How Speedwave works
While What is Speedwave introduces the platform from a user perspective (a single desktop application with chat and CLI interfaces), this guide explores the underlying architecture: how execution environments are orchestrated when you open a project and how components maintain strict separation.
The system architecture is structured around three core principles: complete resource isolation per project, dedicated worker containers for external services, and an MCP gateway mediating all model interactions.
Project-level isolation
Section titled “Project-level isolation”Speedwave organizes all workspaces around distinct projects, each mapped to a specific directory on your host machine. Every project operates within an independent container network, token storage directory, and Claude home folder. This structure prevents activities in one project from affecting another. Even if an individual worker container were compromised, it would have no access to the networks or credentials of other projects. For technical details, see the Isolation model.
Worker architecture and credential separation
Section titled “Worker architecture and credential separation”The container running Claude Code contains no external credentials and has no direct network access to third-party services. When a project connects to an integration such as GitLab or Slack, that service runs inside a dedicated worker container with its credentials mounted in read-only mode. The model reasoning engine never interacts directly with service tokens, preventing credential leakage across integrations.
Communication between Claude and individual workers passes through the Tool Gateway. Regardless of how many integrations are enabled, Claude receives access to exactly two tools: search_tools and execute_code. This keeps context window consumption low as additional services are connected. The Tool Gateway also records audit logs and ensures results undergo PII tokenization before they are returned to the model.
Bundled runtime environment
Section titled “Bundled runtime environment”Every project initializes Claude inside a preconfigured container that provides built-in skills, output formatting rules, status indicators, and baseline settings. It also includes five official Anthropic plugins loaded at startup alongside any signed Speednet plugins enabled for the project. These assets operate alongside team-specific .claude/ directories committed in your repository, with project files taking precedence. See Built-in skills and plugins for a complete inventory.
Claude Code runs with pinned versioning and deterministic CLI flags to ensure consistent behavior across all team environments:
| Parameter | Configuration |
|---|---|
| Version | 2.1.206 (fixed release, never latest) |
| Binary path | /usr/local/bin/claude |
| CLI flags | --dangerously-skip-permissions, --mcp-config, --strict-mcp-config (ignores workspace .mcp.json), --thinking-display summarized, --ide |
| Runtime environment | Telemetry and auto-updates disabled, sandbox active, terminal flicker prevention, clipboard shim, 30-minute MCP idle timeout |
Speedwave does not set default values for ANTHROPIC_MODEL; model selection is governed exclusively by your project claude.env. See Binary authenticity for details on release verification.