Migrate To Sarvam

Migrate Claude Code / Codex to Sarvam Code

Bring another agent tool's accumulated setup into Sarvam Code: skills, custom commands, MCP servers, durable memories, and global instructions. The script does the mechanical work; you make the judgment calls and verify.

What migrates, and how

Workflow

1. Survey first. List what exists before proposing anything:

   ls ~/.claude/skills/ ~/.codex/skills/ ~/.agents/skills/ 2>/dev/null
   ls ~/.claude/commands/ ~/.codex/commands/ ~/.codex/prompts/ 2>/dev/null
   python3 -c "import json;d=json.load(open('$HOME/.claude/settings.json'));print(list(d.get('mcpServers',{})))" 2>/dev/null
   grep -A2 '^\[mcp_servers' ~/.codex/config.toml 2>/dev/null
   du -sh ~/.codex/memories/* 2>/dev/null
   

2. Run the script dry (it is dry-run by default):

   python3 <skill_dir>/scripts/migrate.py
   

   Read the plan line by line. Every skip line is a judgment the script made — sanity-check them.
3. Confirm scope with the user before --apply. The two decisions that matter: skills symlinked vs copied (--copy-mode), and how much memory to bring over (the script migrates the curated store; machine-generated extension stores like chronicle/skysight are deliberately excluded).
4. Apply:

   python3 <skill_dir>/scripts/migrate.py --apply
   

5. Verify (do not skip any):
   • TOML parses: python3 -c "import tomllib;tomllib.load(open('$HOME/.sarvam/config.toml','rb'))"
   • Every touched SKILL.md passes the validator: python3 ~/.sarvam/skills/.system/self-knowledge/scripts/quick_validate.py <skill_dir>
   • Memory parity: diff -rq source vs ~/.sarvam/memories/user/ reports no content differences
   • memory_search finds a known memory under the new paths
   • Hard rules present: grep -c '^- \*\*' ~/AGENTS.md
6. Tell the user to restart Sarvam Code and check /skills and /mcp.

Judgment calls the script encodes (verify, don't blindly trust)

• Thin prompt wrappers are not converted. A Codex prompt that only says "use the $X skill" is skipped when skill X already exists — the skill already carries the behavior. If the wrapper adds real behavior, convert it by hand.
• App-specific MCP servers are skipped. Anything whose command points at ChatGPT/Codex app binaries (node_repl, computer-use, *.app paths) will not work in Sarvam. Name-based and path-based skip lists are in the script; extend them if new cases appear.
• Secrets are never read, copied, or printed. The script never opens secrets.json, .credentials.json, auth.json, or keyring files. MCP env values are written to the target config (that is where servers read them from) but are always redacted in plan output. Never echo them into chat.
• Machine-generated memory extensions are excluded. Stores like chronicle/ or skysight/ (hundreds of observed-screen logs) would pollute memory search. Only curated stores migrate. If the user wants everything, copy by hand and say so.
• Existing Sarvam config is never clobbered. MCP servers already present in ~/.sarvam/config.toml are kept; new ones are appended. Existing skills with the same name are kept. Every overwrite of an existing file takes a timestamped .bak-<stamp> backup first.

Safety rules

• Dry-run is the default. Never run --apply without showing the user the plan.
• Never print secret values from any source file — not in chat, not in logs, not in the plan output.
• Never modify files under ~/.codex/ or ~/.claude/ except the in-place frontmatter repair of SKILL.md files (which is backed up first).
• If a source config fails to parse, skip it with a message — never guess.
• The home directory is usually not a git repo; diff_summary-style receipts may be unavailable. Verify with direct checks instead and say so plainly.

Notes

• Sarvam discovers skills from ~/.claude/skills, ~/.codex/skills, and ~/.agents/skills natively in current builds; linking into ~/.sarvam/skills guarantees discovery everywhere and survives tool removal.
• Sarvam has no custom slash-command system; commands become skills invoked as $name or by description match.
• $ARGUMENTS in converted commands becomes an instruction to read the arguments from the user's message; AskUserQuestion becomes request_user_input.