Ollama Buddy: Local LLM Integration for Emacs

Ollama Buddy

An Emacs package for interacting with local LLMs via Ollama, with support for remote providers (OpenAI, Claude, Gemini, Grok, Copilot, Codestral, DeepSeek, OpenRouter, OpenCode Go) and Ollama Cloud models. Requires Emacs 29.1+.

(use-package ollama-buddy
  :ensure t
  :bind
  ("C-c o" . ollama-buddy-role-transient-menu)
  ("C-c O" . ollama-buddy-transient-menu))

That’s it!, start ollama serve, open Emacs, press C-c o, select [o] to open chat, and go, here is what will be presented:

There are videos demonstrating functionality:

https://www.youtube.com/@OllamaBuddyforEmacs

Want More Details?

For a deeper dive have a ganders at:

Full README – Detailed README tutorials, transient menu guide, and role setup

Demos

See - Videos

In-buffer replacement demo

Demonstrating in-buffer in-line modification, now we are out of the chat buffer!

Tool Calling Demo

LLMs can invoke Emacs functions: read/write files, execute shell commands, search buffers, and more.

Advanced Tool Calling Demo

Demonstrating advanced ollama tool calling with multiple tools working together.

Quick Demo

Submit queries, swap models, view token usage and statistics:

Features

Plan Mode – Structured task planning with org-mode TODO integration. /plan generates a step-by-step plan as TODO headings; cycle states with S-<right>; execute steps with /plan-next (with full tool calling support)
Tool Calling – LLMs can invoke registered Emacs functions (file ops, shell commands, buffer search, calculations). Safe mode restricts to read-only tools; unguarded mode (☠) bypasses all prompts for uninterrupted execution.
Multiple Providers – Local Ollama, Ollama Cloud, OpenAI, Claude, Gemini, Grok, GitHub Copilot, Codestral, DeepSeek, OpenRouter, OpenCode Go, and any generic OpenAI-compatible server (LM Studio, llama.cpp, vLLM, Jan…)
Web Search – Inline @search(query) syntax for real-time web search via Ollama’s API
Interactive Menus – Transient popup menus with role-based command grouping
Roles & Presets – Switchable command configurations (developer, writer, tutor, documenter, custom); each command carries a :destination hint (chat or in-buffer) so rewrites go to the right place automatically
Sessions & History – Save/load conversations, prompt history navigation
Thinking Block Folding – Reasoning model output (<think> tags, DeepSeek API) rendered as collapsible org headings; TAB on the heading to peek mid-stream
File Attachments – Attach files and images (vision models) to conversations
Model Management – Pull, delete, copy models; categorized recommendations; cloud model support
Multishot & Benchmark – Send the same prompt to multiple models simultaneously; benchmark all models in one command with per-model timeout
Parameters – Full control over temperature, top_k, top_p, and all Ollama API options
Token Tracking – Real-time token usage statistics with graphs
Skills – Specialized AI personas and task-specific instructions integrated into the User Prompts library; use `/skill` slash command or `@skill(name)` for inline expansion
RAG – Index local documents and source code; retrieve relevant context via semantic search; inline @rag(query) syntax; configurable embedding backend (Ollama /api/embed or any OpenAI-compatible /v1/embeddings service, e.g. LM Studio)
In-Buffer Replace – Stream LLM output directly into the source buffer, replacing the selected region in-place; accept/reject confirmation; inline diff view (C-c d) with word-level smerge highlighting; automatic code-fence stripping

Alternative LLM Feature Comparison Matrix

A cross-cutting view of key features across the major chat clients, to the best of my knowledge, I will try and keep up-to-date, but things are moving quickly!

