Home / Glossary
The vocabulary of a native agent SDK
Short, plain-language definitions for the concepts behind tokenworm — from the C ABI at its centre to the .tworm sessions it produces.
- Native binary
- A program compiled directly to machine code that runs without an interpreter or virtual machine. tokenworm ships as a ~960KB native binary with zero runtime dependencies.
- Systems language
- A language like Zig, Rust, or C that compiles to native code with manual control over memory and low-level resources. tokenworm is written in Zig 0.16.
- C ABI
- The C Application Binary Interface — a stable, language-neutral calling convention. tokenworm exposes its core through a C ABI (src/lib/ffi.zig) so any language with FFI can drive the agent.
- FFI
- Foreign Function Interface — the mechanism a language uses to call functions written in another language. The Python, TypeScript, and Go SDKs use FFI to call libtokenworm.
- Agent loop
- The canonical while-loop an agent runs: call the model, dispatch any tools it requested, update the conversation, and repeat until the task is done. tokenworm runs this in agent_runner.zig.
- Tool use
- The pattern where a model requests actions (read a file, run a command) that the runtime executes and feeds back. tokenworm ships six built-in tools: read, write, edit, bash, grep, glob.
- LLM provider
- A backend that serves a large language model. tokenworm supports five — OpenCode Zen, Anthropic, OpenAI, Ollama, and MiniMax — switchable at runtime.
- MCP
- Model Context Protocol — a standard for connecting agents to external tools and context servers. tokenworm supports MCP over both stdio and SSE transports.
- SKILL.md
- A file format for packaging reusable agent instructions and workflows as a loadable skill. tokenworm supports SKILL.md skills.
- Lifecycle hook
- A defined point in the agent loop where you can intercept and shape behaviour. tokenworm exposes 18 lifecycle hooks.
- Sandboxing
- Restricting what a process can access. tokenworm runs the bash tool in an OS-native sandbox: bubblewrap on Linux and sandbox-exec (Seatbelt) on macOS.
- .tworm session
- tokenworm’s portable binary session format, storing conversation history, provider metadata, timestamps, token counts, and workspace snapshots. Per the README it is 5x smaller and 20x faster to parse than JSONL.
- Cold start
- The time from launching a program to it being ready to do work. The README reports an 8ms cold start for tokenworm.