- The
copilot.elpackage received performance and stability updates in July and August 2026, with GitHub’s own September 2022 research finding that Copilot users completed tasks 55% faster on average, a figure that carries more weight now that the tooling runs natively inside Emacs. - Emacs developers are connecting to a wider range of models, including OpenAI’s GPT-4 and local Ollama instances, through frameworks such as
eacopandllm.elbypassing vendor lock-in without leaving the editor.
Emacs now runs agentic AI tooling that can read files, execute Lisp functions and generate test suites without leaving the editor, and the package ecosystem supporting it has been moving fast. Updates to copilot.el and the broader copilot-cli in July and August 2026 tightened latency and improved suggestion quality, but the more interesting story is how open-source abstraction layers are letting developers swap models, run inference locally and build multi-step automated workflows inside an environment that predates the internet by a decade.
Copilot.el Gets Faster
GitHub Copilot’s Emacs client, copilot.elreceived updates to its underlying command-line interface in late July and early August 2026 that reduced latency and improved overall responsiveness. The package functions as a direct bridge between the editor and Copilot’s suggestion engine. GitHub’s own September 2022 research found that developers using Copilot completed tasks an average of 55% faster than those without AI assistance, though that figure comes from controlled conditions and real-world gains vary by task type and codebase. For Emacs users, tighter integration means less context-switching between boilerplate and the problems that actually require attention.
gptel: One Client, Many Models
gptel is the most flexible of the current Emacs LLM clients, supporting OpenAI Anthropic and local models running through Ollama or llama.cpp from a single interface. It has been available for AI-powered code completion since November 2022 and has since added tool use, meaning the model can call Emacs Lisp functions directly rather than only returning text. That makes it genuinely agentic in a narrow but useful sense: gptel can generate documentation, refactor a selected region or draft a commit message inside the same session. Emacs users weighing local versus hosted models will find gptel’s backend-switching the most practical way to compare them without reconfiguring their workflow.
llm.el as the Abstraction Layer
llm.el sits underneath several of the tools described here, providing a common interface for talking to LLM backends whether those are API-driven services like OpenAI or self-hosted runtimes like GPT4All and llama.cpp. The library handles embeddings and tool use as well as basic chat, which means packages built on top of it inherit those capabilities without reimplementing them. The practical value is portability: a workflow built against llm.el can switch backends by changing a configuration variable rather than rewriting integration code. How widely that portability is being used in practice across the Emacs community is harder to verify from public material alone.
Local Models and Data Governance
Running inference locally removes the question of whether sensitive code ever reaches an external server. Tools like Ollama and llamafile make this straightforward in Emacs through gptel and the newer ellama package. The emacs-copilot project documents one concrete configuration: WizardCoder 34b running via llamafile on a Mac Studio M2 Ultra, which the project’s maintainers report delivers completions they consider production-quality. Local setups are particularly relevant for teams under strict data governance constraints or developers who want to avoid recurring API costs, though they require capable hardware to match hosted model latency. The trade-offs around Emacs Copilot’s evolving feature set versus a fully local stack are still being worked out by the community.
Agentic Tools in Emacs
The more ambitious end of the Emacs AI ecosystem involves packages where the model does more than complete a line. Emigo lets an LLM interact with the editor environment directly, reading and writing files by executing Lisp functions based on its own reasoning about a task. AI mode for Emacs adds interactive chat with buffer binding, code completion and automated bug detection in a single package. Both represent a shift from suggestion to execution: the model is given tools and a goal rather than a cursor position and a prefix. Whether multi-step agentic workflows inside a text editor hold up under real project conditions is still an open question, and the safety characteristics of multi-agent systems apply here too, even in a single-user setting.
