A coding agent in one C file. ~800 KB, no runtime, no package manager.
Runs offline on a model you own, and on a phone.

_{site  ·  releases  ·  changelog  ·  hako (models)  ·  hako-edit  ·  hako-studio  ·  org}

_{: opens the command menu under the line  ·  / drives the model  ·  every skill is /<skill>}

_{:theme previews as you move; the default is your terminal's own colors}

_{:model lists the provider's own catalog, narrowed with /}

Local first. Auto-defaults to any installed hako model (hako-sho 3B or hako-koi 7B); it runs on your device, no key, no cloud. The agent detects it and just works.

Single binary. One C file. Builds with gcc -lpthread. Curl on PATH for HTTP. Nothing else linked.

14 cloud providers when you want them. Anthropic OAuth (Claude Pro/Max sub), GitHub Copilot, GH Models (free), OpenRouter, plus 9 more over OpenAI-compat. Set a key or run a one-line :login, your call.

_{—— I ——}

Why this exists

Every serious coding agent ships a runtime with it. Hermes pulls Python 3.11, Node, uv, ripgrep and ffmpeg; on Windows it downloads a portable Git too. Claude Code and its neighbors want Node and npm. They are good tools; they are also a language ecosystem you have to keep alive on every machine you want to think on.

hako is the same idea built the other way. One C file. It compiles with

gcc hako.c -o hako -lpthread

and the result is a ~800 KB binary (the browser UI inside) that links libc and nothing else. It talks to fourteen cloud providers when you want them, and to a 3B model on your own disk when you do not. It runs on a laptop, on a FreeBSD box, on Windows, and in a 32-bit Linux emulator on an iPhone: the same source, the same behavior.

	hako	agents that ship a runtime
install	one binary, ~800 KB	interpreter + package tree
runtime deps	libc, pthread, curl(1)	Python and/or Node, plus packages
build from source	gcc hako.c -o hako -lpthread	a package manager and a lockfile
works with no network	yes, on local weights	usually only as a client
your data leaves	only to the provider you chose	varies
license	GPL-3.0	varies

The wager is simple: an agent is mostly a loop, some JSON, and careful text handling. None of that needs a runtime. Keeping it in C is what makes it small enough to live anywhere, and being small enough to live anywhere is the whole point.

