prism

Free AI that codes, beautifully.

No subscription. No credit card. No account needed.
Already pay for ChatGPT, Claude, Grok, Gemini or Copilot? Sign in and use your plan too.

Prism is a coding assistant for your terminal that runs on free AI, or on the AI plan you already pay for. It reads your code, edits files, runs commands and checks its own work. When it builds a web page, it opens it in a headless browser and clicks through it like a person (add, tick, filter, reload) before calling it done.

It runs inside Docker, so it can only touch the one folder you give it.

Try it in 2 minutes

1. Install Docker (guide), then run the install command below.
2. Open a new terminal, cd into any folder, and type prism.
3. Press any key and type: make a snake game I can play in the browser

Something broken? Tell us here.

Quick start

New to Docker? Follow the step-by-step setup guide for Linux, Mac or Windows. It takes about 10 minutes on Linux, a bit longer on Mac or Windows, and each step shows what you should see.

You only need Docker (free). Then one command. It runs install.sh, a short script you can read first:

curl -fsSL https://raw.githubusercontent.com/sjdsjadjajdjsa/prism/main/install.sh | bash

Windows: open the Ubuntu app first (see the Windows steps in docs/SETUP.md) and paste the command there. Mac or Linux: paste it into Terminal.

It checks Docker, downloads Prism to ~/.prism and adds the prism command. Then open a new terminal:

cd ~/my-project
prism

The first start takes a few minutes while Docker builds Prism; later starts take a few seconds.

New to Docker on Windows? See the Windows beginner guide.

Prefer to watch? The demo video (51s, no sound) is a real screen recording on Linux: start Prism, ask for an app, then use it in the browser. The two waits are sped up.

Tried it on a Mac? Please open an issue with how it went.

git clone https://github.com/sjdsjadjajdjsa/prism.git
cd prism
./prism install

Then just type what you want: fix the bug in app.py, make me a to-do list website, explain this code.

Free AI, no setup

A router is a small helper that picks a free AI for each step, and a key is a secret code you get free from an AI company's website, pasted in with prism setup; you don't need a key to start.

Prism works with no keys at all. Out of the box it uses free models from Kilo Gateway, BlockRun and LLM7; the router below hides most of the slowness.

Want it faster and smarter? prism setup walks you through free keys, about a minute each. None of them ask for a card.

Use the plan you already pay for

Prism works with every major AI subscription. Type /login, pick your plan with the arrow keys or the mouse, press Enter and sign in in your browser. Your plan's real models appear in /models, and /model picks one for the chat (/model auto goes back to free models). They only answer when you pick them, so free auto never spends your plan.

If the sign-in page shows a code instead of sending you back, paste the code (or the whole link) into Prism and press Enter. /login shows who you're signed in to; /login rm <name> signs out. Sign-ins are saved on your computer only (subscriptions.json, readable only by you) and renew themselves.

Pay per use instead? /key <name> <secret> adds a key for 22 services (OpenAI, Anthropic, DeepSeek, xAI, Mistral, Gemini, OpenRouter and more), and prism keys add connects any OpenAI-compatible service. Like plans, they only answer when you pick them.

These sign-ins use the same app IDs as each company's own tools, so a company can change or block them at any time, and using them may go against its terms.

How it stays fast

A small router picks the model for every step:

• Races free models. If the first one hasn't started answering in a few seconds, the next joins the race, and the first to answer wins.
• Learns which models are quick and remembers it between runs.
• Skips models that are busy or down, and tries them again later.
• Fast lane: on NVIDIA's Nemotron models, short questions skip the slow thinking step.

Benchmarks

Each row is one run of the same tasks on the same computer. A task only counts if a script confirms the result works.

Against other coding agents

Same five tasks, same free models, only the agent changes. Every agent went through the same router, three runs per task (60 runs in all).

Prism passed every run, but it is not the fastest: opencode was quicker on 4 of 5 tasks and Aider on 3. Prism runs and checks its work, which costs time. Method, versions and every run: bench/harness.

With one model pinned (no router), Prism against Aider:

Free models are shared, so the same run can be 30% faster or slower on another day. Compare rows loosely. The raw numbers are in bench/history.json, python3 bench/chart.py redraws the charts, and bench/run.sh and bench/chat.sh measure your own setup.

Skills

Skills are short playbooks Prism follows for common jobs. It picks the right one by itself. /skills new <name> creates a new one.

/skills lists them. Add your own as .prism/skills/<name>.md in a project or ~/.config/prism/skills/<name>.md. See cli/skills/README.md for the format.

