Amp

Amp is Sourcegraph's AI coding agent. It uses a server-side proxy architecture where all LLM requests route through ampcode.com, supporting multiple models behind selectable modes.

Key Features

• Multi-Model Modes: Smart (Claude Opus 4.6), Rush (Claude Haiku 4.5), Deep (GPT-5.2 Codex), and more
• Server-Side Proxy: All requests route through ampcode.com — no direct provider API keys needed
• Specialized Subagents: Uses additional models for code review, search, and research tasks
• Thinking Support: Extended thinking in Deep mode for complex problems

Getting Started

1. Get an API Key

1. Go to ampcode.com and sign in (or create an account)
2. Navigate to your account settings
3. Generate or copy your API key

2. Configure in Sesame

1. Go to Settings → Credentials (or click the key icon on Amp in the session form)
2. Expand the Amp section
3. Paste your API key
4. Click Save

3. Create a Session

1. Select Amp as the agent
2. Choose a mode from the model dropdown:
   • Smart — Default. Uses Claude Opus 4.6 for maximum capability
   • Rush — Uses Claude Haiku 4.5 for faster, cheaper tasks
   • Deep — Uses GPT-5.2 Codex with extended thinking for complex problems
   • Free — Free tier (experimental, may have limitations)
3. Enter your prompt and start the session

Modes

Amp uses modes instead of direct model selection. Each mode bundles a primary model, system prompt, and tool configuration together on the server side.

Specialized Subagents

In addition to the primary model, Amp uses specialized models for subtasks:

These subagents are managed by Amp's server — you don't need to configure them.

Authentication

Amp uses a single API key (AMP_API_KEY) for authentication. All requests are proxied through ampcode.com, so you don't need separate API keys for the underlying model providers.

System-Wide Configuration

Set the API key as an environment variable:

AMP_API_KEY=your-api-key-here

UI Configuration

Add your API key in Settings → Credentials → Amp.

Permissions

Amp runs with --dangerously-skip-permissions in Sesame. Amp's own permission system uses declarative rules rather than interactive prompts, and does not expose an API for programmatic permission responses. This means Amp has full access to read, write, and execute within the sandbox.

Session Continuity

Amp supports session continuity. When you send a follow-up message in a Sesame session, Amp resumes the previous conversation using its internal session ID via --continue. This preserves context across turns.

Network Requirements

Amp requires access to:

• ampcode.com
• *.ampcode.com

These domains are automatically allowed when using Sesame's sandbox security.

Troubleshooting

"AMP_API_KEY not configured"

Ensure you've added your API key in Settings → Credentials → Amp, or set the AMP_API_KEY environment variable.

"Amp process exited unexpectedly"

1. Check the session logs for stderr output
2. Verify your API key is valid (try logging in at ampcode.com)
3. Ensure the sandbox has network access to ampcode.com

Deep Mode Not Working

Deep mode requires an experimental flag. Sesame handles this automatically by writing amp.experimental.modes: ["deep"] to Amp's settings file before execution. If it still fails:

1. Check the session logs for errors
2. Ensure your Amp account has access to the Deep mode

Bootstrap Failed

If the Amp binary fails to download:

• Check that the sandbox has internet access
• Verify network/firewall rules allow HTTPS to storage.googleapis.com
• Check the session logs for specific download errors

Reference

• Amp Documentation
• Amp Models