SillyTavern — Claude (Subscription) Proxy

Use your Anthropic Pro / Max subscription for SillyTavern chat instead of paying for API credits. A SillyTavern Server Plugin routes requests through the locally-installed Claude Agent SDK — the same subscription mechanism VS Code, Cursor, and Zed use — and ships with a companion "Claude Max" UI extension for one-click connection and Claude-native settings.

What v2 gives you

• Current models — Claude Fable 5, Opus 4.8/4.7/4.6/4.5, Sonnet 4.6/4.5, Haiku 4.5, plus explicit "(1M context)" variants for the tiers that support extended context.
• One-click connect — the Claude Max panel (Extensions drawer) configures SillyTavern's Custom endpoint for you. No URLs to paste.
• Claude-native reasoning effort — low / medium / high / xhigh / max, set in the Claude Max panel. SillyTavern's built-in Reasoning Effort dropdown does not work for Claude models on Custom endpoints (its "Maximum" is downgraded to high client-side, and the field is dropped server-side for non-OpenAI model IDs) — this panel bypasses all of that.
• Thinking display — Claude's reasoning streams into SillyTavern's native collapsible "thoughts" block (reasoning_content). Toggle in the panel; also enable "Show model thoughts" in ST's settings.
• Real multi-turn context — chat history is replayed as a genuine Claude session (synthetic session resume), not folded into one giant string: proper role separation, working prompt caching, better long-RP quality.
• Roleplay isolation — no coding system prompt, no host CLAUDE.md / settings / MCP connectors leaking into your scenes, no agent tools. Your character card and world info are the entire system prompt.
• Stop sequences enforced — the Agent SDK has none, so the plugin scans the stream server-side (\n{{user}}: guards work as expected).
• Resilience — automatic OAuth token refresh, rate-limit retries, 1M-context → base-model fallback with a one-hour probe cooldown.
• Served-model guard — an explicit Fable request is NEVER silently substituted: if the upstream resolves your claude-fable-5 pick to anything else (e.g. Fable temporarily disabled on your plan), the request errors out with a clear message instead of quietly switching to Opus mid-roleplay.
• Quota meter — live 5-hour / 7-day subscription window utilization in the panel, so a long session never hits a surprise lockout.
• Privacy sweep — the transcript the Claude CLI writes for each live turn is deleted after the request; your roleplay does not persist in plaintext under ~/.claude/projects.

Prerequisites

On the same machine SillyTavern runs on:

1. Sign in once with your Pro / Max account (installs nothing into ST):

   npm i -g @anthropic-ai/claude-code
   claude login
   

   (Headless / Docker: generate a token with claude setup-token and set CLAUDE_CODE_OAUTH_TOKEN instead.)
2. SillyTavern config.yaml: enableServerPlugins: true.

Install

From your SillyTavern install directory (the one containing server.js):

node plugins.js install https://github.com/LukaTheHero/SillyTavern-ClaudeSubscription
cd plugins/SillyTavern-ClaudeSubscription
npm install

Restart SillyTavern. The server log should show:

[claude-subscription] installed UI extension v2.0.1 at public/scripts/extensions/third-party/SillyTavern-ClaudeMax
[claude-subscription] standalone listener: http://127.0.0.1:8901/v1
[claude-subscription] initialised — endpoint http://127.0.0.1:8901/v1

The companion Claude Max UI extension is installed/updated automatically on startup. Hard-refresh the browser (Ctrl+F5) after the first install so SillyTavern loads it. Opt out with CLAUDE_SUBSCRIPTION_NO_UI_INSTALL=1.

Alternative: install the panel via the extension dialog

The repo doubles as a regular UI extension (manifest.json at the root), so the Claude Max panel can also be installed from Extensions → Install extension by pasting the same GitHub URL. The auto-installer detects a dialog-installed copy and stands down (and a runtime guard dedupes if both ever load).

⚠️ The dialog installs only the panel. The server plugin (the actual proxy) must still be installed into plugins/ as above — without it, Connect has nothing to connect to.

Connect

1. Open the Extensions drawer → Claude Max.
2. Click Connect to Claude Max.
3. Pick a model from the normal model dropdown (e.g. Claude Fable 5 or Claude Opus 4.8 (1M context)).
4. Chat.

Set effort/thinking in the same panel — changes apply from the next message, no reconnect needed.

