Wisemonkey - A dead simple CLI agent for Linux and macOS

---

Wisemonkey is a simple, open, and hackable AI agent for the Linux and macOS terminal. It connects to any service providing an OpenAI, Anthropic, or Ollama-compatible endpoint. It features session management, persistent memory management, vector store for document embedding, native and MCP tools, skills, and much more.

The sections of this document are:

• Quickstart
• Run from source
• Run in a container
• Configuration
• Soul file
• Usage and commands
• Mid-turn steering
• Global memory
• Rolling chat memory
• Extend agent

Quickstart

Wisemonkey has been tested to work on Linux and macOS.

Requirements

• Python 3.13+
• uv for dependency management

Installation

On Linux or macOS, install uv and run the agent:

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
# Run wmk
uvx wisemonkey

Wisemonkey is available on PyPI.

You can also install it with our script:

curl -fsSL https://codeberg.org/langurmonkey/wisemonkey/raw/branch/master/install.sh | sh

This installs wisemonkey to ~/.local/share/wisemonkey/repository. It adds a wmk binary to ~/.local/bin/wmk. If Docker is installed, it also adds a wmkc binary to ~/.local/bin/wmkc to run the agent in a container (see Run in a container). If you have ~/.local/bin in your $PATH, you can launch the onboarding process to configure the agent interactively:

wmk --onboard

Running

Run the agent with the default session:

# Using uvx
uvx wisemonkey
# If installed, simply do
wmk

For the rest of this document, we assume that wmk is in your path. You can substitute it with uvx wisemonkey if you use the uvx method.

If you need an API key to access the endpoint, put it in the .env file. Wisemonkey looks for the .env file in the following locations, in order:

• Current directory, ./.env
• Config directory, $XDG_CONFIG_HOME/wisemonkey/.env
• Home directory, $HOME/.env

Create the .env file with the API key:

echo "OPENAI_API_KEY=your-api-key-here" > .env
echo "ANTHROPIC_API_KEY=your-api-key-here" > .env

The agent uses python-dotenv to load .env at startup. The openai package reads OPENAI_API_KEY from the environment automatically. You can also set OPENAI_API_KEY in your shell profile. Same goes for ANTHROPIC_API_KEY.

API keys for several endpoints

Wisemonkey is usually driven through a router — OpenRouter, OpenCode, a local vLLM — and each of those wants its own key. A single global variable cannot cover that: you would have to rewrite it on every switch (into shell history), and one endpoint's key would be sent to another the moment the provider changed. So the key is resolved from an ordered, per-provider chain, first non-empty wins:

1. model.api_key_env — the environment variable name given in config.yaml. Explicit always wins.
2. WM_API_KEY_<PROVIDER> — namespaced, so several keys coexist. Providers are GENERIC, OPENAI, ANTHROPIC (WM_API_KEY_GENERIC, …). This is the one that makes router hopping painless.
3. The provider's conventional variable — OPENAI_API_KEY, ANTHROPIC_API_KEY. Existing setups keep working untouched.
4. WM_API_KEY — a generic fallback, tried last, for every key-bearing provider.

Local providers (Ollama, LM Studio) need no key.

The rule is that only step 4 is shared: WM_API_KEY_GENERIC never satisfies an Anthropic request, so switching provider cannot send the wrong credential anywhere.

Recommended setup for routers. Give each endpoint its own variable and point at it:

cat >> .env <<'EOF'
OPENROUTER_API_KEY=sk-or-...
OPENCODE_API_KEY=sk-oc-...
EOF

model:
  provider: generic
  base_url: https://openrouter.ai/api/v1
  api_key_env: OPENROUTER_API_KEY   # name only, never the key itself
  name: anthropic/claude-sonnet-4.5

Switching to OpenCode is then two lines in config.yaml, with both keys still in .env and neither moving to your shell history. If you do not care which key is which, the namespaced variables need no config at all:

WM_API_KEY_GENERIC=sk-or-...   # used by any OpenAI-compatible endpoint
WM_API_KEY_ANTHROPIC=sk-ant-...

In the agent loop, /model configures a whole endpoint in one flow: URL → provider → key variable → model (/url is an alias for the same flow). A canonical URL picks the provider for you, an unknown host asks whether it is OpenAI-compatible, and a local endpoint asks no key question at all. The key step is a list, not free text: it offers the host's conventional name first (OPENROUTER_API_KEY for openrouter.ai, OPENCODE_API_KEY for opencode.ai), marks the variables that are actually set in your environment, and keeps auto-detect as the default.

