Use your own Anthropic API key to power your Claws with Claude. Bring your own key for Claude Sonnet, Opus, and Haiku, with no markup, full control over costs and model selection.

Overview

ClawStaff uses a Bring Your Own Key (BYOK) model for AI providers. Connect your Anthropic API key to power your Claws with Claude, the model family built for careful reasoning, extended context, and safety-conscious operation. Your Claws call the Anthropic API directly at your negotiated rate, with zero markup from ClawStaff. You control model selection, spending, and usage across your entire AI workforce.

This integration supports the full Claude model family: Claude Opus for maximum capability and deep reasoning, Claude Sonnet for the best balance of intelligence and speed, and Claude Haiku for high-throughput, low-latency tasks. Assign different models to different Claws based on what each workflow demands.

What Your Claw Can Do

Choose the right Claude model per Claw. Assign Claude Opus to Claws handling complex analysis, code review, or multi-step reasoning. Use Claude Sonnet for general-purpose work that needs strong intelligence at reasonable cost. Deploy Claude Haiku for high-volume tasks like message triage, quick summaries, and classification.
Use extended context windows. Claude models support large context windows, allowing Claws to process entire codebases, long documents, or extended conversation histories in a single request. This is especially valuable for code review, document analysis, and knowledge synthesis tasks.
Track token usage. The ClawStaff dashboard shows token consumption per Claw, per task, and per time period. Monitor input and output tokens separately to understand where your spend goes.
Set rate limits. Configure per-Claw rate limits to control costs. ClawStaff enforces token-per-minute and request-per-minute caps to prevent runaway spending.
Benefit from careful reasoning. Claude models are designed for careful, step-by-step analysis. Claws powered by Claude excel at tasks requiring precise interpretation of ambiguous instructions, code logic analysis, and generating balanced, well-reasoned responses.

Who Can Reach Your Claw

The Anthropic integration powers the AI model behind your Claw. Access control is determined by the Claw’s scoping level and the whitelisting on its communication channels (Slack, Teams, Gmail, etc.), not by the model integration itself. However, the BYOK model provides its own layer of isolation:

By API key. Each API key is isolated to your Anthropic workspace. Your Claw’s API calls are billed to your account and invisible to other ClawStaff customers.
By workspace-scoped key. Anthropic’s workspace-scoped keys let you further restrict which workspace and rate limits apply to each Claw.

The three Claw scoping levels (Private, Team, Organization-wide) are enforced on the communication channels (Slack, Teams, Gmail) not on the model integration. A Private Claw using Claude Opus is private because only you can message it, not because of the API key. The key isolation ensures that your model usage, prompts, and completions are never shared across tenants.

How It Works

Generate an API key. In the Anthropic Console, create a new API key. You can use workspace-scoped keys for additional access control.
Add the key to ClawStaff. Paste your Anthropic API key into the integration card on the ClawStaff dashboard. The key is immediately encrypted and stored.
Assign a Claude model to each Claw. When configuring a Claw, select which Claude model it should use. Switch models at any time without reconfiguring the Claw’s integrations or workflows.
Claw calls Anthropic directly. At runtime, the Claw uses your API key to call the Anthropic Messages API from within its ClawCage container. Requests go directly from the container to Anthropic’s servers. ClawStaff does not proxy, log, or intercept API traffic.
Monitor usage. Token consumption and request counts are tracked per Claw and displayed in the ClawStaff dashboard. You can also monitor usage in the Anthropic Console.

Security

API key encrypted at rest. Your Anthropic key is stored using AES-256 encryption in ClawStaff’s database. The plaintext key exists only in memory, inside the ClawCage, during task execution.
Key never logged. ClawStaff does not log, cache, or transmit your API key outside the encrypted store and the runtime container. There are no intermediate proxies or relay services.
ClawCage isolation. Each Claw runs in a dedicated Docker container. Your Anthropic API key is injected as an environment variable at runtime and is inaccessible to other Claws, containers, or the host system.
Direct API calls. Requests go from the ClawCage directly to Anthropic’s API. ClawStaff never acts as a middleman, so your key is never exposed to a third-party relay.
Revoke at any time. Delete the key from the Anthropic Console to instantly cut off access. You can also remove or rotate the key in ClawStaff without affecting your Claw configurations.
No data retention. ClawStaff does not store prompts, completions, or any data sent to or received from Anthropic. Only aggregate token counts are recorded for usage visibility.

Cross-Integration Workflows

The Anthropic integration powers every Claw that uses Claude, enabling sophisticated cross-tool automation:

Deep code review. A Claude Opus-powered Claw can review GitHub pull requests with thorough, step-by-step reasoning. It can identify subtle logic errors, suggest architectural improvements, and explain its feedback clearly in review comments. Claude’s extended context window means it can review large diffs without losing track of the broader codebase.
Detailed document synthesis. When paired with Notion or Confluence, a Claude-powered Claw can synthesize information from multiple sources into well-structured documentation. Claude’s careful reasoning produces balanced, accurate technical writing rather than superficial summaries.
Thoughtful triage and classification. A Claude Sonnet-powered Claw can read Slack messages and distinguish between genuine bugs, feature requests, questions, and noise. Claude’s instruction-following precision means fewer misclassified tickets in Jira and fewer false-positive alerts.
High-throughput processing with Haiku. Deploy Claude Haiku for Claws that handle high-volume, lower-complexity tasks: acknowledging Slack messages, labeling GitHub issues, updating Jira fields, or routing requests. Fast, cheap, and accurate enough for classification and extraction work.
Extended context analysis. Claude’s large context window makes it ideal for Claws that need to process entire documents, long Slack threads, or full pull request histories. A single Claw can read an entire sprint’s worth of Jira tickets and produce a thorough retrospective in Confluence.

Anthropic AI Agent Integration