deepseek-migration
Skill1 starsChanged 18 days agoMITView on GitHub
Migrate an existing Claude Code CLI setup onto an isolated DeepSeek-backed (or any Anthropic-compatible) endpoint profile that reuses the same skills, plugins, commands, agents, hooks, statusline, and MCP servers without sharing credentials. Use when someone wants to run Claude Code against DeepSeek, onboard a teammate to a dual-profile setup, clone a `~/.claude` profile into a provider-isolated one, audit which installed plugins survive a non-Anthropic backend, or roll such a profile back.
- Reads credentials
1 other project uses a version of this.
Adopted by
Most recently taken up by
- zhangliang0115/ai-plugintheir own version, first seen Sep 6, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- zhangliang0115/ai-plugin★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 2 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 44
Ask, and do not guess:
- Very specific: line 54
Should repo-level
.claude/settings.jsonbe honoured?: Default here is no (--setting-sources user), to stop an untrusted repo's hooks and … - Distinctive: line 71
Run these in order. Every step is idempotent.
- Very specific: line 122
After the first launch, inside dsd:
- Very specific: line 191
Delete the launchers and the profile directory, and remove only the marked alias block from ~/.zshrc.
- Very specific: line 194
Do not restore a whole zshrc backup — unrelated edits made since will be lost.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
DeepSeek migration
Builds a second Claude Code profile that talks to https://api.deepseek.com/anthropic
while sharing all configuration with the normal profile. Credentials, history,
sessions, and permissions stay isolated; skills, plugins, commands, CLAUDE.md,
hooks, statusline, and MCP definitions are shared.
Two names, used throughout:
| Name | Meaning | Default path |
ccd | existing Claude Code profile | ~/.claude |
dsd | DeepSeek profile being created | ~/.config/deepseek-claude/claude |
Why a separate profile at all
CLAUDE_CONFIG_DIR is a complete config-namespace switch: settings, sessions,
history, credentials, permissions, and project approvals all move with it. So one
profile can point at Anthropic and another at DeepSeek, with no cross-contamination
of auth state, and nothing leaking into the calling shell.
What is shared by symlink or mirror, and what is isolated:
| Shared | Isolated |
skills/ (per-entry symlink) | credentials / Keychain namespace |
plugins/ (whole dir symlink: registry, marketplaces, cache) | history, sessions, project approvals |
commands/ | permissions |
CLAUDE.md | model routing, theme, attribution |
hooks, statusLine (mirrored into profile settings) | |
MCP server definitions (mirrored into profile .claude.json) |
Sharing is not a filesystem sandbox. Both profiles can read the same workspace, so the DeepSeek profile can send repository contents to DeepSeek. Say that out loud before running this on a work machine.
Before you start
Ask, and do not guess:
- Which endpoint and model IDs? Per the official DeepSeek integration guide the
defaults are
deepseek-flash[1m]for the main slot anddeepseek-flashfor the small/subagent slot. DeepSeek also maps incomingclaude-*names server-side:claude-opus*→deepseek-v4-pro,claude-sonnet*/claude-haiku*→deepseek-flash. A corporate gateway or relay may use different names, and may or may not accept the[1m]context suffix. Everything below reads model IDs fromprofile.env, so nothing is hardcoded — confirm the real values first. - Is a key already on disk, or will the user paste one?
- Should repo-level
.claude/settings.jsonbe honoured? Default here is no (--setting-sources user), to stop an untrusted repo's hooks and plugins from loading under the DeepSeek account. Opt back in withdsd --setting-sources user,project,localwhen the repo is trusted.
Never print, echo, log, or commit the API key. It is read from a file inside the launcher and never enters the shell environment of the caller. Do not paste it into chat, a skill file, or a commit. Same for the MCP env tokens copied by the merge step.
Procedure
Commands below use $SKILL_DIR, the directory containing this file:
export SKILL_DIR="${SKILL_DIR:-$HOME/.claude/skills/deepseek-migration}"Run these in order. Every step is idempotent; a failed run can be repeated.
1. Inventory the source profile
python3 "$SKILL_DIR/scripts/audit-plugins.py"Reports every enabled plugin as PORTABLE, MODEL-ROUTED, PROVIDER-BOUND, or
TOOL-RISK, with the file and line that triggered the verdict. Read it before
enabling anything on the DeepSeek side. references/compatibility.md explains what
each verdict means and which Claude Code features have no DeepSeek equivalent.