Endpoints you use regularly can be listed under model.endpoints, and the URL step then becomes a picker with "Enter a new URL..." as its last entry:

model:
  endpoints:
    - name: Local (LM Studio)
      base_url: http://127.0.0.1:1234/v1
      provider: lmstudio          # skips the provider question
    - name: OpenRouter
      base_url: https://openrouter.ai/api/v1
      provider: generic
      api_key_env: OPENROUTER_API_KEY   # the key this endpoint wants
    - https://api.openai.com/v1         # bare URL = entry with no extras

Picking an entry sets its URL, provider and key variable at once, so switching back and forth is one selection. The key list likewise shows every *_API_KEY variable found in .env or the environment, not just the presets. Two entries are always appended for anything new: "Enter a different variable name..." and "Enter a new API key", which prompts for the value and writes it to ~/.config/wisemonkey/.env (mode 600), so a new key never needs a hand-edited file or a restart.

Storing a key in config.yaml is not supported, deliberately: config.yaml gets shared and committed, .env does not. See .env.example.

If a key cannot be found, Wisemonkey names every variable it tried, e.g.

Missing API key for generic. Tried: WM_API_KEY_GENERIC, OPENAI_API_KEY,
WM_API_KEY. Set one in your environment or .env file, or point
model.api_key_env at a variable holding the key, or run /onboard.

Run from source

# Clone the repo, then build the project:
uv build
# Set an API key (any of the names in "API keys for several endpoints"):
export WM_API_KEY_GENERIC=your-api-key
# Run the agent with the default session:
uv run wmk

Run in a container

The wmkc script runs the agent inside a container, so it needs no Python or uv on the host, and the agent only ever sees the directory you launch it from. On first use it builds the image from the Dockerfile (a few minutes) and caches it; later runs start instantly.

# From the directory you want the agent to work in:
cd ~/my-project
wmkc                      # same arguments as wmk
wmkc my-project           # a named session
wmkc --version

The container runs as your user, not root, so files it writes into the workspace and into the session history come out owned by you. Three host directories are mounted:

• the current directory — the workspace; it is the only part of the host the agent can see and modify
• $XDG_CONFIG_HOME/wisemonkey — config.yaml, mcp.json and any .env already in use
• $XDG_DATA_HOME/wisemonkey — sessions, chat history, notes and vector stores, so conversations survive the container

Your HOME, USER, TERM, EDITOR/VISUAL, the XDG variables and your API keys (WM_API_KEY_*, OPENAI_API_KEY, ANTHROPIC_API_KEY) are forwarded from the environment. Keys are passed by name, never as NAME=value, so they do not show up in ps or in your shell history.

The agent runs sandboxed: no Linux capabilities, no privilege escalation, and a bounded process count.

Variable	Effect
WMKC_IMAGE	Use a different image (default wisemonkey-agent:local)
WMKC_BUILD=0	Never build; fail if the image is missing (for scripts)
WMKC_REBUILD=1	Rebuild the image even if it already exists
WISEMONKEY_HOME	Where the Dockerfile lives (set by the install.sh wrapper)

The image ships the agent, not the checkout: editing agent/*.py does not change a container that is already built. Rebuild with WMKC_REBUILD=1 wmkc to run changed code.

Tool hooks (agent.hooks) run inside the container, so a guard binary installed only on the host (e.g. ~/.local/bin/dcg) is not found there — and a hook that cannot start fails open. If you rely on a host-side guard, run wmk directly, or extend wmkc to mount the binary in.

Client/server mode also works, but a daemon started inside a container is private to that container: XDG_RUNTIME_DIR is deliberately not forwarded, so the socket goes to /tmp/wisemonkey-<user>.

Configuration

You can configure the agent interactively before the first run with wmk --onboard. On first run, the configuration file is created in $XDG_CONFIG_HOME/wisemonkey/config.yaml from the default configuration (config.yaml) in the root of this repository.

Additionally, the configuration directory holds the mcp.json (see next section), and the .updates.yml, which holds information about the last update time and status.

Model Context Protocol (MCP)

Wisemonkey also supports MCP. Use the following commands to manage the MCP integration:

• /mcp: Show the current MCP configuration
• /mcp edit: Edit the MCP configuration file (~/.config/wisemonkey/mcp.json)
• /mcp tools: List all MCP tools available. Alias: /tools mcp

MCP servers are started when the agent boots. You need to restart the agent if you add new servers.

