Local Claude Extraction¶

Inkwell can use your local Claude CLI for extraction. This is an explicit, local-only subprocess backend named claude-code; it is separate from the existing claude provider, which calls the Anthropic API with an API key.

Inkwell does not sign you in to Claude, install or bundle Claude Code, read or copy credentials, create setup tokens, or proxy requests. The Claude CLI must already be installed and authenticated independently. Local Claude extraction is never available to Modal, Vercel, or another hosted worker.

Configure And Validate¶

Choose a model explicitly. Inkwell has no compiled Claude model default:

inkwell plugins configure claude-code model sonnet
inkwell plugins validate claude-code --json

The requested value may be a current Claude CLI alias or full model identifier. The result and cache provenance record both that requested value and the primary effective model whose token usage matches Claude's terminal result. Claude may also report auxiliary model work, such as a safe-mode classifier; Inkwell keeps those model identifiers in the attempt record and includes their tokens in total usage. Ambiguous primary-model metadata fails closed. Inkwell does not enable --fallback-model.

Optional bounded settings mirror Local Codex extraction:

inkwell plugins configure claude-code timeout_seconds 180
inkwell plugins configure claude-code max_input_bytes 8000000
inkwell plugins configure claude-code max_stdout_bytes 8388608
inkwell plugins configure claude-code max_stderr_bytes 1048576

Validation calls only claude --version and claude auth status --json. It records installed/authenticated/supported state and the sanitized auth class; it discards account, organization, subscription-tier, and credential fields. Validation does not perform inference or start an authentication flow.

On macOS, Claude Code's saved login is Keychain-backed. A sandboxed launcher may be unable to see that login even when claude auth status --json succeeds in your normal terminal. Treat that as a process-isolation diagnostic, not a logout: run validation from the same unsandboxed host context, and do not log in again solely because a sandbox reports runtime_not_authenticated.

Use Your Local Claude CLI¶

Select the backend explicitly for a local fetch:

inkwell fetch ./source.txt \
  --extractor claude-code \
  --force-extraction \
  --json

Or explicitly select it for the current process:

INKWELL_EXTRACTOR=claude-code inkwell fetch ./source.txt --json

claude-code is excluded from templates, default-provider heuristics, automatic routing, and cross-provider retry. Direct Claude/Gemini providers remain the automatic defaults and the only hosted choices.

Authentication And Isolation Boundary¶

The child process receives HOME only so the Claude CLI can use its own saved first-party subscription login. Inkwell removes ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, CLAUDE_CODE_OAUTH_TOKEN, Claude cloud-provider selectors, cloud credentials, and other provider secrets from the child. It accepts readiness only when Claude reports saved claude.ai authentication against the first-party provider. API keys, setup tokens, Bedrock, Vertex, and Foundry authentication fail closed.

Inkwell does not use --bare, because that mode skips saved OAuth/keychain authentication. Instead it runs documented claude -p through stdin with:

--safe-mode and no user/project/local setting sources;
an empty built-in tool list and denied MCP tools;
a strict empty MCP configuration;
permission mode dontAsk;
disabled skills and slash commands;
no session persistence;
explicit model selection and schema-constrained JSON output;
bounded stdin/stdout/stderr, timeout, process-tree cancellation, and a private mode-0700 temporary working directory.

The prompt never appears in argv. Runtime errors are sanitized and do not carry the prompt, provider response body, environment values, account identifiers, or session identifiers.

This boundary prevents Claude tool use and customization in the tested profile. It is still a same-user local process using the host's Claude installation, not a separate OS-level sandbox. Use it only for content you are willing to send to your local Claude CLI.

Subscription Limits, Cost, And Cache¶

Current Anthropic guidance says Claude Agent SDK, claude -p, and third-party app usage draw from the user's subscription usage limits. The proposed June 15 separate monthly Agent SDK credit was paused and is not available.

Because subscription execution has no attributable per-call charge, Inkwell records token usage but reports an unknown monetary amount:

{
  "cost_usd": 0.0,
  "cost_known": false,
  "billing": {"mode": "runtime_managed", "amount_usd": null}
}

The numeric zero is a compatibility field and does not mean the request was free. Cache keys include the CLI version, runtime protocol, requested model, sanitized auth/billing class, transcript, template, prompt, and schema. Cache hits retain the originating effective model, usage, attempts, and provenance.

Recovery¶

Diagnostic	Recovery
Executable missing	Install Claude Code independently and retry validation
No saved subscription login	Authenticate Claude Code independently, then retry validation
API-key/setup-token/cloud auth	Remove that auth override from the local Claude CLI context and use its saved subscription login
Unsupported version	Update Claude Code and rerun validation
Model missing	Configure `claude-code.model` explicitly
Timeout or output limit	Increase the bounded plugin setting only if the source requires it

Inkwell never repairs, copies, or creates Claude authentication state.