Oh My Sol (OMS) runs in your terminal with the model you already use, Codex and Claude Code included, and it beats both Codex CLI and Claude Code on Solana work.
macOS, Linux and WSL2. Pick one:
curl (recommended) — installs Node, Git, Solana CLI and Surfpool as needed:
curl -fsSL https://barestate.xyz/oms/install | shOpen a new shell, then run oms in your project folder. OMS installs privately under ~/.oms; Solana CLI uses the official Anza installer and its default location under ~/.local/share/solana. No sudo is needed except when installing Git on Linux. On macOS, finish the Command Line Tools prompt if Git is missing, then rerun. Rerun to upgrade OMS. Set OMS_HOME to change the OMS install directory, OMS_VERSION to pin a release, OMS_NO_MODIFY_PATH=1 to leave shell config alone, OMS_SKIP_SOLANA=1 to skip Solana CLI installation, or OMS_SKIP_DOCTOR=1 to defer Surfpool installation.
Homebrew — installs Node, Git and Solana CLI:
brew install barestate-ai/tap/omsnpm / npx — bring Node >=22.19.0, Git and Solana CLI:
npm install -g @barestate/oms --ignore-scripts --no-audit --no-fund
oms
# Or run without a global install:
npx @barestate/omsThe first launch installs Surfpool 1.7.0 and checks its SHA-256. oms doctor --install does the same on its own.
- Solana references. Programs, instructions and accounts, built in and offline.
- Error decoding.
/error 0xbc4names a program error from Anchor, System and SPL tables and from your own IDL. - A local chain. Surfpool runs beside every session.
/chain,/airdrop,/txand/rewinddrive it. - Toolchain repair. OMS resolves Anchor and Kit version and build mismatches.
- Helper agents.
/team <goal>splits work acrossexplore,reviewandbuildhelpers, or your own. - A checker the model cannot game. A task closes only when its pinned checks pass on a fresh local chain.
- A sandbox. On macOS every command runs in Seatbelt, with your keys and home folder out of reach.
Each model ran in OMS and in its own lab's harness. OMS scored higher on every model.
Solana Bench, mean score:
| Model | Environment | Oh My Sol | Own harness | OMS lead |
|---|---|---|---|---|
| GPT-6.1 Sol | basic | 377 | 273 (Codex CLI) | +38% |
| GPT-6.1 Sol | swap | 366 | 283 (Codex CLI) | +30% |
| Claude Opus 5.5 | basic | 303 | 252 (Claude Code) | +20% |
| Claude Opus 5.5 | swap | 258 | 230 (Claude Code) | +12% |
| Claude Sonnet 5.5 | basic | 208 | 100 (Claude Code) | +107% |
| Claude Sonnet 5.5 | swap | 176 | 93 (Claude Code) | +90% |
Fresh Solana tasks, runs passed:
| Model | Oh My Sol | Own harness | Input tokens per solved task |
|---|---|---|---|
| GPT-6.1 Sol | 96% | 76% (Codex CLI) | 42% fewer |
| GPT-6 Astra | 92% | 71% (Codex CLI) | 51% fewer |
| Claude Opus 5.5 | 92% | 79% (Claude Code) | 62% fewer |
| Claude Sonnet 5.5 | 71% | 46% (Claude Code) | 59% fewer |
Charts are in the benchmark docs. To run Solana Bench yourself, see docs/solana-bench.md.
Run oms in a project folder and chat. The agent reads and edits that folder in place, and /diff shows what changed. OMS keeps its own state under ~/.oms/folders/, so the folder stays clean, and running oms there again resumes the conversation.
Switch to the editor with /workbench when you want the files and a shell in front of you, and back with /harness.
| Access you have | Connect in OMS |
|---|---|
| ChatGPT or Codex subscription | /connect, then pick a Codex model in /model |
| Claude subscription | claude auth login once, then /connect and Claude Code in /model |
| MiMo Token Plan | /connect with your Token Plan region |
| OpenCode Go | /connect, then any OpenCode model |
| OpenAI or Anthropic API key | /connect with the key |
With a Claude subscription, OMS runs a native Claude Code chat tied to your OMS conversation and picks it up again when you come back. Every command Claude runs still goes through OMS approvals and the sandbox. docs/claude-code.md covers effort and resume.
A task pins its requirements and verifier files before work starts. The model proposes a requirement as done, and an independent checker decides. It captures the proposed source, starts a fresh offline Surfpool chain and runs the pinned checks. The model cannot edit the pinned files and never closes a requirement itself.
/goal make the escrow tests pass --check "anchor test"
/goal keeps the agent working until the check passes, feeding each failure back to the model. /verify runs the checker on its own, with no model turn.
OMS never holds mainnet keys. Mainnet work comes back to you as a transaction to review and sign.
- macOS: every command runs under a per-run Seatbelt profile. Writes go only to the worktree, a private home and temp.
~/.ssh,~/.config/solana,~/.aws, keypair files and the keychain are unreadable, and the only network destination is the local chain. - Linux: OMS asks you to approve every shell command, with extra warnings for wallet paths, recursive removal and downloads piped to a shell.
- Docker:
--sandbox dockerruns commands in a pinned, offline container for repositories you don't trust.
Agent commands start from an empty environment with a private home, so API keys such as ANTHROPIC_API_KEY and OPENAI_API_KEY never reach them. Provider credentials stay in private configuration and never enter chat or task files.
The agent looks things up through the official Solana MCP, web_fetch and web_search. Signing in with /login turns on hosted search, or you can bring an Exa, Firecrawl, Tavily, Parallel or TinyFish key. Results come back as reference text, and /web off turns all of it off.
| Command | What it does |
|---|---|
/connect, /model |
Add a provider, pick a model |
/goal <text> --check "<cmd>" |
Work until the check passes |
/verify, /status |
Run the checker, see requirements and spend |
/team <goal>, /agents |
Split work across helper agents |
/test, /run <cmd> |
Run tests or a command in the sandbox |
/diff, /usage, /session |
Changes, tokens and session |
/chain, /airdrop, /rewind, /tx |
Drive the local chain |
/error <code>, /program <name> |
Decode errors and programs |
/review |
Offline security report for your program |
/workbench, /harness |
Editor or chat |
/web on|off, /login |
Web access and hosted search |
/help, /exit |
Every command and key, and quit |
Keys: Enter sends, Shift+Enter adds a line, Esc cancels, PgUp and PgDn scroll, Ctrl+K opens the command menu. Ctrl+C cancels running work, and pressing it twice at an idle prompt quits, as do Ctrl+D on an empty prompt and Ctrl+Q. /help keys lists the rest.
npm ci --ignore-scripts
node packages/cli/src/bin.ts doctor --install
npm startnpm test runs every workspace's tests and npm run typecheck checks the types. OMS runs native TypeScript on Node with no build step. docs/quickstart.md covers the source setup and docs/release.md the release archive.
| Package | What it holds |
|---|---|
packages/cli |
The oms command and the terminal app |
packages/engine |
Task controller, sandboxed tools and verification |
packages/contract |
Task contract, digest pinning and the evaluator |
packages/core |
Tracing, budgets, model catalog and usage |
packages/sandbox |
Seatbelt and Docker command sandboxes |
packages/surfnet |
Surfpool lifecycle and rewind |
packages/review |
The offline security review |
packages/signer |
The signing boundary |
bench |
Solana Bench harness |
MIT. See LICENSE and THIRD_PARTY_NOTICES.md.
