FMXExpress has released PasClaw, an open-source AI agent framework written entirely in Delphi Object Pascal. The project compiles to a single native Windows executable with no managed runtime, no Python, and no Node.js required. It ships under the MIT license.
According to the project’s GitHub repository, PasClaw ships as a complete agent framework with a 19-provider LLM catalog spanning Anthropic, OpenAI, Google Gemini, Groq, DeepSeek, Mistral, Ollama, and twelve others. It supports tool calling with built-in filesystem access, shell execution, web search (six providers including DuckDuckGo, Brave, and Tavily), web fetch, and SQLite FTS5 memory search. Mutating tools run serially while read-only tools fan out on worker threads for parallel dispatch.
MCP and Skills
PasClaw implements both MCP transports: stdio via subprocess and Streamable HTTP with SSE and Bearer-token authentication. The framework includes a public server catalog, so pasclaw mcp install replicate or pasclaw mcp install huggingface reads the correct environment variable and writes the authorization header automatically, according to the GitHub documentation.
Skills work through markdown manifests advertised in the system prompt. The model loads skill bodies on demand via fs_read. Installation supports GitHub repositories and ClawHub, with malware-flagged skills blocked at install time.
Eight Chat Channels, HTTP Gateway, Cron
The framework ships with bidirectional integrations for Telegram, Discord, LINE, WhatsApp, Slack, Matrix, IRC, and Email, plus outbound-only support for Microsoft Teams and generic webhooks. An HTTP gateway exposes OpenAI-compatible endpoints at /v1/chat/completions and /v1/responses, the latter enabling Codex CLI to drive its own tools through PasClaw.
An embedded web UI (vanilla ES2020, single HTML file) provides eight tabs: Chat, Memory, Files, MCP, Cron, Skills, Logs, and Settings. The cron system supports scheduled agent tasks with standard cron syntax.
Hashline Edits and Sandboxing
For code editing, PasClaw uses the Hashline format introduced by can1357. Edits anchor to numbered lines with hash verification rather than fuzzy search-and-replace, which the framework’s documentation claims reduces token usage by 20 to 40 percent compared to unified-diff formats across 16 frontier models.
Security defaults include a workspace boundary that locks all filesystem and shell operations to a single directory, an always-on shell denylist blocking sudo, rm, mkfs, and similar destructive commands, and regex-based exceptions for controlled access outside the sandbox.
Why Delphi Matters for Agent Architecture
The choice of compiled Object Pascal is deliberate. PasClaw starts in milliseconds with no interpreter overhead, produces deterministic memory layouts with no garbage collector pauses, and deploys as a single pasclaw.exe binary alongside standard OpenSSL DLLs. The FMXExpress announcement positions this as an alternative to the Python-dominant agent ecosystem, noting that Delphi’s type system catches bugs at compile time that dynamically typed frameworks surface at runtime.
PasClaw can also serve as a backend for Embarcadero KAI, allowing RAD Studio’s integrated AI features to connect to custom agent infrastructure, tools, and model providers rather than being limited to basic chat interactions.
The release signals continued fragmentation in agent framework SDKs. With OpenClaw, Hermes, Claude Code, Codex, and now PasClaw each targeting different developer ecosystems, the architectural patterns of tool-calling agents, MCP integration, and persistent memory are becoming language-agnostic primitives rather than features tied to any single platform.