Tools and plugins (MCP)

Prism can connect MCP servers to give the AI extra tools such as file access, web pages, library docs, git or GitHub. Add a .prism/mcp.json like this to connect the filesystem server:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"]
    }
  }
}

Type /mcp to see what is connected. Put the same file in ~/.config/prism/mcp.json to use a server in every project. See docs/MCP.md for more.

Computer help

prism fix

Prism normally lives in a Docker box and can't see your computer. prism fix runs it on your computer instead, so it can find out why the Wi-Fi drops, the disk is full or an app won't start. It looks around with read-only commands, explains what it finds in plain words, and asks you before every command or file change (y allow, n no). Tested on Linux; Mac should work but hasn't been tried yet.

In your browser

prism web

The same Prism, Shard and beam included, in your browser. It opens by itself; the address is printed in the terminal (usually http://127.0.0.1:4849; if that port is taken it uses the next free one up to 4899, and the terminal shows the exact address). It only listens on your own computer.

Inside the terminal app, /web moves the current chat to your browser (same history and memory). Close the tab or press enter in the terminal to come back, and everything done in the browser is there. The browser version understands the same slash commands, with the same suggestion list. Commands that need a choice (/model, /login, /theme, /resume, /mcp add, /tabs) open a list you can click or drive with the arrow keys, Enter and Esc, and it fits a phone screen too.

Project tree

/tree shows every file with what changed: M changed, ? new, ● Prism edited it, ◦ Prism read it, ★ many files use it. /tree changed or /tree <text> narrows it. In the browser, the Files button opens a live, searchable tree that updates as files change, with each file's functions and which files it imports or is imported by. prism tree --html writes the same view as a page you can open anywhere.

It respects .gitignore, and watching for changes costs nothing while nothing changes. Prism also gets a short version of the map (about 600 tokens), so easy tasks start in the right file.

Web search

Prism can search the web and read pages when it needs facts it doesn't have, like current docs, versions or an error message. It uses DuckDuckGo, so no key is needed; if TAVILY_API_KEY or BRAVE_API_KEY is set, it uses that first. It only reads public pages (never your computer or local network), and it treats what it finds as information, not instructions.

Memory

• PRISM.md: rules for your project that Prism always follows. /init writes one for you. It also reads AGENTS.md and CLAUDE.md.
• Automatic memory: after each task Prism saves what it learned about the project, so a new chat already knows how your project works. /memory lists numbered memories, /memory forget <n> deletes one (or several: forget 2 5), /memory forget all clears everything.
• Every chat is saved. /resume reopens one, prism continue picks up the last.
• Saves free usage: old file reads are trimmed and long chats are summarized, keeping your latest request word for word. /usage shows what you used today.

Commands

Meet Shard

Shard is Prism's mascot. He reacts to what the AI is doing: thinking, working, done, or waiting when every free model is busy. When a task with real steps finishes, he fires the Last Prism beam:

/beam turns the beam on or off, and /quiet hides Shard's one-line comments.

Safety and privacy

• Prism only sees the folder you start it in, and won't start in your home folder or /. The one exception is prism fix, which you start on purpose: it runs on your computer and asks before every change.
• Commands run inside the Docker box, not on your computer. Files it creates belong to you.
• Keys live in keys.conf, readable only by you and ignored by git.
• The router only listens on 127.0.0.1, so other devices on your network can't use it, and it refuses requests from web pages. Changing keys or sign-ins needs a secret (.router-token) that the AI's own commands can't read.
• Free AI services may keep what you send them. Don't paste passwords or private data.

How it's built

Networking: on first run Prism checks whether Docker's normal network can reach the internet. Some Linux firewalls block it; then Prism shares the computer's network instead (docker-compose.host.yml). The choice is saved in .network; delete it to check again.

Thanks

Prism was built with Empryo, whose live map of the code inspired Prism's project tree. It borrows ideas from OpenCode, Crush, jcode, Claude Code and Codex, and is measured against Aider and mini-swe-agent.

Its look is built on Charm's Bubble Tea and inspired by Omarchy. It runs in Docker, searches with DuckDuckGo and checks web pages in Chromium.

It is free because of the free plans from NVIDIA, Groq, Mistral, Google Gemini, OpenRouter, Kilo, LLM7 and Z.ai, and local models through Ollama.

The beam is named after the Last Prism from Terraria; the art and lines are original.

License

MIT. See LICENSE. What changed in each version: CHANGELOG.md.