
Free / trial
Hobby is free with nothing included. Usage is bought as credits at list price, and the customer's own keys or linked subscriptions carry no Amp fee.
Lowest paid plan
Megawatt at $20 a month
Commercial notes
Amp Frontier Corporation spun out of Sourcegraph in December 2025 and sells Amp on its own. Customers can run on their own model keys or linked ChatGPT, Claude, SuperGrok or GitHub Copilot subscriptions with no Amp fee outside Enterprise.
Key ambiguities
The pricing page calls the paid tier Individual and shows only the $20 figure, while the docs name Megawatt at $20 and Gigawatt at $200. The security page lists single sign on as Enterprise only, while the workspace and pricing pages offer it to Teams with one paid member.
Cancellation / refund
Monthly plans. Purchased credits expire after 12 months. Enterprise runs on an order form with a platform fee percentage.
Support SLA / resale
Enterprise adds pooled credits with no paid seats, SCIM, spend limits and reporting, audit logs, network and MCP access controls, IP allowlisting and regional endpoints.
Missing data
Enterprise list prices and the platform fee percentage.
Related vendors
- Cognition — Maker of Devin, an autonomous AI software engineer
- 10Web — Agentic website platform whose AI agents build, host and manage…
- AgentUI — Managed AI app builder for operations teams that generates and hosts…
- Aider — Open source, model agnostic terminal coding agent that edits your…
- Anthropic Claude Code — Anthropic's agentic coding system across terminal, desktop, IDE, web…
- Anything — Anything is an AI app builder whose agent turns a plain language…