A shell with AI superpowers
There's now a Butterfish Neovim plugin!
Butterfish is for people who work from the command line, it adds AI prompting to your shell (bash, zsh) with OpenAI. Think Github Copilot for shell.
Here's how it works: use your shell as normal, start a command with a capital letter to prompt the AI. The AI sees the shell history, so you can ask contextual questions like "Why did that command fail?".
This is a magical UX pattern -- you get high-context AI help exactly when you want it, NO COPY/PASTING.
Once you run butterfish shell you can do the following things from the command line:
- "Give me a command to do x"
- "Why did that command fail?"
- "!Run make in this directory, debug problems" (this acts as an agent)
- Autocomplete shell commands (if the AI 'verbally' suggested a command it will appear)
- "Give me a pasta recipe" (this is a ChatGPT interface so it's not just for shell stuff!)
Feedback and external contribution is very welcome! Butterfish is open source under the MIT license. We hope that you find it useful!
Many AI-enabled products obscure the prompt (instructional text) sent to the AI model, Butterfish makes it transparent and configurable.
To see the raw AI requests / responses you can run Butterfish in verbose mode (butterfish shell -v) and watch the log file (/var/tmp/butterfish.log on MacOS). For more verbosity, use -vv.
To configure the prompts you can edit ~/.config/butterfish/prompts.yaml.
Butterfish works on MacOS and Linux. You can install via Homebrew on MacOS:
brew install bakks/bakks/butterfish
butterfish shell
Is this thing working? # Type this literally into the CLIYou can also install with go install:
go install github.com/bakks/butterfish/cmd/butterfish@latest
$(go env GOPATH)/bin/butterfish shell
Is this thing working? # Type this literally into the CLIThe first invocation will prompt you to paste in an OpenAI API secret key. You can get an OpenAI key at https://platform.openai.com/account/api-keys.
The key will be written to ~/.config/butterfish/butterfish.env, which looks like:
OPENAI_TOKEN=sk-foobar
It may also be useful to alias the butterfish command to something shorter. If you add the following line to your ~/.zshrc or ~/.bashrc file then you can run it with only bf.
alias bf="butterfish"
How does this work? Shell mode wraps your shell rather than replacing it.
- You run
butterfish shelland use your existing shell as normal, this is tested with zsh and bash - You start a command with a capital letter to prompt the LLM, e.g. "How do I do..."
- You can autocomplete commands and prompt questions with
Tab - Prompts and autocomplete use local context for answers, like ChatGPT
This pattern is shockingly effective because your shell history becomes the AI chat context. For example, if you cat a file to print it out then the AI will see it. If you tried a command that failed, the AI can see the command and the error.
Shell mode defaults to using gpt-3.5-turbo for prompting, if you have access to GPT-4 you can use it with:
butterfish shell -m gpt-4> butterfish shell --help
Usage: butterfish shell
Start the Butterfish shell wrapper. This wraps your existing shell, giving
you access to LLM prompting by starting your command with a capital letter.
LLM calls include prior shell context. This is great for keeping a chat-like
terminal open, sending written prompts, debugging commands, and iterating on
past actions.
Use:
- Type a normal command, like 'ls -l' and press enter to execute it
- Start a command with a capital letter to send it to GPT, like 'How do I
recursively find local .py files?'
- Autosuggest will print command completions, press tab to fill them in
- GPT will be able to see your shell history, so you can ask contextual
questions like 'why didnt my last command work?'
- Start a command with ! to enter Goal Mode, in which GPT will act as an Agent
attempting to accomplish your goal by executing commands, for example '!Run
make in this directory and debug any problems'.
- Start a command with !! to enter Unsafe Goal Mode, in which GPT will execute
commands without confirmation. USE WITH CAUTION.
Here are special Butterfish commands:
- Help : Give hints about usage.
- Status : Show the current Butterfish configuration.
- History : Print out the history that would be sent in a GPT prompt.