_{—— II ——}

Overview

• Local models run on the native engine: no ollama, no server. The agent starts the hako engine as a hakm --serve child that keeps the model resident and the system prompt's KV prefix warm (on disk too, so a cold first turn costs seconds, not minutes). Answers stream token by token. No in-process link, no compile flag, no daemon, no socket. Install the engine once (curl -fsSL https://mithraeums.github.io/hakm.sh | sh, or make hakm next to a checkout), hakm pull hako-sho, and hako defaults to the mithraeum provider on first launch: no key, no :login, no config. Small models speak their native Qwen <tool_call> dialect, and the harness repairs the shapes they get wrong. (ollama is also supported as a separate optional provider, :login ollama, for non-hako models.)
• 14 cloud providers when you want them. Anthropic native (SSE + OAuth via Claude Pro/Max), ChatGPT sign-in (any ChatGPT plan, free included), OpenAI function-calling, GitHub Copilot, GitHub Models (free), OpenRouter (PKCE), Nous Portal (device code), and OpenAI-compat aliases for Gemini, Groq, Cerebras, DeepSeek, Mistral, Together, Fireworks, xAI/Grok, custom. Bare :login is a menu: sign-ins first, API keys one level down in :login api; :login <name> goes straight there.
• An input bar that stays out of the way. Drawn in the normal flow of the terminal (no absolute cursor moves, no scroll region), so a resize or a phone keyboard reflows cleanly. Keep typing while the model works; lines queue and run in order. History, ^R search, kill-word, bracketed paste, ctrl+g to finish the line in $EDITOR, ctrl+o for the whole transcript in a pager, ctrl+v to paste an image (or drop its file in) for any provider that can see one.
• Your terminal's colors by default, the suite's themes when you want them. theme=terminal uses the profile's own 16 colors; :theme <name> swaps to one of the shared presets or a custom theme.<name>.<slot> theme from the rc, and bare :theme opens a picker that recolors as you move (enter keeps it, esc puts the old one back, and it fits even a phone-sized window). The pick is written to the rc, so hake and the studio see it too. Truecolor where supported, 256-color fallback otherwise.
• HAKO.md project context. Per-project <cwd>/.hako/HAKO.md is auto-loaded into the system prompt, the CLAUDE.md equivalent for hako-code. Binary-safe + size-capped (200KB).
• Trust-gated tools. read_file, list_dir, grep, glob, write_file, edit_file (str_replace), edit_lines (line range), run_shell (120s, up to 600 per call), web_fetch, read_skill, delegate (a subagent with its own context: it reads only unless you are in auto mode, cannot start another, and stops after 10 minutes). Per-call [y]/[n]/[a] approval, and writes show a diff of what they are about to change before you answer. Untrusted dir = all tools refused; :trust once per project.
• Persistent sessions + Agent Skills. Per-cwd session id, 7-day resume, append-only JSONL history. Skills follow the open standard (a folder, a SKILL.md with name and description, loaded by progressive disclosure), so a skill written for any compatible agent works here. Read from ./.hako/skills, ./.claude/skills, ./.agents/skills, ~/.hako/skills, ~/.claude/skills and ~/.agents/skills, nearest first; a skill in two places loads once. Install a whole repository of them with :skill install owner/repo.
• Drop-in in a repo you already set up. AGENTS.md or CLAUDE.md in the working directory is read as project context (after .hako/HAKO.md), and skills already sitting in .claude/skills are picked up as-is. Nothing to duplicate, nothing to migrate.
• Several models on one chat. Run hako in two panes, :join the second to the first's chat, and each keeps its own model: @name <text> asks another (@ completes the live ones), a line without @ goes to your own unless :route <name> sends it elsewhere, and each model hears what the others said, as it is said. A model writes to another the same way, with a line that starts with @name, and three models in a row hand it back to you. Memory notes say which one saved them, a hako never writes over a file another changed since it read it, and :say is for the people on the chat, not the models.
• Scriptable. hako -p "task" --output-format json prints one JSON object and exits 0 or 1; -p - takes the prompt on stdin. Enough for another tool to drive hako as a subprocess.
• Measurable. hako --bench tools/bench/tools-v1.jsonl scores tool-call reliability against any provider and prints JSON: well-formed (exactly one parseable call) and correct (right tool, right argument), including restraint tasks where calling nothing is the right answer. It runs the real parser on a real turn with execution stubbed, so the number is about hako rather than about the benchmark.
• Asks when it should. ask_user puts a real choice in front of you, rendered through the same arrow-key picker as :model and :theme. Tool approvals use it too.
• Self-update. hako --update pulls the latest GitHub release, verifies sha256, atomically replaces the binary. No reinstall, no rebuild.
• Streaming everywhere. Anthropic SSE, every OpenAI-compatible provider and the local engine stream live tokens, including through tool calls. The spinner names the tool in flight (reading main.c); anim_style= picks one of eight.
• Context that does not run out. Token accounting per provider, and at 70% of the window the middle of the conversation folds into a digest while the opening request and recent turns stay verbatim. :compact to force it, :usage to see it.
• Skills as modes. Skills load as one-line descriptions and fetch their body on demand, so installing ten costs almost nothing. :use <skill> promotes one to a mode: its full instructions go into every reply until :use none, and it survives a restart.

_{—— III ——}

Build & Run

Build

gcc hako.c -o hako -lpthread        # one-liner

make                                      # cross-OS Makefile (-O0: fast to build, the agent waits on I/O)
make RELEASE=1                            # -O2, what releases ship
make UNIVERSAL=1                          # macOS arm64 + x86_64 fat binary
make WEB=1                                # embed hako studio, for `hako --serve`
make test                                 # the gate: no model, no network, no credentials
make asan                                 # ASan + UBSan build → ./hako_asan

Deps: libc + pthread + curl(1) on PATH. No third-party libraries linked. More gates: make test-tty (a pty harness with a fake engine; python3 + pyte), make test-serve, make fuzz (the parsers under ASan), make win-check, make wasm-check.

