Glossary
The words that keep coming up in our articles and tool pages, defined in a few lines, with a link to the source whenever the definition is factual.
Coding agent
AI software that does more than suggest code: it reads a project's context, turns a request into a multi-step plan, then executes and checks those steps (editing code, running tests, fixing). This is the definition Gartner used for its first Magic Quadrant for coding agents (May 2026). Examples in the directory: Claude Code, Codex, OpenCode, Aider, Cline.
Agentic
Describes a use or a tool where the model acts over several steps with tools (terminal, browser, APIs) instead of answering in a single message, as in “agentic coding” or “agentic tasks”. Vendors measure these abilities with tests such as DeepSWE or Terminal-Bench.
MCP (Model Context Protocol)
An open-source standard that connects AI applications to external systems: data sources (files, databases), tools (search engines, calculators) and workflows. Its authors compare it to a USB-C port for AI applications (modelcontextprotocol.io). The “MCP support” row on our tool pages says whether a tool can plug into it.
MoE (mixture of experts)
A model architecture (Mixture of Experts) where only part of the parameters work on each token. Example: Meta's Muse Spark Open has 816 billion parameters in total, but only 42 billion are active, with 8 of 256 experts activated per token (model card). The benefit: a very large model that is cheaper to run than a dense model of the same size.
Open weights vs open source
An open-weight model can be downloaded, but its license may set conditions. Open-source software must meet the Open Source Initiative definition, which notably forbids discrimination against persons or fields of endeavor. Example: the Muse Spark Open license requires displaying “Built with Muse Spark Open” and a separate agreement with Meta to offer it as an API above $20M in revenue (Muse license): open weights, but not open source. DeepSeek V4.1-Flash-1017 is released under the MIT license (Hugging Face).
Context window
The maximum amount of text, measured in tokens, that a model can take into account in one request: instructions, files, history and the answer. Example: GPT-6.1 Sol has a 1,050,000-token context window, with up to 128,000 output tokens (OpenAI). A larger window lets you load more code but costs more: some vendors charge more above a threshold.
Tokens
The unit of text a model processes, roughly 3 to 4 characters of English text according to Zed. Your prompt, files, history, tool results and the answer all use tokens (OpenAI). APIs charge per million tokens, with different input and output prices: see our model comparison.
SWE-bench
A family of tests measuring how well a model resolves real GitHub issues. The original version has 2,294 tasks from 12 Python repositories; SWE-bench Verified is a human-filtered subset of 500 tasks (swebench.com). The score is the percentage of tasks resolved.
DeepSWE
A long-horizon software engineering test published by Datacurve: the agent must understand a repository, build a feature and validate it with tests. Version 1.1 (June 2026) has 113 tasks and grades the committed code in an isolated environment for more reproducible results (Datacurve). Note: scores announced by vendors often use their own agent harness and reasoning effort, so they are not strictly comparable.