Tool hooks (agent.hooks)

Wisemonkey can run external guard programs around every tool call — before it runs (pre_tool_use) and after (post_tool_use). The protocol is compatible with Claude Code / Codex hooks: the hook gets JSON on stdin and answers on stdout, so an already-installed tool like dcg (the destructive command guard) works with no adapter.

agent:
  hooks:
    pre_tool_use:
      - matcher: "run_command|spawn"   # regex over tool names; empty or "*" = all
        hooks:
          - type: command
            command: /usr/local/bin/dcg   # use an absolute path
            timeout: 5                   # seconds
            honor_unsafe: false          # see below
    post_tool_use: []                    # logging/audit; output is ignored

• A pre_tool_use hook may deny, ask or allow via {"hookSpecificOutput": {"permissionDecision": "...", "permissionDecisionReason": "..."}} on stdout, or block with exit code 2 and the reason on stderr. Silent stdout means "no opinion". A deny is fed back to the model as a tool error (like a declined confirmation), so the turn continues and the model can adapt; an ask turns into an ordinary wisemonkey confirmation prompt.
• Shell tools (run_command, spawn) are presented to a hook under the Claude name Bash with the command in tool_input.command — that is the name a stock dcg matches on.
• A hook that cannot start fails open with a warning (a broken install does not refuse every call). A hook that hangs past timeout or prints garbage fails closed, because it has not cleared the call; set fail_open: true on a handler to invert that.
• honor_unsafe: true makes a handler skip while agent.unsafe is set. The default (false) is that hooks always run, so agent.unsafe never silently disables a guard that did not ask to be switched off.

File references (@path)

You can attach files directly to a prompt with @-references: writing @agent/agent.py (or the backtick-quoted @\agent/agent.py``) in your message inlines the file content into the context sent to the model — no extra tool-call round-trip needed.

• Small files (up to agent.at_file_max_chars, default 8000) are inlined as fenced code blocks appended to your message.
• Large files inject only a stub (path, size, first lines) so the model knows the file exists and can read it itself with the read_file tool.
• Directories inject a shallow listing.
• Non-existent or handle-like tokens (@langur, a@b.com) are left untouched.

The typed text stays as-is on screen and in chat history — only the message sent to the model carries the attachment. Set agent.at_file_max_chars: 0 to disable the feature. Path completion works after @ too (@ag → @agent/).

Soul file

Wisemonkey supports a soul file — a markdown file that defines the agent's identity and persona. It is injected into the system prompt before the workspace instructions, so it takes precedence over AGENTS.md without replacing it.

Two soul files are considered, in order:

1. Global soul — $XDG_CONFIG_HOME/wisemonkey/SOUL.md. Applies to every session on the machine.
2. Workspace soul — SOUL.md in the working directory or any parent directory (same walk-up lookup as AGENTS.md). Applies to the project.

Both are loaded when present; the global soul comes first. If neither exists, nothing is added.

<!-- ~/.config/wisemonkey/SOUL.md -->
# Soul

You are Wisemonkey, a terse and precise engineering assistant.
You prefer concrete answers over hedging, and you never invent facts.

Configure the file names (or disable with an empty string) in config.yaml:

agent:
  soul_file: SOUL.md          # Workspace soul file name
  global_soul_file: SOUL.md   # Global soul file name in the config directory

The soul file is identity, not task instructions. Keep workspace rules in AGENTS.md and task-specific procedures in skills.

Usage and commands

Run the agent, and then you can enter your prompt. You can use the following key bindings during input:

• Enter: submit the prompt
• Alt / Shift / Ctrl + Enter, or Ctrl + j: add a new line
• Ctrl + q: quit

Newline keys and your terminal

A terminal sends the same byte for Enter, Shift+Enter and Ctrl+Enter, so the agent cannot tell them apart unless the terminal is asked to. On start-up it enables the kitty keyboard protocol (supported by kitty, Ghostty, Alacritty, WezTerm, foot and Rio, no configuration needed) and disables it again on exit, so your shell is left exactly as it was.