Install

curl -fsSL https://mithraeums.github.io/hako.sh | sh       # the agent
curl -fsSL https://mithraeums.github.io/install.sh | sh    # the whole suite: agent, editor, engine

Detects OS and arch, downloads the latest release, verifies the sha256 sidecar, drops hako into /usr/local/bin (or ~/.local/bin). Where there is no release binary (iSh is 32-bit x86), it builds from source instead (apk add build-base git curl first). No package manager. No daemon. No telemetry.

macOS universal2 · arm64 + x86_64

Linux x86_64 · arm64

FreeBSD x86_64

Windows x86_64 · MinGW

iSh i686 · built from source

First turn

Local, zero config (recommended): the engine, one model, launch:

curl -fsSL https://mithraeums.github.io/hakm.sh | sh   # the engine → ~/.hako/bin/hakm
hakm pull hako-sho                                     # 3B weights, ~1.9 GB, resumable
hako                                                   # finds them and runs locally
> :trust                                               # grant tool access in this project
> list files in this directory

Or wire a cloud provider:

hako
> :login anthropic       # OAuth, uses your Claude Pro / Max subscription
> :login github-models   # free for any GitHub account
> :login gemini          # paste a free-tier API key
> :model gemini-3.8-flash

Or one-shot:

hako -p "summarize README.md"

Self-update:

hako --update           # check + atomic-replace if newer

_{—— IV ——}

Key Bindings & Commands

Normal Mode

Key	Action
← →	move cursor
↑ ↓	history prev / next
Home / End · ^A / ^E	line start / end
^U / ^K / ^W	kill to start / end / word
^R	reverse-incremental history search
^L	clear screen
Backspace · Delete	delete back / forward at cursor
^C	cancel the line, or stop the turn in flight
^D (empty)	EOF / exit
^G	finish the line in $EDITOR (editor= in ~/.hakorc)
^O	the whole transcript, expanded, in a pager (q closes)
Shift+Tab	cycle ask → auto → plan
Tab	complete commands, names and paths; with the menu open, take the selected entry
Esc	close the command menu (the line stays); stop a reply being read aloud

Bracketed paste: multi-line pastes land as one entry.

Commands (inside prompt):

: runs hako itself; / drives the model. Typing either opens a menu under the line (↑/↓ to move, Tab to take, Esc to close); the wrong prefix still works.

:help [cmd]   :q   :update   :remote [on|off|pair]   :mcp [reload]   :memory   :e   :o
:name <chat>   :color <name|off>   :sessions [clear [all]]   :resume <id>   :session [new]
:history   :clear   :undo   :edit   :todo [clear]
:agents   :join [hako or chat]   :as <name>   :route <name|off>   :say <text>   @name <text>
:theme [<name>]   :mode ask|auto|plan   :auto   :tools   :toolgate   :toolmode   :trust   :mouse   :editor   :voice [on|off]
:model [<id>]   :models [refresh]   :provider [<name>]   :pull <model>
:login [<prov>|api]   :logout [<prov>]   :accounts   :usage
:skills [reload]   :skill install <url>   :skill uninstall <name>   :use <skill|none>
/compact   /retry   /<skill> [text]      (every installed skill is a / command)

Bare :provider / :model / :theme (no argument) open an arrow-key picker; give an argument to set directly. :help lists every command from the same table the menu, the editor's Rei pane and the studio use. A command that would race a running turn (/retry, :clear, :model) runs as soon as the turn ends.

_{—— V ——}

Configuration

Auth · Trust · State

~/.hako/state                              provider, model, settings (mode 0600, no secrets)
~/.hako/credentials                        per-provider api_key / oauth tokens, sealed (XChaCha20-Poly1305, 0600)
~/.hako/credentials.key                    its key where there is no keychain (0600; on a Mac the key is in the login keychain)
~/.hako/skills/                            markdown behaviors (flat or dir-dispatcher)
~/.hako/input_history                      line editor history (last 500)
~/.hako/projects/<encoded-cwd>/state       per-project session id, turn count
~/.hako/projects/<encoded-cwd>/trust       sentinel: tools allowed in this cwd
~/.hako/projects/<encoded-cwd>/sessions/   append-only JSONL session history
~/.hakorc or ~/.config/hako/hakorc         user config ($HAKORC first)
<cwd>/.hakorc                              per-project config, read once the folder is trusted
<cwd>/.hako/HAKO.md                        per-project context (auto-loaded into system prompt)

