AmpvsCosine
Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.
Amp
Amp Frontier Corporation (spun out of Sourcegraph) · Terminal agent
#111MCP
Panel
5.51 spec wins
- Reliability
- 5.7
- Usefulness
- 6.5
- Cost
- 5.2
- Longevity
- 4.8
“Twenty dollars a month buys twenty dollars of agent usage, which is at least an honest exchange rate.”
Cosine
Cosine AI · Terminal agent
#100MCP
Panel
5.81 spec wins
- Reliability
- 5.7
- Usefulness
- 6.8
- Cost
- 5.2
- Longevity
- 5.5
“It used to be called Genie and is now named after a function that always comes back around, which is a bold choice for a startup.”
Spec by spec
| Spec | Amp | Cosine |
|---|---|---|
| Architecture | ||
| Category | Terminal agent | Terminal agent |
| Runs | local, cloud, sandbox | local, cloud, sandbox |
| Platforms | macos, linux, windows, web | macos, linux, windows, web |
| Context window | not documented | not documented |
| Protocols | ||
| MCP client | Yes | Yes |
| MCP server | No | No |
| Capabilities | ||
| Runs terminal commands | Yes | Yes |
| Multi-file edits | Yes | Yes |
| Git operations | Yes | Yes |
| Browser control | YesOrbs preinstall an `agent-browser` tool (https://ampcode.com/docs/orbs/customizing) and the agent "can install dependencies, run your app, spawn browsers, and test its work end to end". | YesBrowser tools drive Chrome or Chromium over the DevTools Protocol, which requires launching the browser with --remote-debugging-port and setting cdp_url in ~/.cosine/config.toml. |
| Sandboxed execution | YesIsolation comes from cloud orbs, where "every orb thread gets a fresh, isolated environment" with a whole machine to itself; the local CLI has no container or OS-level sandbox. | YesRemote environments are isolated containers built from a Dockerfile; local runs are not sandboxed. |
| Multi-agent orchestration | Yes | Yes |
| Headless / CI mode | YesNon-interactive `amp -x` execute mode is documented at https://ampcode.com/docs/cli/execute-mode. | No |
| Models | ||
| Backbone | OpenAI GPT-5.6, Anthropic Claude Fable 5.1, fast models | Lumen Scout, Lumen Outpost, GPT-5.5, Claude Sonnet 4.6, Claude Opus 4.7, Gemini 3.1 Pro, DeepSeek, Kimi K2.6, MiniMax M2.7, Qwen |
| Bring your own model | YesAmp's Model Routing settings accept your own Anthropic API key for the workspace, but you still pick an Amp mode (low/medium/high/ultra) rather than an arbitrary model, endpoint or cloud deployment. | YesLimited to ChatGPT and Claude models accessed through your own provider account. |
| Local models | NoNo custom base URL, LLM gateway or local endpoint is documented; all inference is routed through Amp's own service. | No |
| Cost | ||
| Pricing model | mixed | subscription |
| Starts at | $20/mo | $19/mo |
| Free tier | No | No |
| Bring your own key | Yes | YesYou can sign in with your own ChatGPT or Claude Max subscription, which routes billing to that provider instead of consuming Cosine credits. |
| Openness | ||
| Open source | No | No |
| License | proprietary | proprietary |
| GitHub stars | n/a | n/a |
Which one would each critic pick
| Critic | Amp | Cosine | Pick |
|---|---|---|---|
| El Juez | — | — | not enough reviews |
| El Amigo | 6.8 | 6.3 | Amp — Pick Amp if you want a sandbox without running Docker yourself; pick Claude Code or Codex CLI if you want a bigger included quota and one vendor. |
| El Crítico | 6.0 | 6.0 | no preference |
| El Profesor | 5.5 | 5.5 | no preference |
| La Inversora | 6.0 | 6.0 | no preference |
| La Jefa | 5.0 | 5.3 | Cosine — Sixty engineers on the entry plan is $1,140 a month before a single credit is topped up, and there is no headless mode to put any of it in a pipeline. |
| El Hacker | 4.0 | 5.8 | Cosine — Closed source, but it takes my own Claude or ChatGPT subscription, and the browser tools want a cdp_url in ~/.cosine/config.toml pointing at a Chrome I launched. |
Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.