Your terminal	Newline keys
kitty, Ghostty, Alacritty, WezTerm, foot, Rio, iTerm2	Shift+Enter, Ctrl+Enter, Alt+Enter, Ctrl+j
Terminals sending xterm modifyOtherKeys	Shift+Enter, Alt+Enter, Ctrl+j
macOS Terminal.app	Alt+Enter (needs Use Option as Meta Key), Ctrl+j
GNOME Terminal, PuTTY, screen	Ctrl+j
tmux	as the outer terminal, plus set -g extended-keys on and set -g extended-keys-format csi-u

When a terminal ignores the protocol, Shift+Enter simply arrives as Enter and submits, so Ctrl+j is the reliable one. Nothing is printed about this: the fallback is silent.

During inference, you can cancel the turn and return to the input prompt with Ctrl + c, or type into the footer to steer the running turn (see mid-turn steering).

Questions asked mid-session

Commands like /model and /config ask you things (a URL, a key variable, a yes/no). Those questions are fully line-edited — arrows, Backspace, Ctrl+u, history — with the question itself styled like the rest of the UI. The one thing that does not work there is a modified Enter: the questions are read by readline, which cannot tell Shift+Enter from a plain Enter, so Enter submits.

Sessions

Internally, Wisemonkey uses sessions to separate different memory histories. Sessions are named by the user. By default, the agent uses the default session. You can start in a different session (either create a new one, or restore it if it exists) by passing its name as a positional argument:

# Start in a specific session named 'my-project'
wmk my-project

The default session's name is default, so the following two commands are equivalent:

# These two commands start the 'default' session
wmk
wmk default

You can also list the existing sessions with -ls:

# List sessions
wmk --ls           
Sessions:
- my-project - ~/.local/share/wisemonkey/sessions/my-project
- default - ~/.local/share/wisemonkey/sessions/default

Or from inside a session, with /sessions (which also shows message counts and marks the current one):

/sessions

Sessions can be switched without restarting, with /resume:

/resume                    # resume the most recently used other session
/resume my-project         # switch to 'my-project'

/resume rebinds the chat history, notes, user profile and vector store of the running process, and sends the previous session to disk first. An unknown name creates a new session, exactly like wmk my-project. It is not available in client/server mode: there the daemon owns the session, so start a server for the target session (wmk --server my-project) and attach to that instead.

Sessions contain:

• The input history
• Chat memory (see chat memory)
• Vector store (see document embedding)
• Notes (see session memory)
• User profile (see session memory)

The configuration file can be session-specific. When a session starts, the agent first looks for <session_dir>/config.yaml. If it exists, that file is used; otherwise, the global config <config_dir>/wisemonkey/config.yaml is used.

Sessions are matched by the directory name in the sessions location (~/.local/share/wisemonkey/sessions). You can rename a session by just renaming the directory!

Client/server mode

Wisemonkey can run as a client/server split: a server owns the session (chat memory, profile, notes, vector store, model connection), and clients are thin frontends that connect to it and render the UI.

Start a persistent daemon server with --server:

# Start a daemon server for the 'default' session
wmk --server
# Or for a named session
wmk --server my-project

The server binds a Unix domain socket at $XDG_RUNTIME_DIR/wisemonkey/$SESSION.sock (e.g. /run/user/1000/wisemonkey/default.sock) and stays running until stopped.

Any client started afterwards tries to attach to an existing daemon for its session first. If a server is up, the client runs as a thin remote client: turns, slash commands, cancellation, and memory stats all flow over the socket, while shell commands (!) run locally and their results are recorded in the server's chat history. If no server is up, the client falls back to the normal mode, where an ephemeral server is spawned for the duration of the client's lifespan.

# Terminal 1: start the daemon
wmk --server my-project
# Terminal 2: attach to it (same session name!)
wmk my-project

Clients attach by session name: a client for session2 will never connect to a daemon running session1 — it simply starts its own independent session.

This also means several clients can share a single session state through one daemon, and the session's chat history survives client restarts while the daemon keeps running.

vi mode

You can enable vi mode for the current session with the command /vi on, or permanently in the configuration.

External editor---In vi mode, exit INSERT mode (Esc), then press v to edit your prompt in an external editor (uses your $VISUAL or $EDITOR variable).

Slash commands

There are a few commands available to use in the agent loop. You can list them with /help. Also, use /[command-name] help (e.g. /config help) to show additional help for a command.