Feature	gptel	chatgpt-shell	ellama	org-ai	ollama-buddy
Zero-config Ollama	No	No	No	No	Yes
Works in any buffer	Yes	Yes	Yes	Yes	Yes
Dedicated chat buffer	Yes	Yes	Yes	Yes	Yes
Org-mode native output	Yes	Yes	Yes	Yes	Yes
Streaming	Yes	Yes	Yes	Yes	Yes
Vision / image input	Yes	Yes	Yes	No	Yes
Thinking block support	Yes	Yes	Yes	No	Yes
Tool calling	Yes	Yes	Yes	No	Yes
MCP support	Yes	No	Yes	No	No
RAG / document indexing	No	No	Yes	No	Yes
Web search integration	No	No	No	No	Yes
In-buffer replace / rewrite	Yes	No	Yes	No	Yes
Inline code completion	No	No	No	No	Yes
Session save/load	Yes	Yes	Yes	Yes	Yes
Transient menu	Yes	No	Yes	No	Yes
Prompt history	No	Yes	No	No	Yes
Multi-model comparison	No	No	No	No	Yes
Slash commands	No	No	No	No	Yes
Image generation	No	Yes	No	Yes	No
Speech input/output	No	No	No	Yes	No
Project integration	No	No	No	No	Yes
Ollama Cloud models	No	No	No	No	Yes
Curl backend fallback	No	No	No	No	Yes
MELPA	Yes	Yes	Yes	Yes	Yes
Emacs requirement	27.1	28.1	28.1	27.1	29.1

Updates

What’s New (v7.4.2)

Kimi K2.6 cloud model support.

Kimi K2.6 Cloud Model – Added kimi-k2.6:cloud to ollama-buddy-cloud-models (appears in C-c m with the ☁ indicator). Moonshot AI’s native multimodal agentic model, running on Ollama’s cloud (NVIDIA Blackwell hardware) with long-horizon coding, coding-driven design, autonomous execution, and swarm orchestration capabilities. Requires a paid Ollama cloud subscription (https://ollama.com/upgrade) – without one, prompts return an authentication error pointing to the upgrade page. Most other entries in ollama-buddy-cloud-models still run on the free tier.

What’s New (v7.4.1)

Response post-processing hook and a thinking-block point-anchor fix.

=ollama-buddy-response-post-process-functions= – New hook run on the streamed response region just before markdown-to-org conversion. Each function is called with (START END) buffer positions and may edit the region freely.
=ollama-buddy–normalize-numbered-lists= – Default hook entry that splits glued numbered-list items of the form N. [CATEGORY] ... onto their own lines when dense models (notably cloud gemma) emit them as a single paragraph. Idempotent on already-clean input.
Thinking-block point anchor fix – Fixed the first response token being inserted at the top of the folded thinking block instead of after *** Response. Both the API-style (message.thinking) and tag-style (</think>) paths now re-anchor to (point-max) after finalize-thinking-block returns.
Emacs Blog Reviewer prompt – New user prompt that reviews an Emacs blog post for grammar and technical accuracy (function/variable names, key bindings, elisp syntax, deprecated APIs, version claims). One-line-per-issue output format.

What’s New (v7.4.0)

Directory-scale project annotation via structured output.

=/annotate-dir= – Batch-annotate every source file in a directory. Each file is attached, sent as its own request, and the returned JSON is auto-applied to .simply-annotations.el when the response finishes. Prefix arg recurses. Cancel with /annotate-cancel.
Context-aware batching – Estimated token cost (chars / 4) is checked against the current model’s context window; oversized files are skipped. Configurable via ollama-buddy-annotate-context-reserve (default 0.5).
Structured-output skill auto-load – The batch installs the Annotate Project skill and its #+FORMAT: JSON schema before starting, so structured output is guaranteed. The schema now allows file, defun, and line tags.
=C-c A= – New chat-buffer binding for ollama-buddy-annotate-apply-last-response.
New cloud models – Added deepseek-v3.2:cloud, gemini-3-flash-preview:cloud, qwen3-next:80b-cloud, and nemotron-3-super:cloud.

What’s New (v7.3.0)

Direct launch support for external agents.

Direct Agent Launch – The /launch system now supports external tools that are not built-in ollama launch integrations. Registry entries with :direct t are invoked as standalone executables in a terminal, bypassing ollama launch entirely. Direct agents only require their own executable on PATH (no ollama dependency).
Oh My Pi (omp) – Added to the built-in agent registry as the first direct agent.
Flexible Model Passing – Direct agents receive --model MODEL by default. The :model-flag property lets entries customise or disable model passing.
Detached Terminal Sessions – Launched terminals are fully decoupled from Emacs; closing Emacs no longer kills running agent sessions.
Standalone Launch – M-x ollama-buddy-launch-external (or transient v) launches an agent from the current directory without requiring the chat buffer. In Dired, opens in the displayed directory.

