Code. Automate. Flourish. Explore repositories, plan and edit code, run approved tools, coordinate bounded agent tasks, and inspect verification results through one model-independent runtime, with any model from a 4B model on your laptop to the frontier.
curl -fsSL https://raw.githubusercontent.com/FAHTHIZ/oasis/v1.1.0/scripts/install.sh | shInstallers for Windows, macOS (Apple silicon and Intel) and Linux (x86_64 and arm64). The packaged app does not require a separate Node.js or Git installation. Installers are unsigned; verify them with the published SHA-256 checksums.
For Android phones. Install Termux from its official source, then use the guided one-command setup. No native Android APK is published yet.
Install OASIS on AndroidFor a terminal-first workflow, scripts, and automation. Requires Node.js 22 or newer. Then run oasis setup, and oasis web --open for the browser workspace. With Homebrew: brew install fahthiz/oasis/oasis-cli after tapping the repository.
Steer it while it works, approve exactly what it runs, and review every change before you keep it.
Real OASIS runtime, UI and tools; model replies are scripted so the demos are reproducible (scripts/demo).
The model-free harness benchmark feeds the runtime outputs that real models actually produce and scores what happens. These are first scores on held-out cases the runtime was never tuned on.
What the model sends
No harness layerOASISMalformed or text-only tool callsFences, comments, Python literals, Hermes/Qwen tags
Edits whose old text driftedIndentation, spacing, quotes, CRLF, BOM
Destructive shell commands, approvals offEveryday commands still run
Placeholder rewrites, weakened tests, leaked secrets0 false alarms
npm run bench:harnessFull methodologyMost agent failures are harness failures. OASIS fixes each one in the runtime, so the fix holds whichever model you connect.
It said it was done, but never ran the tests.
A verification gate: the agent must review its diff and run a relevant check before it can finish. Any later edit resets it.
It ran rm -rf while I wasn't watching.
Catastrophic commands are refused even with approvals off, including when hidden in bash -c, eval or $(…). High-risk ones always ask.
It changed the test to match the broken code.
Weakened-test detection catches skips, .only, deleted tests, removed assertions and exact checks swapped for toBeDefined().
It replaced my function with “rest unchanged”.
A change guard blocks placeholder edits and flags large, unexplained deletions.
Ten patches, and the same error every time.
Fix-thrash detection tells the agent to step back on the third identical failure, and to change approach or ask you on the fifth.
I have to wait for it to finish before I can correct it.
Type while it works. The agent reads your message at its next step and keeps going. Nothing you send is dropped.
OASIS combines a terminal-first coding agent with tools and a browser workspace that help you understand, change, and verify software.
Inspect files, make a plan, apply focused edits, run approved commands, and summarize what changed across multi-step tasks.
Connect supported providers and compatible endpoints instead of being locked to one model vendor. Tool calling must be supported by your chosen model.
Use Model Context Protocol (MCP), portable Agent Skills, hooks, custom commands, layered memory, bounded subagents, and a shared dependency-aware task board.
Work from the CLI or open the local browser dashboard for projects, tasks, notes, agent chat, visible tool activity, and approval prompts.
Use command approvals, file snapshots, checkpoints, rewind, saved sessions, and cost limits to keep multi-step work easier to review.
Use headless and JSON workflows, CI scripts, the GitHub Action, and the Node.js SDK for repeatable developer automation.
OASIS does not bundle a model or provider subscription. During setup, choose a supported provider that matches your tools, budget, and preferred workflow. Small local models get a compact prompt, a 13-tool profile and tool-call repair.
Privacy note: The keyless search fallback may fetch results from DuckDuckGo. Search requests go to the search provider you use; your prompts and any code/context you include are sent to the model provider you configure. You can configure a supported search provider with your own key. Do not send secrets you do not want those providers to receive.
OASIS includes safeguards, but agent-driven code still needs human review. Understand what is enabled before using it on unfamiliar or important projects.
This page rebuilds itself on every commit to main. These are the latest changes, straight from the repository.
OASIS is Apache-2.0 licensed. Explore how it works, what it supports today, and how to contribute.
Download the app and point it at a project, or read the code first. It is free and open source, and it runs with the model you choose.