Correcting a turn

A prompt you already sent is easy to reach but hard to take back: ↑ re-sends it verbatim, and the bad exchange stays in the agent's context either way, so the model keeps building on it. Three commands deal with that:

Command	What it does
/edit	Open the last prompt in $EDITOR, and re-run it — the edited prompt replaces the old exchange, so the model never sees both versions
/retry	Re-run the last prompt unchanged. --drop also removes the old answer first, so the re-run starts from the same context the original did
/undo [n]	Drop the last n exchanges from the context entirely, without running anything

/undo drops whole exchanges — the prompt, the answer, and any tool calls and results that belonged to it — never a fragment, which would leave a tool result in the prompt with no matching call.

Slash commands are still written to history.txt, but they are filtered out of what ↑ walks through, so the commands you ran early on no longer push your real prompts out of reach.

Shell commands

Prefix an input with ! to run it directly in the shell instead of sending it to the model:

!git status
!ls -la
!docker ps

The command runs synchronously (60 s timeout by default), its stdout, stderr, and exit code are printed, and the command together with its result is appended to the session chat memory, so the model can use it as context on the next turn. Since you type the command yourself, no confirmation is requested — this is the same trust model as running it in a terminal.

Mid-turn steering

You don't have to wait for the agent to finish a turn before adding to it.

While the model is working, Wisemonkey reserves the bottom four terminal rows as a sticky footer:

──────────────────────────────
 ⇒ stealth/space-bunny  Mem…
 ⠹ Processing prompt...
 Ctrl+C: cancel turn   |   type + ↵ to steer

The footer stays pinned while the response scrolls, using the terminal's own scroll region (DECSTBM) — no output interception, so streaming, tools, and markdown rendering all behave exactly as they do otherwise.

Steering a running turn

Type in the footer's input row while the turn runs and press Enter. The line is queued and shown as ↳ queued: <text> (next tool result).

It does not have to wait for the turn to end. Wisemonkey delivers it to the model at the next tool-call seam — right after the current batch of tools finishes, just before the model is asked again. The line becomes a regular user message, so it is persisted to chat history and counts tokens like any other user message, and it is echoed on screen at the moment the model actually receives it:

⤷ Read the CSV first, then plot the two columns. [steered mid-turn]

That makes a genuine mid-turn correction rather than a new turn: the model already has its earlier work in context when your line lands.

If the turn finishes before another tool call happens, nothing is lost. The line stays queued, and it runs as an ordinary next prompt.

Details

• The footer input row is deliberately not a line editor: no history, no completion, and nothing is echoed to stdout while you type. It only draws the transient row. The stage spinner and the 💡 Thinking... indicator live on the bottom hints row, to the right of the key bindings, so they never overwrite what you are typing.
• Ctrl + c still cancels the running turn, even mid-turn. The key is deliberately left enabled in cbreak mode for exactly this reason.
• Lines you type are captured as raw keystrokes, so an accidental keypress is harmless; it just does not appear until you press Enter.
• Steering works in both modes: local, and attached to a daemon server (the client forwards the line over the socket, and the server confirms delivery before the client drops its own copy).
• Nothing is lost on cancellation — a queued line survives a cancelled turn.

Tool output

Each tool prints one line when it is activated:

↸ Activating tool:  read_file  (agent/core.py)

That is the only line a successful, quick tool produces. A "finished in 0.0s" line under every call turns a fifteen-tool turn into fifteen lines of noise, so the result line is printed only when it tells you something:

• Failures are always reported, with the error text and the duration.
• Slow tools — at least agent.tool_slow_threshold seconds (default 1 s) — print Tool X finished in Y.Ys, so the tool that made you wait is the one you can name.

agent:
  tool_slow_threshold: 1.0   # 0 = always print a result line; a huge value = errors only

This is purely about what you see. The tool's full result still goes to the model and into chat history either way.

You can turn the whole thing off with agent.steer_midturn: false in the configuration.

If your terminal renders the footer oddly, set agent.footer_debug_bytes: true and look at $XDG_STATE_HOME/wisemonkey/footer.log, which records every byte the footer writes.

Session memory

Persistent memory follows XDG Base Directory spec in ~/.local/share/wisemonkey/session/$SESSION_NAME:

• user_profile.json---User information
• notes.json---Persistent notes (added via save_note tool)
• tmp/---Scratch space for tool subprocesses (see below)

Every command the agent runs (run_command, spawn) gets TMPDIR pointed at <session_dir>/tmp, created 0700 on first use. Anything a command leaves in the temp directory---mktemp, Python's tempfile, a compiler's intermediates---therefore stays in the session instead of the shared, world-readable /tmp, and is cleaned up with the session. Writing there needs no confirmation, because the agent already owns it; a path that is literally /tmp still prompts.

