
How to move your Solana work from Claude Code or Codex CLI to Oh My Sol
Short answer
Keep your subscription and your repo. Install Oh My Sol with npx @barestate/oms doctor --install, start it, run /connect to sign in with the Claude subscription you use in Claude Code (or your ChatGPT/Codex subscription), pick a model with /model, and turn your test command into /goal <text> --check "<command>". Oh My Sol already calls the official Solana MCP and carries a reference pack built from the solana-dev skill.
- Who it’s for
- Solana developers who use Claude Code or Codex CLI today and want to try a Solana-specific harness without a new subscription or a new workflow.
- Last checked
- , against Oh My Sol 0.1.0
What carries over, and what does not?
Most of what you rely on in Claude Code or Codex CLI either comes with you or is already built into Oh My Sol (OMS).
| What you have today | In Oh My Sol 0.1.0 |
|---|---|
| Claude subscription in Claude Code | Carries over. OMS runs it through the official Claude Code CLI as claude-code/sonnet, claude-code/opus or claude-code/haiku. The login must be a Claude account, not an API key, Bedrock or Vertex |
| ChatGPT or Codex subscription | Carries over, as the openai-codex provider |
| OpenAI or Anthropic API key | Carries over. Enter it in /connect, or import OPENAI_API_KEY or ANTHROPIC_API_KEY from a private file with /connect import <path> |
| Your repository | Carries over. Each task runs in its own managed Git worktree, so your checkout is not edited in place |
| Solana Developer MCP | Built in. OMS calls the official server (solana_* tools) when web access is on. program_autofixer is not exposed |
| solana-dev skill | Built in as reviewed static reference: the system prompt carries a pack built from the Solana Foundation solana-dev 2.7.0 skill and two Helius skills |
| Claude Code hooks, built-in tools and extra MCP servers | Turned off inside OMS’s Claude chat. Every command goes through OMS approvals and its sandbox instead |
| Your test command | Becomes the check in /goal <text> --check "<command>" |
Custom model endpoints, models.json | Not supported. Custom endpoints and a models.json file are disabled |
How do I switch, step by step?
You need Node 22.19.0 or newer, Git and the Solana CLI, on macOS or Linux, or on Windows under WSL 2.
- If you use a Claude subscription, make sure Claude Code is signed in with your Claude account:
claude auth login claude auth status - Install the pinned Surfpool and run the health check, once:
Doctor treats Node, Surfpool 1.7.0 and the Solana CLI as blocking. Its model-access warning only looks for API key variables, so you can ignore it if you sign in with a subscription.npx @barestate/oms doctor --install - Start OMS. The first launch creates a starter workspace under
~/.oms/starter, and nothing is sent to a model until you ask for something:npx @barestate/oms - Connect your model and pick one:
/connect → Check local Claude Code subscription login (or the OpenAI Codex subscription, or an API key) /model → claude-code/opus - To work in your own repository from the shell, start a task there. This example uses the Claude subscription:
oms run --workspace /path/to/repo --task first --provider claude-code --model sonnet - Optionally run
/login. The OMS account is used only for hosted web search, and you can skip it.
The quickstart and connect guide cover what to do if a step fails.
How do I bring my test command over?
In Claude Code or Codex CLI, the model usually decides when it is finished. In OMS, the model can only propose that it is done; a separate checker runs your requirements against a fresh offline Surfpool chain and closes the task only if they pass. Give it the command that proves the work:
/goal <text> --check "<command>"/statusshows requirements, proof and budget, and/usageshows tokens, spend, time and caps./verifyruns the independent check without a model turn, which is useful after you edit files by hand./testruns the project tests in the sandbox, and/goal clearremoves the goal.
A run stops on a repeated identical failure, at the attempt cap (5 by default), or when the budget runs out. The defaults are 2,000,000 tokens, $5 of API spend and 30 minutes. See keeping an agent working until the tests pass.
When should I keep using Claude Code or Codex CLI?
- Non-Solana work. OMS is built only for Solana programs and clients.
- Workflows built on Claude Code hooks, subagents or extra MCP servers. Those are turned off inside OMS’s Claude chat.
- Custom model endpoints. OMS 0.1.0 disables them.
What should I know before switching?
- Use Docker mode on Linux. On macOS, commands run under Seatbelt, which always blocks keypair folders such as
~/.config/solana. On Linux, use Docker mode (--sandbox dockerorOMS_SANDBOX=docker); in host mode OMS asks you to approve each shell command the model runs. - Your keys stay with you. OMS works against a local Surfpool chain and never holds your mainnet keys. Mainnet work comes back to you as a transaction to review and sign yourself.
- What you gain. OMS beats both Codex CLI and Claude Code on the same model, on Solana Bench (377 against 273 on GPT-6.1 Sol, 303 against 252 on Claude Opus 5.5) and on fresh Solana tasks, with 42 to 62% fewer input tokens per solved task. The benchmarks overview has the charts.
Frequently asked questions
Do I need a new model subscription to use Oh My Sol?
No. Oh My Sol (OMS) does not supply a model. It runs a Claude subscription through the official Claude Code CLI (provider claude-code), a ChatGPT or Codex subscription (provider openai-codex), or an OpenAI or Anthropic API key. Your plan’s limits apply, and Oh My Sol does not fall back to an API key when a Claude plan runs out.
Do I have to uninstall Claude Code or Codex CLI?
No. To use a Claude subscription, Oh My Sol needs Claude Code installed and signed in with a Claude account. Keep using Claude Code or Codex CLI for non-Solana work; Oh My Sol is built only for Solana.
Does the Solana Developer MCP still work in Oh My Sol?
Oh My Sol calls the official Solana MCP server itself, through its solana_* tools, when web access is on for the session. You do not add it yourself. /web off turns web search, fetch and the Solana MCP off for the session. The server’s program_autofixer tool is not exposed.
Do I need to install the Solana dev skill in Oh My Sol?
No. The OMS system prompt carries a reviewed reference pack built from the Solana Foundation solana-dev 2.7.0 skill and the Helius helius 1.1.5 and svm 1.2.0 skills. Inside the OMS Claude chat, Claude Code’s own tools, hooks and extra MCP servers are turned off, so every command goes through OMS approvals and its sandbox.
How do I bring my test command over?
Make it the check: /goal <text> --check "<command>". Oh My Sol keeps the model working until that command passes, or until the same failure repeats twice, attempts reach the cap (5 by default) or the budget runs out. /verify runs the check without a model turn, and /status shows requirements, proof and budget.
Does Oh My Sol do better than Claude Code or Codex CLI on Solana?
Yes. OMS beats both. On fresh Solana tasks it passed 96% of runs against 76% for Codex CLI with GPT-6.1 Sol, 92% against 79% for Claude Code with Claude Opus 5.5, and 71% against 46% for Claude Code with Claude Sonnet 5.5. On Solana Bench it averaged 377 in the basic environment against 273 for Codex CLI on GPT-6.1 Sol, and 303 against 252 for Claude Code on Claude Opus 5.5.
Try it on your own program
Oh My Sol 0.1.0 runs on macOS and Linux, and on Windows through WSL 2. You need Node 22.19.0 or newer and Git. You bring your own model access: an API key, a ChatGPT subscription or a Claude subscription.
npx @barestate/oms doctor --install
npx @barestate/oms