Mithraeum builds quiet software for people who write code by hand. Three instruments under one suite — a modal editor (hake), a terminal agent (hako), and the local models that run them (hakm) — on a from-scratch C inference engine, no llama.cpp, no ollama. Bring your own keys, or run our hako-sho-stock weights on-device — no cloud, no telemetry, no daemon.
Your text, your keys, your weights. No telemetry. No silent network. The cursor is a private place.
One file. One C source. No JavaScript runtime to swallow your editor. The tool fits in a head.
Any model. Any provider. Any prompt. The agent is a peer at your terminal, not a stranger in the cloud.
A single-file modal text editor written in C99. Vim-bound, language-aware, multi-pane. Embeds the hako agent in a split pane so you can talk to a model without leaving your file.
— Modal editing · 40+ syntax modes · 17 themes · OS-aware clipboard fidelity · Zero runtime dependencies beyond libc + curl.
A standalone terminal AI agent in a single C file. Auto-defaults to any installed hako model (hako-sho-stock 3B or hako-koi-mini-stock 7B), no key, no cloud. Adds 13 cloud providers (Anthropic OAuth, Copilot, GH Models, OpenRouter, OpenAI-compat) when you want them.
— Persistent sessions · skills as markdown · trust-gated tools · prose tool mode for small models · sha-verified self-update.
Local-first model family. sho (mini · 3B · Qwen2.5-Coder + hako wrap) and koi-mini (mid-small · 7B · Qwen2.5-Coder + hako wrap) live today, tool-aware out of the box. koi (mid · 14B/32B fine-tune) queued. samurai (max · 50B+ desktop) reserved.
— No cloud inference. ChatML-compatible. Honest framing: stock is stock, fine-tunes name their base + corpus. Native C engine — own loader, Q4_K/Q6_K + int8 kernels, BPE. No llama.cpp, no ggml, no ollama. Runs in-process inside the agent.
Each installer fetches the latest signed release for your OS and arch, verifies the sha256 sidecar, drops the binary into ~/.local/bin (or /usr/local/bin if writable). No package manager, no daemon, no telemetry. Same script shape across the suite — different binary at the end.
hako —curl -fsSL https://mithraeums.github.io/hako.sh | sh
hake —curl -fsSL https://mithraeums.github.io/hake.sh | sh
hakm —git clone https://github.com/mithraeums/hako
cd hako/engine && make # libc + libm + pthread only — no deps
Convert a Qwen GGUF once with tools/gguf2mlf.py, drop the .mlf2 in ~/.hako/models/, launch hako. The agent runs it in-process via the linked engine — no key, no :login, no ollama, no cloud round-trip.
Anthropic OAuth (Claude Pro/Max sub), GitHub Copilot, GH Models (free), OpenRouter PKCE, plus 9 more over OpenAI-compat. Single :login <name> walks you through.
Persistent sessions, JSONL history, trust-gated tools. Skills are markdown — drop a folder, the agent loads it. Notes you keep on disk where you can read them.
hako --update pulls the latest release, verifies the sha, atomically replaces the binary. No reinstall, no rebuild.
Each skill is a folder in mithraeums/skills containing a SKILL.md and any helpers. hako-code loads them on startup. Open a PR to add yours — the rite is the README.
Extracts text and structure from a PDF on disk. Handed to hako-code so the agent can answer questions about a document without re-uploading it.
Open a PR against mithraeums/skills. Folder name becomes the invocation.
Real mithraeum provider (distinct from ollama), force-prose tool mode for koi-class models, auto-defaults to local hako-sho-stock when present. Fixed aiExtractStringValue \uXXXX handling so the prose tool parser actually sees <tool> blocks. Renames: hakoCLAW → hako-code, binary hakoc → hako, ~/.hakoc/ → ~/.hako/ with auto-migrator.
CLI rename hako → hake, source hako.c → hake.c, macros HAKO_* → HAKE_*, splash gains EDIT subtitle. OS-aware clipboard transport (pbcopy / wl-copy / xclip / clip) — fixes tabs+indents mangled on yank-out → external paste. Subprocess wire updated to launch the renamed hako agent.
From-scratch runtime runs the real Qwen2.5-Coder weights end to end: own GGUF→MLF2 loader, own Q4_K/Q6_K dequant + matmul (int8-activation × int4-weight fast path, AVX2), own BPE tokenizer — libc + libm + pthread only. Links into the agent as libhakm.a and runs in-process (no subprocess, no server, no port). hako-sho-stock (3B) + hako-koi-mini-stock (7B) convert + run; ~2.1× from the int8 dot. We skipped the "vendor llama.cpp first" step and built the whole stack.
First skill landed. Defines the SKILL.md schema hako-code loads against. PRs open for community contributions.