- Emacs Lisp 99.8%
- Shell 0.2%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
don't add to buffer-undo-list when replaying past session messages it gave the warning: combine-change-calls: buffer-undo-list has been truncated **Cause:** Your Doom config has advice on `comint-output-filter` (`~/.config/emacs/lisp/doom-emacs.el:1477`) that sets `buffer-undo-list` to nil after comint output. The replay runs inside `combine-change-calls`, which needs the undo list it saved at the start to still be there at the end. Doom's reset wiped it, so it printed the warning. |
||
| screenshots | ||
| .dir-locals.el | ||
| .gitattributes | ||
| .gitignore | ||
| AGENTS.md | ||
| Bugs.org | ||
| LICENSE | ||
| opencode-api.el | ||
| opencode-common.el | ||
| opencode-convert.el | ||
| opencode-diff-parser-test.el | ||
| opencode-diff-parser.el | ||
| opencode-edit-hooks.el | ||
| opencode-ert.el | ||
| opencode-flycheck.el | ||
| opencode-format-tool-calls.el | ||
| opencode-forms.el | ||
| opencode-permission.el | ||
| opencode-session-control.el | ||
| opencode-session-input.el | ||
| opencode-session-output.el | ||
| opencode-sessions-test.el | ||
| opencode-sessions.el | ||
| opencode.el | ||
| README.org | ||
| run-ert-tests.sh | ||
Emacs OpenCode UI
An emacs UI to opencode - an open source AI coding agent. This provides two sets of advantages over the official TUI:
-
A better TUI and GUI
Emacs is a mature TUI and GUI framework, that while janky in its own way, is far less janky than the TUIs the new agentic coding tools have written from scratch. This package builds on a solid foundation of comint, vtable, diff-mode, hideshow,
markdown-ts-mode, emacs' completion system, and more, to offer a (IMO) nicer UI. Also if you're an emacs user, the UI is more consistent: goto next or previous prompt, history search, and everything else works the same as in any other repl or shell based on comint mode, completion and filtering works the same as everywhere else in emacs, and everything is just a text buffer where all your usual editing and other commands work as expected. -
Emacs integration
- runs
opencode-file-edited-functionsafter edit tool calls. The default hook will reload file after edits to emacs lisp files. You can also addopencode-fix-parensto fix parentheses with parinfer ifparinfer-rust-modeis installed after edits to lisp files. -
calls
opencode-files-finished-editing-functionswith a list of all files changed in the last assistant turn, including by shell commands, according to the server's snapshot diff, so it requires a git repository and gitignored files aren't included. Addopencode-indent-fileshere to automatically fix indentation withindent-region, andopencode-flycheck-edited-filesto run flycheck and report all errors and warnings in the edited files to the LLM.⚠️ WARNING: opencode-fix-parens, opencode-indent-files, and opencode-flycheck-edited-files are all disabled by default, but if you use them they call revert-buffer, discarding changes in emacs in favor of llm edits. I use git worktrees or just stash, if I am editing at the same time as an llm.
You can set project specific hooks in
opencode-project-files-finished-editing-functions. We includeopencode-uv-pytestandopencode-emacs-ertwhich will run all tests from pytest and ERT and report errors to the LLM. They are defined usingopencode-run-command-diagnosticwhich you can use to run any shell command and report errors to the LLM. To set up a python project to run pytest after LLM edits add your project directory tosafe-local-variable-directoriesand then create a.dir-locals.elin the project directory with:((nil . ((opencode-project-files-finished-editing-functions . (opencode-uv-pytest))))) - add any emacs buffer to chat context with
opencode-add-buffer-dwim - add files from dired to context with
opencode-add-file-dwim - add selected region to context with
opencode-add-region - add function at point to context with
opencode-add-defun - start a new session to fix all flycheck errors in the current buffer with
opencode-fix-flycheck - start a new session to diagnose and fix failed tests from the ERT results
buffer with
opencode-fix-ert, bound tooinert-results-mode. By default it will diagnose all failed tests, but with a prefix argument will just diagnose the test at point. - integration with magit is possible,
opencode-new-worktreewill create a new git branch and worktree for the current project, and start an opencode session in it `file`andfile:linereferences in LLM output are turned into clickable buttons- use
dabbrev-expandin the chat window to complete long variable or function names from your code buffers - configure how your custom MCP and tools are displayed with
opencode-tool-formatters - I use with eval-elisp-mcp when developing in elisp.
- For python development integrated with opencode.el, with features like AI autodebugging exceptions, check out Swanky Python
- runs
Demo
A brief demo showing starting opencode, viewing a past session, sharing it, toggling an MCP, selecting the agent, model, and variant, performing some edits, receiving a notification when the response is finished if the session buffer isn't focused, reverting edits, and forking the session.
Installation
When this package is a bit more stable I'll package it on MELPA, for now install
from git. This package requires Emacs 31.1 with the markdown and markdown-inline
Tree-sitter grammars (install with M-x markdown-ts-mode-install-parsers), plus
magit, plz, plz-media-type, and plz-event-source.
