July 22, 2026

It's been a little over a month since the 0.55 update, but plenty has landed since. Let's go through the highlights as of v0.63.

What's agent-shell?

agent-shell is a native Emacs mode to interact with AI agents powered by ACP (Agent Client Protocol).

Supported agents

The roster continues to grow.

This time we get two new agents supported:

Additionally:

  • Claude regains rendering thinking output, since its ACP server started omitting by default.
  • Cursor now runs on the official Cursor CLI (agent acp). Note a recent CLI is required, as the acp subcommand was only added in 2026.
  • Codex now uses @agentclientprotocol/codex-acp.
  • Cline now accepts a default model and session mode via defcustoms (#693 by @nhojb).
  • Goose no longer requires an OpenAI key by default (#711 by @hoyon).
  • Qwen now has explicit support for OpenAI-compatible keys.

Preferred agent

agent-shell-preferred-agent-config gained finer control over agent selection. Set to an agent identifier to skip the picker entirely, or wrap it to keep the picker with a preselected default:

  • (setq agent-shell-preferred-agent-config 'claude-code) always starts Claude, no prompt.
  • (setq agent-shell-preferred-agent-config '(preselect . claude-code)) keeps the picker but preselects Claude as the default.

Custom markdown renderers

The new markdown renderer introduced in 0.55 is now extensible. Third-party packages can claim and render specific markdown constructs (source blocks, inline code ranges, etc) through agent-shell-markdown-render-functions.

Thanks to Andrea Alberti for driving the first integration in agent-shell-math-renderer, which renders LaTeX math equations as SVGs.

agent-shell faces

If you've wanted more theming control over agent-shell buffers, this is now possible via agent-shell-faces.el. Additionally, check out agent-shell-markdown.el for Markdown theming.

Less chatter, more grouping

If your agent session requires lots of tool use (fairly typical), you may notice that shell output is fairly verbose or chatty. This can be quite distracting, so from now on, tool usage as well as thinking are grouped together under an "Activity" section (collapsed by default). If you're a fan of the chatty output, not to worry. Use agent-shell-activity-group-expand-by-default to expand the lot by default.

The new "Activity" section has customisable headers (via agent-shell-activity-group-header-label-function). You now have three ways of rendering the activity header:

Descriptive

Collapsed

Expanded

(setq agent-shell-activity-group-header-label-function
      #'agent-shell-activity-group-descriptive-label)

Count

Collapsed

Expanded

(setq agent-shell-activity-group-header-label-function
      #'agent-shell-activity-group-count-label)

Tally

Collapsed

Expanded

(setq agent-shell-activity-group-header-label-function
      #'agent-shell-activity-group-tally-label)

Inline image and rich content rendering

agent-shell can now render agent-supplied image content inline, whether base64-encoded, a remote URL, or a resource_link/resource content block (#676 by @melito). Audio and other binary resources now render as links that can open externally.

Per-agent MCP servers

You can now attach MCP servers per agent kind through its config using the :mcp-servers field, as requested in #593.

Diff improvements

Section titles now summarize lines added and removed.

In addition, pressing RET in a diff buffer jumps to the file location where the change would apply, while multi-diff permission requests are now supported.

Improved out-of-turn handling

Some agents emit notifications outside of the usual prompt session request/response turn. Historically, these have been dropped by agent-shell. These out-of-turn notifications should now render as expected. Please file a bug if you continue to run into issues.

New copy commands

A few new commands for grabbing hidden content out of the shell buffer:

  • M-x agent-shell-copy-as-markdown copies buffer text as markdown.
  • M-x agent-shell-copy-source-block-at-point copies the source block at point.
  • M-x agent-shell-copy-link-url-at-point copies the URL of the link at point.

Viewport tweaks

The viewport picked up a few refinements:

  • agent-shell-viewport-dismiss-on-send dismisses the viewport once you send.
  • Submitting prompts via C-u C-c C-c now enables you to continue queuing additional requests.
  • History navigation now preserves your in-progress prompt.

Session choices

The new agent-shell-session-choices-function lets you customize what options are presented when starting a new shell. This is useful if you'd like to hide some of these options:

Agent-specific notification adapters

A new notification-adapter mechanism enables agent-specific logic to preprocess notifications before handing back to agent-agnostic handling, starting with a Cursor notification adapter (#702 by @aburtsev).

Narrow to block

M-x agent-shell-narrow-to-block narrows the buffer to the most recent N blocks (defaulting to 1), handy for focusing on the latest exchange (#672 by @arthurgleckler).

Keeping the system awake while busy

To keep a long agent turn from being cut short by idle sleep, agent-shell now keeps the system awake while an agent is busy, releasing it as soon as the turn finishes (the display may still blank). It won't override a hard sleep such as closing a laptop lid. This needs the system-sleep library (Emacs 31.1+) and can be disabled via agent-shell-inhibit-system-sleep. Thanks to @shipmints for proposing the idea.

Artist support

M-x agent-shell-artist-insert-text-drawing lets you sketch simple text drawings to drop into a prompt. Check out Bending Emacs episode 14 for a demo, crafting iOS UI.

New third-party packages

The agent-shell package family keeps expanding. Recent additions:

Maintenance and housekeeping

Keeping up with the project requires daily work. Luckily, I'm getting a bit better at time management since my new 24-hour job, so I've been catching up on general project maintenance, delivering bugfixes and feature requests, merging pull requests, and general bookkeeping.

At its most backed-up period, in mid-May, agent-shell's open issues peaked at 75 while pending PRs peaked at 29.

As of today, we are down to 13 open issues and 4 open PRs.

Side note: this chart was generated using the /github-activity skill shared in my emacs-skills repo.

agent-shell needs your support

Vendor-neutral tooling matters, and there are a couple of ways to help keep agent-shell going. Some cost money, others just a click. All are appreciated ;)

agent-shell is just me, an indie dev, while the tools it competes with have well-funded teams behind them. If this project is useful to you, please consider sponsoring the project. And if your employer benefits from your agent-shell use, nudge them to chip in too, they can typically contribute at a scale individuals can't.

GitHub stars for Anthropic credit

Anthropic offers 6 months of free Claude Max 20x for qualifying open-source projects with at least 5,000+ GitHub stars. Starring agent-shell costs nothing and can save me some money, so if you don't mind a couple of clicks, the project can really use another GitHub star.

Pull requests

Thank you to all contributors for these improvements!

Liking agent-shell? Would like to see it evolve? Consider sponsoring the effort.

powered by LMNO.lol

privacy policy · terms of service