2. Bootstrap the profile
bash "$SKILL_DIR/scripts/bootstrap.sh" --dry-run
bash "$SKILL_DIR/scripts/bootstrap.sh"Creates, with correct modes (dir 0700, key 0600, launchers 0700):
~/.config/deepseek-claude/api-key~/.config/deepseek-claude/profile.env— endpoint + model IDs, the only place to edit~/.config/deepseek-claude/claude/— the profile itself~/.local/bin/deepseek-claude— launcher, safe permission prompts~/.local/bin/dsd— same, plus--dangerously-skip-permissionslink-ccd-skills.sh,merge-ccd-config.py— sync helpers, re-run on every launch- a marked alias block in
~/.zshrc(alias name viaDSD_ALIAS, defaultdsd)
Pass --force only to replace an existing launcher; without it an existing launcher is
left alone and a .new file is written for diffing. That protects a working setup.
3. Confirm what got shared
bash "$SKILL_DIR/scripts/verify.sh"Checks modes, link counts, JSON validity, and that hooks/statusLine/MCP actually
landed in the profile. Then verify.sh --live makes one real API call through the
profile, which is the only check that proves the endpoint and model IDs are right.
4. Migrate plugins
Plugins need no reinstall — plugins/ is a directory symlink into the shared
~/.claude/plugins, so the registry, marketplaces, and cache are the same objects.
What must be mirrored is the enablement, which merge-ccd-config.py does by
copying enabledPlugins and extraKnownMarketplaces into the profile's
settings.json on every launch.
After the first launch, inside dsd:
/plugin # confirm the list matches ccd
/mcp # confirm servers are present; re-auth OAuth ones
/status # confirm model + endpointOAuth logins are deliberately not copied. Each OAuth MCP server (GitHub, Notion,
Linear, Sentry, …) needs one manual /mcp re-auth in the new profile. Say so up front
so nobody reads the empty auth state as a broken migration.
Rules for the audit verdicts:
PORTABLE— enable as-is. Markdown skills, commands, LSP plugins, statuslines, and hooks that only shell out locally.MODEL-ROUTED— enable, then spot-check. Agent files pinningmodel: opus|sonnet|haikuare remapped throughANTHROPIC_DEFAULT_*_MODEL, so they run on whatever tier you configured. Tier semantics change even when the plumbing works: an "Opus" reviewer may now be the same model as everything else. Quality assumptions in those prompts no longer hold automatically.PROVIDER-BOUND— disable unless verified. Anything callingapi.anthropic.comor the Anthropic SDK directly bypasses the profile entirely and will either fail or bill the wrong account.TOOL-RISK— enable only after testing the specific tool. Web search server tools, image input, and thinking-budget controls are the usual casualties.
To disable one without touching the shared profile:
claude plugins disable <plugin>@<marketplace> # run inside dsdThat writes to the profile's settings.json, but merge-ccd-config.py re-copies
enabledPlugins from ccd on every launch and will undo it. For a permanent
difference, add the plugin to DISABLED_PLUGINS in the profile's profile.env
(exposed as DSD_DISABLED_PLUGINS), or drop the plugin from ccd's enabledPlugins.
5. Pin the model IDs
Edit ~/.config/deepseek-claude/profile.env. Defaults the launcher understands:
DSD_BASE_URL=https://api.deepseek.com/anthropic
DSD_MAIN_MODEL=deepseek-flash[1m] # opus / sonnet / fable slots
DSD_SMALL_MODEL=deepseek-flash # haiku slot + subagents
DSD_EFFORT_LEVEL=high
DSD_AUTO_COMPACT_WINDOW=786432Append [1m] to a model ID only if the endpoint advertises a 1M-token window and
accepts the suffix. If requests 400 on the model name, drop the suffix first — that is
the most common misconfiguration.
6. Hand over, with the guardrails
Tell the user explicitly:
- Billing is DeepSeek's, not the Claude subscription. Both profiles can be used on the same machine, but they are different bills.
- Isolation is configuration, not security. Same filesystem, same workspace.
--setting-sources useris what keeps repo-level hooks and plugins out. Changing it widens the trust boundary.- Rotate the key by replacing
api-keyonly, keeping mode0600. ~/.config/deepseek-claude/contains the key and mirrored MCP tokens. Never commit, sync, or share it.
7. Rollback
Delete the launchers and the profile directory, and remove only the marked alias block
from ~/.zshrc. bootstrap.sh --rollback does exactly this.
Do not restore a whole zshrc backup — unrelated edits made since will be lost.
The current setup keeps dated backups under backups/ for reference only.
Driving it from a fresh machine
Greenfield, no existing ~/.claude, works too. bootstrap.sh tolerates a missing
source profile: the link and merge helpers no-op, and the user gets an empty but
functional DeepSeek profile to build up from claude plugins marketplace add ….
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