Settings reference (Claude Max panel)

Setting	Default	What it does
Reasoning effort	Auto	How hard Claude reasons before replying, low → max. Auto sends nothing (model default, ≈high). Higher = better consistency on complex scenes, slower replies, more quota.
Thinking mode	Adaptive	Whether extended thinking happens at all. Adaptive: model thinks only when the message warrants it. Always on: every reply. Off: none — except Fable 5 / Opus 4.7+ always think (can't be disabled), and thinking is auto-disabled on other models when Max response length < 2048 tokens.
Show reasoning	On	Display only — doesn't change whether thinking happens. On: the thinking summary streams into ST's collapsible "thoughts" box (enable "Show model thoughts" in ST too). Never added to chat history or re-sent as context.
Identity mode	Off	Off: your character card is the entire system prompt. On: prepends Anthropic's Claude Code preamble (the framing the claude CLI uses) — fixes model self-identification ("are you Opus or Sonnet?") at the cost of extra tokens and a coding-assistant flavor. Leave off for RP.
Session resume	On	How history reaches Claude. On: replayed as a real multi-turn session — better turn awareness, working prompt caching (faster + less quota burned re-reading context); swipes/edits handled naturally. Off: history flattened to one User:/Assistant: text block (troubleshooting only).

Known limitations (Agent SDK)

• Temperature / Top-P / Top-K are not supported on the subscription path — the Agent SDK exposes no sampling controls. (Adaptive-thinking models reject them anyway.)
• Assistant prefill is emulated: a trailing assistant message ("Start Reply With", continue) becomes a continuation instruction rather than true Messages-API prefill. Works well in practice; the model occasionally paraphrases instead of continuing verbatim.
• Embeddings return 501 — use a separate embedding source.

Environment overrides

Variable	Default	Purpose
CLAUDE_SUBSCRIPTION_PORT	8901	Listener port
CLAUDE_SUBSCRIPTION_HOST	127.0.0.1	Listener host
CLAUDE_SUBSCRIPTION_USE_RESUME	1	0 forces the fold path
CLAUDE_SUBSCRIPTION_MAX_TURNS	1	SDK maxTurns
CLAUDE_SUBSCRIPTION_CLAUDE_PATH	–	Explicit claude executable path
CLAUDE_SUBSCRIPTION_NO_UI_INSTALL	–	1 skips UI-extension auto-install

Changing the port or host? Update Endpoint (advanced) in the Claude Max panel to the new http://<host>:<port>/v1 before clicking Connect — the panel's Connect button, per-request settings injection, and quota meter all use that stored endpoint.

Endpoints

Method	URL	Purpose
GET	http://127.0.0.1:8901/status	SDK + credential health
GET	http://127.0.0.1:8901/v1/models	Model list (incl. 1M variants)
GET	http://127.0.0.1:8901/v1/usage/quota	Subscription window utilization
POST	http://127.0.0.1:8901/v1/chat/completions	Chat (SSE + JSON)
POST	http://127.0.0.1:8901/v1/embeddings	Always 501
GET	http://<sillytavern>/api/plugins/claude-subscription/status	Browser health check

Direct API users: the plugin accepts the standard reasoning_effort body field, or a claude_subscription: { effort, thinking, thinking_budget, show_reasoning, identity_mode, use_resume } object for full control. thinking_budget (tokens) only applies when thinking: "on" and the model is not adaptive-only (Opus 4.7+/Fable ignore it); it is clamped to ≥ 1024 and ≤ max_tokens − 512.

API-billing fallback (optional)

Enter a real sk-ant-* key as the Custom API key and the plugin forwards it as ANTHROPIC_API_KEY — billing that key instead of the subscription. Anything else in the key field is ignored (subscription auth).

Troubleshooting

"Failed to load @anthropic-ai/claude-agent-sdk" — run npm install inside the plugin directory (without --omit=optional; the Claude CLI ships inside the SDK as a platform package) and restart SillyTavern.

Auth errors mid-chat — the plugin auto-refreshes the OAuth token once per request; if it still fails, run claude login on the SillyTavern host as the same OS user that runs server.js.

Model list didn't populate — click Connect again; check http://127.0.0.1:8901/status in a browser.

1M variant quietly serving 200k — extended context needs Extra Usage on some plans; after one failure the plugin serves the base model for an hour, then probes again. Watch the server log for the cooldown message.

License

GNU AGPL v3.0 or later.