If you do not have OpenAI free credits then you will need a subscription and
you will need to pay for OpenAI API use. Autosuggest will probably be the most
expensive feature. You can reduce spend by disabling shell autosuggest (-A) or
increasing the autosuggest timeout (e.g. -t 2000).
Flags:
-h, --help Show context-sensitive help.
-v, --verbose Verbose mode, prints full LLM prompts
(sometimes to log file). Use multiple times
for more verbosity, e.g. -vv.
-V, --version Print version information and exit.
-b, --bin=STRING Shell to use (e.g. /bin/zsh), defaults to
$SHELL.
-m, --prompt-model="gpt-3.5-turbo"
Model for when the user manually enters a
prompt.
-A, --autosuggest-disabled Disable autosuggest.
-a, --autosuggest-model="gpt-3.5-turbo-instruct"
Model for autosuggest
-t, --autosuggest-timeout=400 Delay after typing before autosuggest (lower
values trigger more calls and are more
expensive). In milliseconds.
-T, --newline-autosuggest-timeout=2500
Timeout for autosuggest on a fresh line, i.e.
before a command has started. Negative values
disable. In milliseconds.
-p, --no-command-prompt Don't change command prompt (shell PS1
variable). If not set, an emoji will be added
to the prompt as a reminder you're in Shell
Mode.
-l, --light-color Light color mode, appropriate for a terminal
with a white(ish) background
-H, --max-history-block-tokens=512
Maximum number of tokens of each block of
history. For example, if a command has a very
long output, it will be truncated to this
length when sending the shell's history.
If you're in Shell Mode you can start an agent to accomplish a goal by
triggering Goal Mode. Start a command with !, as in !Fix that bug. Goal
Mode will populate a command in your shell, which you can execute with Enter,
or you can edit the command, or give feedback to the agent by doing a shell
prompt (by starting a command with a capital letter). Goal Mode will exit
if it decides the goal is met or impossible, or you can manually exit with
Ctrl-C.
You can trigger Unsafe Goal Mode by starting a command with !!, which will
execute commands without confirmation, and is thus potentially dangerous.
How well does this work? Mileage will vary. Your success rate will be higher with simpler goals and more guidance about how to accomplish them.
The advantages of this feature are that the agent can see your shell history and so it has context of what you're doing manually and can take over. If a command fails the agent will tweak it and try again.
Some disadvantages are that the agent is biased towards specific versions of
commands and may have to experiment to get it right, for example the flags
for grep on MacOS are different than on most Linux implementations. The agent
isn't very effective at manipulating large text files like code files, so you
will want to be conscious of the context it needs to be successful.
Here are some goals that work well:
!Recursively list the golang files in this directory!Find the hidden files in this directory and ask me if I want to delete them. This will generally print some things and then wait for user input (provided by prompting starting with a capital letter).!Show me what process is using the most memory
Here are some goals that work sometimes:
!Run make in this dir, debug problems!Install python dependencies for this project!Create a list of the top 3 hacker news headlines, including a link. Use the pup command to parse them out of HTML
Butterfish uses OpenAI models by default, but you can instead point it to any
server with a OpenAI compatible API with the --base-url (-u) flag. For example:
butterfish prompt -u "http://localhost:5000/v1" "Is this thing working?"
This enables using Butterfish with local or remote non-OpenAI models. Notes on this feature:
- In practice using hosted models is much simpler than running your own, and Butterfish's prompts have been tuned for GPT-3.5/4, so you will probably get the best results using the default OpenAI models.
- Being OpenAI-API compatible in this case means implementing the Chat Completions endpoint with streaming results.
- Butterfish will add your token to requests to the chat completions endpoint, so be careful about accidentally leaking credentials if you don't trust the server.
- Options for running a local model with a compatible interface include LM Studio and text-generation-webui.
Shell Mode is the primary focus of Butterfish but it also includes more specific command line utilities for prompting, generating commands, summarizing text, and managing embeddings of local files.