Manual
Clone this repository and add it to your load path.
(add-to-list 'load-path "/path/to/opencode.el")
(require 'opencode)
With straight.el
(use-package opencode
:straight (opencode :type git :host codeberg :repo "sczi/opencode.el"))
Emacs use-package
(use-package opencode
:vc (:url "https://codeberg.org/sczi/opencode.el.git" :rev :newest))
In Doom Emacs
add to packages.el and run doom sync
(package! opencode
:recipe (:host codeberg :repo "sczi/opencode.el"))
Connecting
Automatic
M-x opencode will try to connect to an opencode server at
opencode-host:opencode-port or will start a headless opencode server if one
isn't already running. Then it opens a Session Management buffer for the current
project.
Manual
M-x opencode-connect will prompt for the host and port of an opencode server to
connect to.
Project Management
In opencode, sessions are grouped by project, where each git repository is a
different project. Different git worktrees for the same repository will be
grouped in the same project. Sessions started in directories outside of any git
repository will go in the global / project. opencode-select-project will fuzzy
complete on all projects and open the Session Management buffer for the project
you select. When creating a new git repository or adding a git remote, you need
to run opencode-reload-locations for the OpenCode server to pick up the changed
project.
Session Management

Here you see a table view with information on all sessions for a given project.
You can configure what columns are shown with opencode-session-control-table-columns.
With a claude subscription plan, the cost column will show an estimate in dollars at API pricing, and show the estimated percent of the 5 hour limit a session has used.
Below the table is the actual usage of your 5 hour and weekly limits, as last reported by Claude Code. Until a claude-agent session makes a request, it is read from the most recent one, so can under report if you use claude outside of opencode.
PRs are welcome to add this information for providers of other subscription plans you may have.
Shortcuts to act on the session at point:
| Key | Description |
|---|---|
x |
kill session |
R |
rename session |
RET or o |
open the session |
Buffer-wide shortcuts:
| Key | Description |
|---|---|
n |
new session |
M |
toggle MCP - fuzzy completes on all available MCPs |
v |
toggle verbose - whether to show subagent sessions, as a tree under their parent |
Chat Sessions

This is what session buffers look like. In the modeline at the bottom right we
see the current agent, current model, model variant if one is selected (like
low/medium/high thinking level), the percent of context used up so far, and an
indicator for if the session is idle or busy. User prompts are shown with a blue
margin highlight, information about tool calls in green, and reasoning in brown.
These colors can be configured through the opencode-*-margin-highlight faces.
Reasoning, tool output, and the diffs of edits are folded with hs-minor-mode,
leaving their first line visible. Click the ellipsis or use hs-toggle-hiding
(C-c @ C-c) on the first line to expand them, and hs-show-all (C-c @ C-M-s)
to expand everything.
| Key | Description |
|---|---|
C-c C-y |
yank code block at point |
C-c y |
yank-media, adds an image from clipboard to context |
C-c Y |
Copy the full conversation as a Markdown transcript, including reasoning and tool outputs. Useful to gather context in OpenCode, and then iterate on a plan in webchat if webchat has separate usage limits. |
C-c C-c |
cancel the current request |
C-c C-p |
respond to a permission request |
C-c x |
delete this session |
TAB |
cycle through available primary agents |
C-c r |
rename this session |
C-c n |
start a new session in the same project |
C-c l |
list and select from all sessions in the same project |
C-c c |
fuzzy select a child session to open - children are all subagent sessions spawned by this session |
C-c C |
compact session |
C-c t |
scroll to the top of the text part of the last assistant response |
C-c p |
open parent session - when in a subagent session, open the session that spawned it |
C-c f |
select a file to add to context, if called with a prefix argument then select a directory |
C-c b |
/btw — ask a side question about the conversation without adding to it, the answer opens in a new buffer |
C-c B |
select a buffer to add to context |
C-c m |
select model |
M-TAB |
cycle to the next reasoning level |
M-S-TAB |
cycle to the previous reasoning level |
C-c M |
select an MCP to toggle |
C-c F |
fork session from message at point - use this when a model responds badly to avoid polluting context with the bad response |
C-c R |
revert all edits done by this session after the message at point |
/ |
when at the start of a user prompt, select a slash command to run |
@ |
reference a sub-agent |
Notifications
If a session buffer and frame is not active, emacs will show a notification when that session's response has finished, so that you can go work on other stuff and be notified when the agent has finished.
