Advanced & Technical FAQ

What can I do in the Python sandbox?

The run_sandboxed_code step executes Python in a secure sandbox. You can use common data processing libraries, call ai.generateText to generate text using an LLM, call ai.generateObject to produce structured JSON output from an LLM, use rankByEmbeddings for semantic ranking of items, and use createExcel, createPDF, and createWordDoc to generate documents. You cannot make outbound HTTP requests from the sandbox (use a separate connection step for external calls) and you cannot access the file system.

How does context management work in AI Agents?

Chase Agents manages conversation context automatically to stay within the language model's context window. As conversations grow long, the platform uses a combination of smart truncation and summarization to compress older parts of the conversation while preserving the most recent and relevant context. Tool results from older steps are compacted if they are large and old enough. This happens transparently without interrupting the conversation.

How do I connect Claude Desktop to my workspace?

Chase Agents exposes an MCP HTTP endpoint that Claude Desktop and other MCP clients can connect to. Get your workspace API key from the workspace settings panel. In Claude Desktop's MCP configuration, add Chase Agents as an MCP server using the Chase Agents MCP endpoint URL and your workspace API key as the authentication token. Once connected, all your workspace's agents, automations, and connections are available as tools in Claude Desktop.

Can I run multiple workspaces?

Yes. A single Chase Agents account can be an admin of up to two workspaces (for non-whitelisted users). Some organizations use separate workspaces for development and production environments, or to isolate different business units. Each workspace has independent billing, connections, automations, and members. You can switch between workspaces from the workspace selector in the top navigation.

How does the automation rate limit work?

Each automation is limited to 5 executions per 5-minute rolling window. This limit prevents a runaway webhook from consuming your entire action quota in seconds. If your use case requires higher frequency execution, design the automation to process batches of items per run rather than triggering once per item. A single run that processes 100 records counts as one run against the rate limit regardless of how many actions it consumes.

What models does Chase Agents use?

Chase Agents uses different AI models for different internal functions: a primary model for automation execution and agent conversations, specialized models for title generation, summarization, evaluation, and auto-fix. Model selection is managed internally by the platform and optimized for each task type. You do not need to configure models for standard operation. Enterprise customers can discuss custom model configurations with the team.

How is the accept_external_input step used?

The accept_external_input step pauses an automation and waits for data to be submitted to a specific endpoint. This is useful for building forms or data collection flows where you want to pause an automation mid-run, give an external user a URL to submit their data, and then resume processing once they have submitted. It is different from seek_human_approval_for_next_step in that it collects arbitrary data rather than a binary approve or reject decision.

Where can I find the API documentation for Chase Agents?

The Chase Agents MCP API documentation is available through the platform itself: connect an MCP client to your workspace and list the available resources to see the full tool documentation. For the Chase Agents Automations API used for external integrations, contact charles@chaseagents.com or visit the developer section of the platform for API reference details.