Examples:
butterfish prompt "Write me a poem about placeholder text"
echo "Explain unix pipes to me:" | butterfish prompt
cat go.mod | butterfish prompt "Explain what this go project file contains:"> butterfish prompt --help
Usage: butterfish prompt [<prompt> ...]
Run an LLM prompt without wrapping, stream results back. This is a
straight-through call to the LLM from the command line with a given prompt.
This accepts piped input, if there is both piped input and a prompt then they
will be concatenated together (prompt first). It is recommended that you wrap
the prompt with quotes. The default GPT model is gpt-3.5-turbo.
Arguments:
[<prompt> ...] Prompt to use.
Flags:
-h, --help Show context-sensitive help.
-v, --verbose Verbose mode, prints full LLM prompts
(sometimes to log file). Use multiple times for
more verbosity, e.g. -vv.
-V, --version Print version information and exit.
-m, --model="gpt-3.5-turbo" GPT model to use for the prompt.
-n, --num-tokens=1024 Maximum number of tokens to generate.
-T, --temperature=0.7 Temperature to use for the prompt, higher
temperature indicates more freedom/randomness
when generating each token.
Use the -f flag to execute sight unseen.
butterfish gencmd -f "Find all of the go files in the current directory, recursively"
> butterfish gencmd --help
Usage: butterfish gencmd <prompt> ...
Generate a shell command from a prompt, i.e. pass in what you want, a shell
command will be generated. Accepts piped input. You can use the -f command to
execute it sight-unseen.
Arguments:
<prompt> ... Prompt describing the desired shell command.
Flags:
-h, --help Show context-sensitive help.
-v, --verbose Verbose mode, prints full LLM prompts (sometimes to log
file). Use multiple times for more verbosity, e.g. -vv.
-V, --version Print version information and exit.
-f, --force Execute the command without prompting.
If necessary, this command will split the file into chunks, summarize chunks, then produce a final summary.
butterfish summarize README.md
cat go/main.go | butterfish summarize
> butterfish summarize --help
Usage: butterfish summarize [<files> ...]
Semantically summarize a list of files (or piped input). We read in the file,
if it is short then we hand it directly to the LLM and ask for a summary. If it
is longer then we break it into chunks and ask for a list of facts from each
chunk (max 8 chunks), then concatenate facts and ask GPT for an overall summary.
Arguments:
[<files> ...] File paths to summarize.
Flags:
-h, --help Show context-sensitive help.
-v, --verbose Verbose mode, prints full LLM prompts (sometimes to
log file). Use multiple times for more verbosity,
e.g. -vv.
-V, --version Print version information and exit.
-c, --chunk-size=3600 Number of bytes to summarize at a time if the file
must be split up.
-C, --max-chunks=8 Maximum number of chunks to summarize from a specific
file.
butterfish exec 'find -nam foobar'
butterfish index .
butterfish indexsearch "compare indexed embeddings against this string"
butterfish indexquestion "inject similar indexed embeddings into a prompt"
Here's the command help:
> butterfish --help
Usage: butterfish <command>
Do useful things with LLMs from the command line, with a bent towards software
engineering.
Butterfish is a command line tool for working with LLMs. It has two modes: CLI
command mode, used to prompt LLMs, summarize files, and manage embeddings, and
Shell mode: Wraps your local shell to provide easy prompting and autocomplete.
Butterfish stores an OpenAI auth token at ~/.config/butterfish/butterfish.env
and the prompt wrappers it uses at ~/.config/butterfish/prompts.yaml. Butterfish
logs to the system temp dir, usually to /var/tmp/butterfish.log.
To print the full prompts and responses from the OpenAI API, use the --verbose
flag. Support can be found at https://github.com/bakks/butterfish.
If you do not have OpenAI free credits then you will need a subscription and you
will need to pay for OpenAI API use. If you're using Shell Mode, autosuggest
will probably be the most expensive part. You can reduce spend by disabling
shell autosuggest (-A) or increasing the autosuggest timeout (e.g. -t 2000).