Commands
| Command | Description |
|---|---|
| opencode | Open opencode sessions control buffer for the current project directory. Connects to an existing server if one is running, otherwise starts a new one if ‘opencode-auto-start-server’ is non-nil. |
| opencode-connect | Connect to opencode server, prompting for HOST and PORT. Call ON-CONNECT once connected. |
| opencode-disconnect | Disconnect from opencode server, optionally log EVENT. |
| opencode-select-project | Completing read to prompt which project to select. |
| opencode-select-open-session | Select among open session buffers. |
| opencode-new-worktree | Create a new git branch, and worktree prompting for a name. Then open an opencode session in it. |
| opencode-new-session | Create a new session. With a prefix argument it will ask for TITLE. Without it will use a default title and then automatically generate one. If CALLBACK is given, it will be called with the session after it is created. If INITIAL-PROMPT is given, it is sent after CALLBACK is called. (fn &key TITLE CALLBACK INITIAL-PROMPT) |
| opencode-select-idle | Select a session that hasn’t been visited since it went idle. |
| opencode-visit-last-idle | Open the most recent session to notify as idle. |
| opencode-add-file-dwim | If in Dired, add all marked files, or file at point if none marked. Otherwise add the current buffer’s file. Otherwise prompt for file in current project. |
| opencode-add-region | Add the active region to context. If the buffer visits a file, the region is sent with its file name and line numbers. |
| opencode-add-defun | Add the defun at point to context, as a region. If the buffer visits a file, the defun is sent with its file name and line numbers. |
| opencode-add-buffer-dwim | Add current buffer or prompt for one to add. |
| opencode-fix-flycheck | Start a new opencode session to fix the current buffer’s Flycheck errors. |
| opencode-reload-locations | Make the server resolve the project of each directory again. Run this after a directory’s project changes, such as on ‘git init’, its first commit or adding a remote. |
| opencode-unrevert-all | Unrevert all reverted messages in the current session. |
| opencode-fix-ert | Start a new opencode session to diagnose and fix ERT test failures. Uses all unexpected results in the current ERT results buffer. With a prefix argument AT-POINT, only use the test at point. |
Question Tool


OpenCode includes a question tool that models can use to ask clarifying
questions of the user. If the session buffer is active when the questions are
raised, it will immediately prompt each question with options for responses. The
options the model provides are suggestions, feel free to type your own answer.
If the session buffer or frame is not active at the time, then emacs will notify
that a question is awaiting response. You can switch to the session buffer that
most recently raised a notification with opencode-visit-last-idle, and as soon
as you visit the buffer it will prompt the questions. For single choice
questions, selecting an option will immediately submit that response. For
multiple choice questions, select the options you want and then press C-c C-c to
submit.
Other requests for input from the server, like web search consent and MCP server elicitations, are prompted the same way. Text and numbers are read in the minibuffer. Elicitations aren't tied to a session so they're prompted right away.
Security
Opencode's permission system, especially for bash, is not nearly as fine-grained
as claude's. However, I don't think this really matters. Any coding agent with
the ability to write and execute code can execute potentially malicious code by
design and should be untrusted. One approach I'd recommend is to develop inside
virtual machines or at least unprivileged containers managed by incus. If you
use btrfs or ZFS for your incus storage pool, then incus copy is almost instant.
So you can have a VM with emacs and opencode and your project, but without
secrets or sensitive data, and spawn a copy to do agent-assisted development in.
When you finish, push the changes to a git repo in a shared folder, and delete
the VM. If you only need one copy at a time, you can just snapshot and restore
rather than creating and deleting copies. Then from your somewhat more secure
and non AI-infected development environment, review and then pull the changes
from the shared folder.
opencode.el sets a random password for the opencode server we start, to guard against a history of vulnerabilities in OpenCode's server.
Performance
Make sure you are using plz-event-source version >0.1.5-pre with a major
performance fix, and set both comint-scroll-show-maximum-output and
comint-move-point-for-output to nil.
Bugs
Known bugs in other packages that affect users of opencode.el are documented in Bugs.org. For bugs in opencode.el itself, see the issue tracker.
Configuration
opencode-commandandopencode-serve-commandcan be used to configure how to run a headless opencode serveropencode-hostandopencode-portare the default host and port to use withM-x opencode-connectopencode-server-passwordconfigures how to authenticate to the opencode server- the
opencode-facesgroup configures faces used when displaying sessions opencode-toast-functionis used to show toast notificationsopencode-file-edited-functionsis a hook run after edit tool callsopencode-files-finished-editing-functionsis a hook run after the assistant response finishes, with all files changed during itopencode-project-files-finished-editing-functionsis a buffer local hook run after the assistant response finishes, that you can use for example to run your test suite and report errors back the to the LLMopencode-session-deleted-functionsis a hook run when a session is deletedopencode-tool-formatterscan be used to control how tools are displayedopencode-shown-outputslists what to show unfolded:reasoning, tool names likeshelloredit, orshell-non-zeroto show shell tool output only when the command failed. Shell commands you run yourself are always shown. Shown outputs can still be folded the same wayopencode-open-session-same-windowcontrols whether to open sessions in a new window or in the current windowopencode-session-control-table-columnsconfigures what columns are shown in the session management bufferopencode-mode-line-process,opencode-header-line-format,opencode-overflow-formatandopencode-mode-line-formatconfigure the session information shown in the mode line and header line