That scratch directory and the working directory you launched the agent in are the whole of "the session". Every tool that would reach past them asks you first --- the shell tools (run_command, spawn) and the file tools (read_file, write_file, patch_file, list_dir, find_files, search_content) alike, so reaching an outside path is an explicit decision whichever tool names it. /unsafe turns those prompts off for unattended runs.

Lifecycle:

• Memory is loaded into the system prompt each turn
• save_note tool adds notes during a session
• save_memory tool explicitly persists memory to disk
• Memory is auto-saved when the agent exits (interactive mode)

Document embedding

Wisemonkey can embed documents into a per-session vector store, allowing the agent to search and reference their contents during conversation. Use /embed to add a document:

/embed ~/documents/research_paper.pdf
/embed ./notes.md

The agent uses the search_knowledge tool to query embedded documents when answering questions about previously indexed files. Supported formats include PDF, Markdown, and plain text. Embeddings are powered by the configured embedding model and stored in the session directory under vectordb/.

Chat history

In addition to persistent memory, the agent maintains a chat history of recent user input and assistant output pairs. This provides context that survives beyond the LLM's context window. Here is how it works:

• Each user message and assistant response is stored in memory
• Reasoning is omitted from chat memory
• Automatically compacted when exceeding the configured character limit
• The user can trigger the compaction any time with /history compact
• Chat memory is attached to the system prompt on each turn, as the last section (after identity, workspace instructions, memory, and skills) so the static prompt prefix stays stable and provider prompt caches are not invalidated on every turn
• The agent displays the last 10 exchanges, with long messages truncated
• /history prints the whole session chat history; /history n prints only the last n turns (a prompt plus its final answer, with tool calls and results left out, like every other count in the /history-* commands)
• Tool results are truncated in the formatted history unless chat_history_full_tool_results is enabled (see below)
• Large file reads can be limited with the read_file tool's max_lines parameter (like head), or read as a window with offset (a 1-based start line, like tail -n +N)
• Targeted edits use patch_file, which can match text (whitespace-tolerant, exact-match required) or patch by line range with start_line/end_line
• search_content searches file contents by literal substring (default) or by Python regular expression (regex: true), with context lines, glob filters, depth limits and a reported result cap
• The token count is live: the footer shows a per-turn Δ while the model streams, the status line shows the exact delta, provider-reported prompt tokens and (if model.pricing is configured) the turn's cost
• All color comes from one palette (agent/palette.py); the Rich theme, the prompt, the sticky footer, and other components cannot drift apart
• Long-running commands can be backgrounded with spawn, watched incrementally with poll_job, and stopped with kill_job (which signals the whole process group and returns the output printed before death)

Persistence:

• Chat history is persisted to ~/.local/share/wisemonkey/session/$SESSION_NAME/chat_history.json
• Automatically loaded on startup
• Saved after every exchange (user input or assistant response)
• Compacted history is also persisted to disk

Configuration:

agent:
  max_chat_history: 80000                     # Maximum history tokens to keep for context
  chat_history_full_tool_results: false       # Include full tool call/result content in the prompt
  chat_history_tool_result_max_chars: 500     # Max chars per tool result when the above is false

Chat history is accounted in tokens, counted exactly with tiktoken (o200k encoding) when available, falling back to a chars/4 estimate otherwise. The token accounting used for compaction matches what is actually injected into the prompt: when chat_history_full_tool_results is false, tool results are counted (and injected) truncated to chat_history_tool_result_max_chars, so a single large tool result does not trigger premature compaction.

Protocol

The client/server wire protocol (Unix domain socket, newline-delimited JSON) is specified in docs/PROTOCOL.md. It is stable enough to build third-party frontends against: attach to a session's socket, send an attach request, and drive turns via prompt while streaming events come back.