See "butterfish shell --help".
v0.1.12 darwin amd64 (commit 0c115fa) (built 2023-09-27T19:12:29Z) MIT License -
Copyright (c) 2023 Peter Bakkum
Flags:
-h, --help Show context-sensitive help.
-v, --verbose Verbose mode, prints full LLM prompts (sometimes to log
file). Use multiple times for more verbosity, e.g. -vv.
-V, --version Print version information and exit.
Commands:
shell
Start the Butterfish shell wrapper. This wraps your existing shell, giving
you access to LLM prompting by starting your command with a capital letter.
LLM calls include prior shell context. This is great for keeping a chat-like
terminal open, sending written prompts, debugging commands, and iterating on
past actions.
Use:
- Type a normal command, like 'ls -l' and press enter to execute it
- Start a command with a capital letter to send it to GPT, like 'How do I
recursively find local .py files?'
- Autosuggest will print command completions, press tab to fill them in
- GPT will be able to see your shell history, so you can ask contextual
questions like 'why didnt my last command work?'
- Start a command with ! to enter Goal Mode, in which GPT will act as
an Agent attempting to accomplish your goal by executing commands,
for example '!Run make in this directory and debug any problems'.
- Start a command with !! to enter Unsafe Goal Mode, in which GPT will
execute commands without confirmation. USE WITH CAUTION.
Here are special Butterfish commands:
- Help : Give hints about usage.
- Status : Show the current Butterfish configuration.
- History : Print out the history that would be sent in a GPT prompt.
If you do not have OpenAI free credits then you will need a subscription and
you will need to pay for OpenAI API use. Autosuggest will probably be the
most expensive feature. You can reduce spend by disabling shell autosuggest
(-A) or increasing the autosuggest timeout (e.g. -t 2000).
prompt [<prompt> ...]
Run an LLM prompt without wrapping, stream results back. This is a
straight-through call to the LLM from the command line with a given prompt.
This accepts piped input, if there is both piped input and a prompt then
they will be concatenated together (prompt first). It is recommended that
you wrap the prompt with quotes. The default GPT model is gpt-3.5-turbo.
promptedit
Like the prompt command, but this opens a local file with your default
editor (set with the EDITOR env var) that will then be passed as a prompt in
the LLM call.
summarize [<files> ...]
Semantically summarize a list of files (or piped input). We read in the
file, if it is short then we hand it directly to the LLM and ask for a
summary. If it is longer then we break it into chunks and ask for a list of
facts from each chunk (max 8 chunks), then concatenate facts and ask GPT for
an overall summary.
gencmd <prompt> ...
Generate a shell command from a prompt, i.e. pass in what you want, a shell
command will be generated. Accepts piped input. You can use the -f command
to execute it sight-unseen.
exec [<command> ...]
Execute a command and try to debug problems. The command can either passed
in or in the command register (if you have run gencmd in Console Mode).
index [<paths> ...]
Recursively index the current directory using embeddings. This will
read each file, split it into chunks, embed the chunks, and write a
.butterfish_index file to each directory caching the embeddings. If you
re-run this it will skip over previously embedded files unless you force a
re-index. This implements an exponential backoff if you hit OpenAI API rate
limits.
clearindex [<paths> ...]
Clear paths from the index, both from the in-memory index (if in Console
Mode) and to delete .butterfish_index files. Defaults to loading from the
current directory but allows you to pass in paths to load.
loadindex [<paths> ...]
Load paths into the index. This is specifically for Console Mode when you
want to load a set of cached indexes into memory. Defaults to loading from
the current directory but allows you to pass in paths to load.
showindex [<paths> ...]
Show which files are present in the loaded index. You can pass in a path but
it defaults to the current directory.
indexsearch <query>
Search embedding index and return relevant file snippets. This uses the
embedding API to embed the search string, then does a brute-force cosine
similarity against every indexed chunk of text, returning those chunks and
their scores.
indexquestion <question>
Ask a question using the embeddings index. This fetches text snippets from
the index and passes them to the LLM to generate an answer, thus you need to
run the index command first.