~/.hakorc

One file for the whole suite (hako, hake, the studio): the first that exists of $HAKORC, ~/.hakorc and ~/.config/hako/hakorc, then the project's own .hakorc once the folder is trusted (it may not set ai_api_key, ai_endpoint, ai_auto_approve, editor or a *_cmd). code.<key> is hako's own and beats <key>; edit. and studio. are the others'. Unknown keys are left alone. The pickers write the one key they change, in place.

ai_provider=mithraeum        # or anthropic, openai, gemini, ollama, groq, … (:providers)
ai_model=hako-sho
theme=terminal               # the default: your terminal's own colors. code.theme=nord for hako only
theme.mine.base=nord         # a custom theme: a preset to start from,
theme.mine.accent=#b89656    # then any slot as #rrggbb, ansi0-15 or default
voice=0                      # 1 reads each finished reply aloud (voice_rate, voice_cmd)
update_check=1               # 0: never ask GitHub for a newer release
mouse=0                      # shared with hake; default off here, it takes the scroll wheel
ai_auto_approve=0            # 0 = ask before tools (shift+tab cycles ask/auto/plan)
show_splash=0
anim_style=pump              # braille dots bar pulse bounce ghost arrows blocks pump
editor=vim                   # for ctrl+g; else $EDITOR
shell_timeout=120            # run_shell deadline in seconds (max 600)
ai_max_tokens=4096
ai_stream=1
prompt_cache=1               # Anthropic prompt caching
context_window=0             # override the model's window (0 = from the catalog)

A key set here beats a stale value in ~/.hako/state; a project's own saved state still wins inside that project.

Three auth paths:

• OAuth (subscription / account-bound): :login anthropic (Claude Pro/Max), :login copilot (GitHub Copilot Pro/Business), :login github-models (free for any GitHub user), :login openrouter (PKCE, auto-issue an OR API key), :login nous (Nous Portal, device code). Inference bills against your subscription where applicable; no per-token API key needed.
• API-key paste: :login openai, :login gemini, :login groq, :login cerebras, :login deepseek, :login mistral, :login together, :login fireworks, :login xai, :login anthropic-api, :login openrouter-api, :login nous-api, :login custom. Opens provider console in your browser, prompts with input hidden, persists into the cred store.
• Local: hako's own models need no login: hakm pull hako-sho, then :provider mithraeum. Ollama works too: :login ollama (ensure ollama serve is running) or :login ollamacloud (paste an Ollama key).

Run :providers for the full grouped list with ◎ (active) and * (saved login) markers. :models lists the active provider's live catalog (its own listing endpoint, cached 24h, :models refresh refetches; installed .mlf2 weights for mithraeum), falling back to a built-in list when offline. Bare :model opens the picker, where / filters as you type. :accounts lists saved logins; :logout [<provider>] wipes one. Mid-chat :provider X swaps secrets and flattens wire-format-specific message bodies so the conversation survives.

Resolution order: HAKO_API_KEY env → <PROVIDER>_API_KEY env → ~/.hako/credentials (per provider). Legacy CLAW_* / HAKOC_* env names are read as fallback with a one-shot deprecation warning.

First run in any directory asks for trust. Untrusted = no tool access at all.

Connection Matrix

Provider	id (:provider <id>)	Auth	Cost	Wire format	Notes
Anthropic	anthropic · claude	OAuth	sub	native + SSE	Claude Pro/Max subscription; live-tested
ChatGPT	chatgpt · codex	OAuth	sub	OpenAI Responses	any ChatGPT plan, free too; device sign-in
Anthropic API	anthropic-api · claude-api	paste	$/tok	native + SSE	regular API key, separate from sub
GitHub Copilot	copilot · github-copilot	OAuth	sub	OpenAI-compat	Copilot Pro/Business; device flow
GitHub Models	github-models · ghmodels	OAuth	free	OpenAI-compat	rate-limited; any GH account
OpenRouter	openrouter	OAuth (PKCE)	$/tok	OpenAI-compat	auto-issues user-scoped key; :free tier
OpenRouter API	openrouter-api	paste	$/tok	OpenAI-compat	paste existing key
Nous Portal	nous	device code	$/tok	OpenAI-compat	many vendors' models; public catalog
Nous Portal API	nous-api	paste	$/tok	OpenAI-compat	paste a key from portal.nousresearch.com
OpenAI	openai · gpt	paste	$/tok	function-calling
Gemini	gemini · google	paste	free/$	OpenAI-compat	generous free tier on AI Studio
Groq	groq	paste	free	OpenAI-compat	fastest hosting for Llama family
Cerebras	cerebras	paste	free	OpenAI-compat	ultra-fast inference
DeepSeek	deepseek	paste	$/tok	OpenAI-compat	DeepSeek-Chat / Reasoner
Mistral	mistral	paste	$/tok	OpenAI-compat	Mistral Large / Small / Codestral
Together	together	paste	$/tok	OpenAI-compat	aggregator
Fireworks	fireworks	paste	$/tok	OpenAI-compat	aggregator
xAI	xai · grok	paste	$/tok	OpenAI-compat	Grok API at api.x.ai
Ollama (local)	ollama · local	none	local	native	needs ollama serve running
Ollama Cloud	ollamacloud · ocloud	paste	$/tok	native	ollama.com hosted
custom	custom	paste	depends	OpenAI-compat	set ai_endpoint in .hakorc

