One responsibility per block
The studio provides blocks for Codex, Claude Code, Claude Chat, OpenAI Chat, CLIs, and Python scripts. Choose the block that matches your tool and the kind of result you need. Each integration has its own prerequisites.
Configure an agent
First, check that the tool works in your terminal or that its API access is correctly configured. In the block, check the model, instruction, and working directory. Use named inputs to distinguish the instruction, data, and context.
With Codex, instructions can reference an input as @in or @nom_de_l_entree. The working directory determines the command’s local context. Limit the agent’s permissions to the files and actions it needs.
Connect an MCP server
The instance’s MCP settings let you manage server references. An MCP block stores an mcp_ref reference; addresses and secrets are resolved from local configuration. They do not need to be copied into every blueprint.
An example configuration is provided in config/mcp_servers.example.json. Local values are saved in user/config/mcp_servers.local.json.
Stay in control of your data
The studio and its projects run locally. A block that calls a remote service sends the data that service needs. Check each integration’s inputs and permissions before giving it sensitive data.
API costs and agent subscriptions are handled by their providers. Using Bloxmith for free does not make those services free.
Read, transform, and produce
File blocks feed your local content into the graph. Python is suitable for deterministic transformations and validation. Display and save blocks make results available to inspect or reuse in another step.