Run "butterfish <command> --help" for more information on a command.
A goal of Butterfish is to make prompts transparent and easily editable. Butterfish will write a prompt library to ~/.config/butterfish/prompts.yaml and load this every time it runs. You can edit prompts in that file to tweak them. If you edit a prompt then set OkToReplace: false, which prevents overwriting.
> head -n 8 ~/.config/butterfish/prompts.yaml
- name: shell_system_message
prompt: 'You are an assistant that helps the user with a Unix shell. Give advice
about commands that can be run and examples but keep your answers succinct. Here
is system info about the local machine: ''{sysinfo}'''
oktoreplace: true
- name: shell_autocomplete_command
prompt: |-
You are a unix shell command autocompleter. I will give you the user's history, predict the full command they will type. You will find good suggestions in the user's history, suggest the full command.
If you want to see the exact communication between Butterfish and the OpenAI API then set the verbose flag (-v) when you run Butterfish, this will print the full prompt and response either to the terminal or to a log file.
The butterfish summarize command gives you a semantic summary of a file. For example you can run butterfish summarize ./go.mod, and it will open that file and give you an English-language summary of what's in it.
When summarize runs, it wraps the file contents in the prompt (also there's some functionality for when it won't fit, but let's ignore that). In other words, it says something like "this is a raw text file, summarize it: '{content}'". But maybe this prompt isn't working well for you, or you want it to assume more things about the file, or you want the output to be different than a completely generic summary.
In that case, you can open ~/.config/butterfish/prompts.yaml, find the prompt named summarize, and edit it. Once you edit you should set oktoreplace to false.
Let's try it - change the summarize prompt to say something like "Summarize in spanish", set oktoreplace, and then run butterfish summarize [file].
Remember that if you run Butterfish in verbose mode (with -v), you will see the prompt when you run it!
Example:
butterfish index .
butterfish indexsearch 'Lorem ipsem dolor sit amet'
butterfish indexquestion 'Lorem ipsem dolor sit amet?'
Butterfish supports creating embeddings for local files and caching them on disk. This is the strategy many projects have been using to add external context into LLM prompts.
You can build an index by running butterfish index in a specific directory. This will recursively find all non-binary files, split files into chunks, use the OpenAI embedding API to embed each chunk, and cache the embeddings in a file called .butterfish_index in each directory. You can then run butterfish indexsearch '[search text]', which will embed the search text and then search cached embeddings for the most similar chunk. You can also run butterfish indexquestion '[question]', which injects related snippets into a prompt.
You can run butterfish index again later to update the index, this will skip over files that haven't been recently changed. Running butterfish clearindex will recursively remove .butterfish_index files.
The .butterfish_index cache files are binary files written using the protobuf schema in proto/butterfish.proto. If you check out this repo you can then inspect specific index files with a command like:
protoc --decode DirectoryIndex butterfish/proto/butterfish.proto < .butterfish_index
Let's say you have a software project repository that you want to embed, we'll call this project helloworld. First we can index it:
butterfish index /path/to/helloworld
That will run recursively, and you should see output as it calculates embeddings. Once those embeddings exist, go to the directory and check if you can use them:
cd /path/to/helloworld
butterfish indexsearch "printf 'hello world'"
This will search for embeddings that match the string you hand it. Hopefully these are relevant results!
Often you want to not only do that index search, but hand the results into a GPT prompt so that you can ask a question. In that case butterfish indexquestion uses the prompt both to search the embeddings, as a prompt to GPT to ask a question.
I've been developing Butterfish on an Intel Mac, but it should work fine on ARM Macs and probably work on Linux (untested). Here is how to get set up for development on MacOS:
brew install git go protobuf protoc-gen-go protoc-gen-go-grpc
git clone https://github.com/bakks/butterfish
cd butterfish
make
./bin/butterfish prompt "Is this thing working?"