Anthracode

~/terminal

npm i -g anthracode-ai

what's different

feature	Anthracode	OpenCode	Claude Code
4-tier persistent memory (session / espace / account / skill)	yes	no	single tier
Background process manager (dev servers, watchers)	yes — first class	no	no
Native PTY sessions (Win ConPTY, real shells, live output)	yes — pty_start/exec	no	Bash tool only
Ask · Plan · Build agent modes + /stats dashboard	yes — Tab to switch	build · plan	no modes
Skills hub (install plugins from any GitHub repo)	yes	no	marketplace only
Git-init and git safety warnings from the TUI	yes — auto-warn + undo	manual	manual
LSP diagnostics + Exa web search (on by default)	yes — both default	LSP only	opt-in
ANTHRACODE_* env-var overrides (model, key, paths)	yes	partial	partial
Native Windows install (no WSL, no Docker)	yes — npm i -g	WSL recommended	WSL recommended
TUI runs anywhere (no IDE)	yes	yes	yes
Open source	MIT	MIT	proprietary
Bring your own model	any provider	any provider	Anthropic only
Cloud lock-in	none	none	API required

make it yours

Pick a personality. The mockup updates live. Save your favourite as your editor theme later.

accent

single sharp colour

density

vertical rhythm

motif

surface decoration

caret

prompt glyph

state amber / comfortable / brackets / ▍

~/anthracode

anthracode init

→memory dir createdok

→config writtenok

refactor the auth middleware

plan3 steps · est 6 min

read auth.ts and 2 dependents
extract verifyToken into lib/jwt.ts
update callers, run tests

browser companion

chrome extension

new

Anthra Selector

DevTools panel that lets the agent click any element on the page and get compact React context — source file, owner component, props, DOM path.

anthra-selector.vercel.appopen

features in motion

Eight things Anthracode does that most agents don't. Each card shows the prompt to type — the GIF shows what happens.

loading…

/public/features/background-process-manager.mp4

live demo

Background process manager — in motion.

Watch the agent spin up a dev server, track its PID and port, hand the URL back, and keep it alive after the prompt returns. Same workflow you'd do across three terminals and a sticky note, collapsed into one TUI lane. Kill it, restart it, or list everything still running with a single slash command.

›Run npx serve . in the background and tell me the port

loading…

/public/features/memory.mp4

4-tier memory

Knows what you decided last week.

Four scopes — session, espace, account, skill — each a plain markdown file you can read, diff, or git-version. Tell it once that you prefer pnpm and TypeScript and it never forgets, even three projects later. The thing Claude Code can't do without you re-priming the agent every conversation.

›Remember that I prefer TypeScript over JavaScript and always use pnpm

loading…

/public/features/skills.mp4

skills hub

Extensible by drop-in.

Install a skill from the marketplace or any GitHub repo and it shows up as a slash command in the TUI. Each skill bundles a prompt, a set of helpers, and the permissions it needs — version-pinned, auditable, and trivially uninstallable. Caveman, graphify, security-best-practices and more, all one command away.

›claude plugin install caveman@caveman

media · drop at

/public/features/git-init.mp4

git from the TUI

/git-init without leaving the prompt.

Drop into an empty folder, type /git-init, and the agent confirms before initialising. Stays in the TUI for the whole flow — no second terminal, no copy-pasting commands. Pairs with the safety warning below so you never lose un-versioned work to an over-eager edit.

›/git-init

media · drop at

/public/features/git-warning.mp4

git safety

Warns before it touches anything irreversible.

Open Anthracode in a folder with no .git and the sidebar lights up immediately. Every destructive action prompts an undo block first — diff preview, revert button, plain-language explanation of what would change. The agent gets out of the way when you know what you're doing and stops you cold when you don't.

›Why am I seeing a git warning?

media · drop at

/public/features/lsp-web.mp4

LSP + web search

Type-aware. Internet-aware. Both on by default.

TypeScript, Go, Rust, Python — LSP diagnostics flow straight into the agent's context, so it sees the same red squiggles you do. Exa-powered web search lights up automatically for questions about library versions, error messages, or anything happening in the world. No flags to flip, no extra config.

›Search the web for the latest Bun release notes and summarize them

media · drop at

/public/features/env-vars.mp4

ANTHRACODE_* env

Override anything per-invocation.

Swap models, API keys, memory paths, or temperature without touching a config file. Perfect for one-off experiments, CI pipelines, or A/B-ing two providers on the same prompt. Every flag a CLI argument supports has an env-var equivalent, and they compose with whatever you already export in your shell.

›ANTHRACODE_MODEL=claude-opus-4-6 anthracode

media · drop at

/public/features/windows-native.mp4

Windows native

PowerShell. No WSL required.

Installs and runs on Windows with one npm command. No Linux subsystem, no Docker Desktop, no virtualisation tax. The TUI renders correctly in PowerShell and Windows Terminal, paths use the native separator, and process management works through Win32 APIs — not a Linux shim pretending to be Windows.

›npm install -g anthracode

faq

The questions people ask before they install. Long-form answers live in the docs.

What is Anthracode?
A hybrid AI coding agent that runs in your terminal. Takes the openness of OpenCode and pairs it with the memory + task model that makes Claude Code feel sticky. MIT licensed, no cloud account required.
How is it different from Claude Code or OpenCode?
Claude Code is closed-source and tied to Anthropic. OpenCode is open but stateless between sessions. Anthracode is open AND keeps a 4-tier memory (session · espace · account · skill), plus a real background process manager and a skills hub.
Which models can I use?
Anthropic, OpenAI, and any OpenAI-compatible endpoint — local LLMs via Ollama, vLLM, LM Studio, plus OpenRouter, Together, Groq. Bring your own key. Anthracode never proxies your prompts.
Where does memory live?
Locally. .anthracode/memory/ in each project for session and espace tiers, ~/.anthracode/ for account and skill tiers. Plain markdown files you can read, diff, and git-version. Nothing leaves your machine unless you enable Pro cloud sync.
Does it work on Windows?
Yes — native Win32 binary. Runs in PowerShell and Windows Terminal without WSL, Docker, or any Linux subsystem. Paths use the native separator.
Is it really free?
Yes. MIT licensed. Every TUI feature is free forever. Anthra Pro ($9.99/mo) only adds cloud sync, hosted background processes, priority model queues, a private skills marketplace, and workspace seats.