Three wire formats (native Anthropic, native Ollama, OpenAI-compat). Quickest paths: (a) Claude Pro/Max → :login anthropic. (b) ChatGPT (free or paid) → :login chatgpt. (c) Copilot Pro → :login copilot. (d) Free, no card → :login github-models, :login gemini, :login groq, or :login ollama. (e) Pay-as-you-go → any paste row.

Tool modes

Some smaller / non-tool-tuned models (Mistral 7B, Phi-4, DeepSeek-R1 distills, smaller Gemma / Llama variants) don't reliably emit OpenAI/Anthropic function-calling JSON. Toggle ReAct mode:

:toolmode prose        # model emits <tool name="X">{...}</tool> blocks in prose (aliases: react, xml)
:toolmode native       # back to native function-calling (default)

ReAct mode injects a tool schema in the system prompt and parses <tool>...</tool> blocks from the model's response. Each tool call is executed and the result is appended as <observation tool="X">...</observation> for the next turn. Works with any instruct-tuned model.

_{—— VI ——}

Skills

~/.hako/skills/
├── style.md                  flat skill, whole file injected
└── corp/                     directory skill, SKILL.md is the dispatcher
    ├── SKILL.md
    └── agents/
        ├── CEO.md
        ├── DEV.md
        └── QA.md

Directory skills inject only SKILL.md plus a <files> manifest. The agent reads inner files on demand via read_skill(skill, path): no trust gate (skills are user-installed), path-traversal blocked.

:skill install mithraeums/skills      # every skill in the repo, into ~/.hako/skills
:skills                               # what is loaded
:use brief                            # one of them, as a mode, until :use none

Browse the catalog: mithraeums/skills.

_{—— VII ——}

Change Log

Latest: v0.4.2

v0.4.2 brings back the FreeBSD build that v0.4.1 lacked. Everything else is v0.4.1:

• Shared chats that hold together. Another hako's lines show as they are said, even mid-turn. Models talk to each other with @name lines, and three in a row hand it back to you. Arrivals and leaves are said, :join takes a chat's name, and each line is logged once.
• Sealed sign-ins. ~/.hako/credentials is encrypted (XChaCha20-Poly1305), its key in the login keychain on a Mac or a 0600 key file elsewhere.
• Current models. ChatGPT sign-in works again (default gpt-6-luna), and every provider's default, list, prices and context sizes are up to date.
• A smoother screen. Every redraw is one frame: menus, the spinner and streaming no longer flicker or hop. In Rex the pane shows hako working, waiting on you or done.
• A question from the model no longer leaves a turn stuck.

Every earlier release, and the full detail, is in CHANGELOG.md.

_{—— VIII ——}

Contributing

If you share the belief that simplicity empowers creativity, feel free to contribute.

Contribution is welcome in the form of:

• Forking this repo
• Submitting a Pull Request
• Bug reports and feature requests

Please ensure your code follows the existing style.

Thank you for your attention.

This project started out of curiosity and as a branch of hako-edit, a C-based modal text editor. If you hit any issues, feel free to open an issue on GitHub. Pull requests, suggestions, or even thoughtful discussions are welcome.

_{— SEE LICENSE —  ·  GPL-3.0}

_{— deus sol invictus mithras —}