Ampcode Release Notes

Amp can now use 432,000 tokens of context with Claude Sonnet 4, more than 2x the previous limit imposed by the model. Your threads can keep going for longer, with more iterations, more context, more replies, and larger files, before you need to compact or start a new thread.

We've rolled it out to all Amp users in the latest CLI and editor extension. We'll ship support for 1 million tokens of context soon.

You should still start new threads for new tasks and use small threads where possible.

Amp thread with 432,000 tokens of context

Note 1: Why does it say 400k in this screenshot? Because the context window is made up of 400k input and 32k output tokens.

Note 2: It remains to be seen what those 400k tokens are made of. Quantity isn't quality, and we're not sure yet whether the model behaves the same at 360k tokens as it does at 60k.

Instead of only allowing you to configure which Bash commands are allowed to run and which aren't, Amp now gives you the ability to define which tool it's allowed to run, and with which arguments.

When Amp wants to call a tool, it first checks your list of permissions to find an entry that matches the tool and the tool call arguments. The matched entry then tells Amp to either allow the tool call, reject it, ask you for confirmation, or delegate the decision to another program. If no matching entry is found, Amp checks the built-in permission list that contains sensible defaults.

You can define these granular tool-level permissions by adding entries to the amp.permissions list in your configuration file. Here's an example:

"amp.permissions": [
  // Ask before running a command containing "git commit"
  { "tool": "Bash", "matches": { "cmd": "*git commit*" }, "action": "ask"},
  // Reject command containing "python " or "python3 "
  { "tool": "Bash", "matches": { "cmd": ["*python *", "*python3 *"] }, "action": "reject"},
  // Ask before running any MCP tool
  { "tool": "mcp__*", "action": "ask"},
  // Delegate everything else to a permission helper (must be on $PATH)
  { "tool": "*", "action": "delegate", "to": "my-permission-helper"}
]

You can read more about the new tool-level permissions in the manual.

Response times of Amp Tab, our in-editor completion engine, are now 30% faster, with up to 50% improvements during peak usage.

We worked together with Baseten to optimize our custom deployment. The new infrastructure delivers roughly 2x performance improvements by switching to TensorRT-LLM as the inference engine and implementing KV caching with speculative decoding.

This new infrastructure also includes a modified version of lookahead decoding that uses an improved n-gram candidate selection algorithm and variable-length speculations, which reduces both draft tokens and compute per iteration compared to standard implementations.

Chart showing Amp Tab latency improvements over time

The Amp CLI has a new flag: --execute.

-x for short.

That flag turns on execute mode, which allows for programmatic use of the Amp CLI.

The CLI sends the given prompt to the agent and waits until the agent is done. Then it prints the agent's last message.

$ amp -x "Give me a list of targets in this Makefile. Only the names."
git
zsh
ghostty
atuin
jj
all

$ amp -x "Rename all .markdown files to .md then print the new names."
- readme.markdown → readme.md
- ghostty.markdown → ghostty.md

You can also pipe input to it:

$ echo "what package manager do we use here? reply only with name." | amp -x
pnpm

Execute mode is automatically turned on when you redirect stdout:

$ echo "list all markdown files in here" | amp > output.txt
$ cat output.txt
README.md
CONTRIBUTING.md
YOURE_ABSOLUTELY_RIGHT.md

You can also use execute mode with threads continue:

$ amp -x "what is 2+2?"
4

$ amp threads continue -x "add 5 to that"
9

$ amp threads continue -x "now minus 3"
6

Everything is changing and everything has changed, even in something that's decades old: the terminal.

Mere months ago, running agents in your terminal was considered a curiosity.

Now it's something we all do. Some of us casually; others many, many, many times a day — in multiple windows, multiple split panes, multiple checkouts of repositories.

Today we released a new version of our Amp CLI.

It's our first step into a future of which we don't know much, but this: agents in the terminal are part of it — and we are building a CLI to unleash them.

You can use it right now:

$ npm install -g @ampcode/cli

$ amp login

$ amp

Four instances of the Amp CLI in action

There's something new in Amp's toolbox: a tool called oracle. Behind that tool is a read-only subagent powered by one of the most powerful models today: OpenAI's o3.

o3 is slightly slower than the model behind Amp's main agent, Sonnet 4. It's also slightly more expensive and less suited for day-to-day agentic coding. But it is impressively good at reviewing, at debugging, at analyzing, at figuring out what to do next.

That's why we made it available as a tool, so the main agent and the oracle can work together — hand in hand, or, well, the agentic equivalent: token by token, one writing code, the other analyzing and reviewing.

We consciously haven't pushed the oracle too hard in the system prompt, to avoid unnecessarily increasing costs for you or slowing you down. Instead, we rely on explicit prompting to get the main agent to consult the oracle.

Here are some prompts we used:

• "Use the oracle to review the last commit's changes. I want to make sure that the actual logic for when an idle or requires-user-input notification sound plays has not changed."
• "Analyze how the functions foobar and barfoo are used. Then I want you to work a lot with the oracle to figure out how we can refactor the duplication between them while keeping changes backwards compatible."
• "I have a bug in these files: ... It shows up when I run this command: ... Help me fix this bug. Use the oracle as much as possible, since it's smart." …

Today, we're launching Amp Tab, our new in-editor completion engine, designed to anticipate your next actions and reduce the time spent manually writing code.

It uses a custom model that was trained to understand what you are trying to do next, based on your recent changes, your language server's diagnostics, and what we call semantic context.

Amp Tab can suggest regular single or multi-line edits to change entire code blocks, next to your cursor or farther away, somewhere else in your current document.

We're releasing Amp Tab now as a research preview, free to use for all Amp users. Enable it by adding the following to your VS Code settings:

{
	"amp.tab.enabled": true
}