PorkiCoder is a meta harness for coding agents. Run Claude Code, Codex, and Grok Build side by side as living terminal cards — prompt, interrupt, and supervise each session without restarting the terminal underneath.
// core philosophy What a meta-harness is
Easier means faster and smarter. Intuitive means you never read the manual first.
// porkiconsult · mcp New
The bundled MCP server exposes consult_claude_code, consult_codex, and consult_grok_build. Any of those three CLIs can now act as the master and ask either of the other two for a focused second opinion without spending PorkiCoder API credits or opening another window.
Bounded by design. Every call needs your explicit approval, receives only the question and context you choose, and runs read-only in a private temporary workspace. The worker cannot inspect your repository, edit files, run shell commands, browse, call MCP tools, or silently fall back to a paid API.
These subscription-backed tools have no repository access and no paid-API fallback.
claude-fable-5 runtime with Max effortporkicoder-consult MCP server auto-registers with all three masters, with no manual MCP configuration// porkicoder research New paper
The frozen Google FLAN-T5-small download never saw our tabs. After a small extra training pass, the same 77-million-weight graph scored 7.28 versus 3.12 for Google's original weights on 500 held-out coding-agent tasks — a paired +4.17 (465 better, 9 ties, 26 worse).
Two later never-seen holdouts: continuing on a leak-checked mix reached 7.59 / 85.7% and 7.55 / 84.5% versus the previous trained file. Scores are from an identity-blind Gemini 3.5 Flash-Lite judge, not humans. English coding-session titles only; do not mix these Terminal-board numbers with the sealed Stack Overflow result.
// terminal overlord · approval log
Forgot to hit /permissions allow all with Codex before starting work? Turn on Overlord. HOGSPHERE handles supported approval prompts as they appear without killing the session, and records every approval in a visible log.
// terminal tool · render_copyable
Ask your agent to turn a command, snippet, URL, or handoff into a clean one-click card inside the terminal. No hunting through scrollback or carefully selecting terminal text.
// features
Keep the real terminal. Change how you see it, supervise the run, and bring in more agents when one answer is not enough.
Keep two agents as living cards and one as a raw CLI, side by side. Drop any pane back into its live terminal without disrupting the grid, then switch it back when you want the card again. Same PTY, no restart.
Send a prompt, interrupt the run, grow the composer for longer instructions, and paste images into supported agent cards. The input writes into the terminal session already running underneath, so avatar mode never becomes a separate chat.
Run Kimi K3 as a native Grok Build agent with tools, maximum reasoning, and long context through PorkiCoder's managed gateway. Remote inference on our own US servers keeps the provider credential out of the app and terminal environment.
Tabs follow the shell's live working directory. A private on-device model can name a submitted shell command or human agent prompt, while manual names stay authoritative.
Pair two Macs for terminal-only access. The host is off by default; an opt-in nearby preview can prefer a direct LAN path. If that direct path drops, PorkiCoder reconnects through its end-to-end encrypted cloud relay. Settings shows and can end detached sessions.
Living cards show WORKING, YOUR TURN, IDLE, SHELL, and STALLED with elapsed time, context, output, turns, and queued follow-ups.
PorkiCoder auto-registers PorkiConsult and Render Copyable with supported coding harnesses, so the same controls follow every session.
Card extraction and the 77M-parameter naming model run on your Mac. The meta-harness reads the real terminal session, keeps turn labels local, and uses managed routes only for features that need them.
Shortcuts, Overlord gates, remote terminals, tab naming, and the full PorkiConsult tool reference.
Open the docs// start here
Short, practical guides for the features that change how a coding-agent session feels. Each starts with the fastest path, then explains the controls and trust boundary.
Living terminal bot cards
Launch Claude Code, Codex, or Grok Build normally, then press ⌘P to see the same PTY as a live card. Prompt, interrupt, paste images, or return one pane to raw terminal view without restarting anything.
Read the tutorialTerminal Overlord
Type codex overlord, claude overlord, or grok overlord. HOGSPHERE handles supported approval prompts, yields to your keyboard, and records every autonomous approval.
PorkiConsult
Ask the current CLI to consult one of the other two local subscriptions. The worker sees only the question and context you approve, runs read-only, and returns advice to the session already in front of you.
Read the tutorial// blog
Research and field notes on agent orchestration, terminal workflows, and reliable multi-agent systems.
The Need for Speed in Mid-August 2026 PorkiCoder coordinates the agent layer around real terminal sessions. Living cards expose runtime state, Terminal Overlord handles supported approval prompts, PorkiConsult brings in bounded second opinions, and...
Read the full article →// download
Signed and notarized for Apple Silicon. Launch Claude Code, Codex, or Grok Build the way you already do; PorkiCoder wraps the session instead of replacing it.
Version 2.17.61 New Signed & notarized 356.5 MB
// click the rain