Structure

Wisemonkey is built to be modular and hackable. Here is an overview of the main parts and their mapping to the file system.

wisemonkey/
├── agent/                  # Core agent code.
│   ├── agent.py            # REPL frontend: agent loop, prompt handling, key bindings.
│   ├── client.py           # Client-side connection to a daemon server.
│   ├── server.py           # Daemon server (UDS) owning the session Core.
│   ├── ipc.py              # IPC protocol: messages, payloads, transports.
│   ├── emitter.py          # Turn event emitter: core callbacks -> IPC events.
│   ├── commands.py         # Slash commands (e.g. /embed, /quit).
│   ├── completion.py       # Smart path completion for the prompt.
│   ├── at_files.py         # @file/@dir reference expansion.
│   ├── shellcmd.py         # `!` shell command handling.
│   ├── startup.py          # Startup banner and info rendering.
│   ├── tokens.py           # Token counting (tiktoken).
│   ├── history.py          # Input history handling.
│   ├── config.py           # Configuration loading and handling.
│   ├── console.py          # Rich console output with themed formatting.
│   ├── output.py           # Output adapter protocol (REPL, IPC).
│   ├── footer.py           # Sticky footer (DECSTBM scroll region) + spinner.
│   ├── steer.py            # Mid-turn steering input (cbreak key capture).
│   ├── keys.py             # Terminal key protocol (modified Enter, kitty keys).
│   ├── core.py             # Core agent functions, like API connection and tool calls.
│   ├── mcp.py              # MCP server support.
│   ├── mdstream.py         # Streaming markdown renderer for the REPL.
│   ├── memory.py           # Session memory, paste file creation.
│   ├── router.py           # API router implementation for OpenAI, Ollama, and Anthropic.
│   ├── skills.py           # Skill loading and management.
│   ├── tools.py            # Tool definitions.
│   ├── update.py           # Update management.
│   ├── utils.py            # Utility functions.
│   └── vectorstore.py      # Vector store wrapper.
├── tools/                  # Tool implementations available to the model.
│   ├── basic.py            # Basic and example tools.
│   ├── files.py            # File read/write tools.
│   ├── memory.py            # search_knowledge tool.
│   ├── network.py          # URL fetching.
│   ├── screenshot.py       # Screen capture tool (base64 JPEG, user-confirmed).
│   ├── terminal.py         # Shell command execution.
│   └── vectorstore.py      # Vector store tool handler.
├── skills/                 # Skill definitions. Add new skills here.
│   ├── example.md
│   └── rolldice.md
├── tests/                  # Contains all `unittest` tests.
│   └── [...]
├── config.yaml             # Default config file.
├── README.md
├── wmkc                    # Run the agent in a container (wrapper around docker).
├── Dockerfile              # Container image for wmkc.
├── pyproject.toml
├── install.sh              # Installer script.
└── .env.example

Extend the agent

This agent is simple enough that it can be easily customized and extended by adding new tools, commands, and skills.

If you create a cool new tool, skill, or slash command, consider contributing it via a merge request!

Adding tools

Create a file in tools/ or use one of the existing ones. To create a tool, create a method and decorate it with @tool(name, description, params):

from agent.tools import tool

@tool(
    name="my_tool",
    description="Does something useful. Be exhaustive here, as it is what the LLM will read to know about your tool.",
    parameters={
        "type": "object",
        "properties": {
            "input": {
                "type": "string",
                "description": "The input parameter."
            }
        },
        "required": ["input"],
    },
)
def my_handler(args):
    input = args.get("input", "no input provided")
    return {"result": f"{input}"}

Tools are auto-discovered on startup.

Adding slash commands

The process is very similar to tools. You need to create your method, preferably in agent/commands.py, and decorate it with @cmd(name, description, aliases, examples, can_complete).

A slash command must return, in that order, ok:bool, msg:str, content:str, markdown:str:

1. ok: a bool indicating if the command succeeded or failed.
2. msg: an optional short status message. It is printed with OK or ERROR.
3. content: an optional str with the Python Rich-formatted content, it is printed to the output.
4. markdown: an optional str formatted in Markdown, it is printed to the output.

@cmd(
    "/my-command",
    "This is the description",
    aliases=["/mycmd"],
)
def _cmd_my_command(agent, params) -> (bool, str, str, str):
    """This command returns a message but no content"""
    return True, "This is awesome!", None, None

Decorated commands are automatically registered, and auto-completed in the input prompt.

Adding skills

Skills are loaded from the skills/ directory of the wisemonkey project, as well as the ./skills/ folder within the current working directory. Skill loading can be disabled entirely with agent.skills: false in the config — no skills directories are created or scanned, and no skills enter the system prompt.

To create a new skill, add a .md file in skills/ with YAML front matter, following the agentskills.io standard:

---
name: my-skill
description: What this skill does
---

# My skill

## When to use

...

## Steps

1. ...

The front matter name and description are parsed and shown in the skills list. The body is injected into the system prompt.