What’s New (v7.2.0)

Small model warning for agent launch.

Small Model Launch Warning – When launching a local model below 7B parameters via /launch or the Model Management Launch button, a confirmation warning now explains that coding agents prepend 20,000+ tokens of system prompt to every request, which can cause small models to appear frozen or take many minutes to respond. Configurable via ollama-buddy-launch-small-model-threshold (default 7; set to nil to disable). Cloud models skip the warning.

What’s New (v7.1.4)

Expanded launch agent registry.

Full =ollama launch= Agent Registry – The built-in agent registry now includes all frontends supported by ollama launch: Cline, Droid, OpenCode, OpenClaw, and Pi, in addition to Claude Code and Codex. Each agent only appears when its executable is found on PATH.

What’s New (v7.1.3)

Fix process-filter crash on cloud usage limit responses.

Cloud limit crash fix – Sending a prompt to an Ollama cloud model when the session was at 100% used to spam error in process filter: Wrong type argument: listp, 82 and never surface a real message. The non-2xx response body’s chunked-transfer chunk-size prefix (e.g. 82\r\n{...}) was being parsed as an integer and crashing alist-get downstream. The streaming filter now strips leading non-={= bytes before parsing and requires the parsed value to be a cons; handle-http-error is also defensive against non-alist input. Cloud usage limits now produce a clean error in the chat buffer.

What’s New (v7.1.2)

GLM 5.1 cloud model support.

GLM 5.1 Cloud Model – Added glm-5.1:cloud to ollama-buddy-cloud-models (appears in C-c m with the ☁ indicator). GLM-5.1 is Z.ai’s next-generation flagship model for agentic engineering, with significantly stronger coding capabilities and support for tools and thinking.

What’s New (v7.1.1)

Gemma 4 model support.

Gemma 4 Cloud Model – Added gemma4:31b-cloud to ollama-buddy-cloud-models (appears in C-c m with the ☁ indicator).
Gemma 4 Recommended Models – Added gemma4:e2b, gemma4:e4b, gemma4:latest and gemma4:26b to the “Efficient & Capable” category shown via C-c L. gemma4:26b is a Mixture-of-Experts model with only 4B active parameters, so it runs comfortably on 16GB GPUs (with partial CPU offload) and on Apple Silicon with shared memory. The dense gemma4:31b is still left off the recommended list.
Gemma 4 Vision Support – All Gemma 4 variants are registered as vision models, so the ⊙ indicator lights up and image attachment works with them.

What’s New (v7.1.0)

Thinking toggle and Ollama offline detection.

Thinking Toggle – /think slash command (or transient T) controls whether reasoning models emit a thinking block. When disabled, models skip reasoning for faster, cheaper responses. Header shows ✦ (enabled) or ✧ (disabled).

What’s New (v7.0.0)

Plan mode — structured task planning with org-mode TODO integration.

Plan Mode – /plan activates a planning workflow where the LLM generates structured plans as org-mode TODO headings. Cycle states with S-<right> (TODO → DOING → DONE | SKIP). Execute steps with /plan-next (one at a time, with full tool calling) or /plan-all (all remaining). /plan-done marks a step complete; /plan-status shows progress; /plan-stop deactivates.
Tool Calling Progress – Status line and echo area now show live feedback: ⚒ Preparing write_file… during streaming, ⚒ Running write_file… during execution, ⚒ Tools complete when finished.
Enhanced =write_file= Tool – Reports file size and line count in both the echo area and tool results.

What’s New (v6.0.0)

Launch external frontends and comprehensive code quality improvements.

=/launch= Slash Command – Launch any Ollama model in an external terminal with a frontend tool (e.g. Claude Code). Runs ollama launch claude --model MODEL with auto-detected terminal emulator. Detects via $TERMINAL, $TERM_PROGRAM, xdg-terminal-exec, or PATH scan.
Launch in Model Management – Each model in C-c M now has a Launch button when frontends are detected.
Welcome Screen Indicator – Shows ⚡ launch: claude (via kitty) when launch frontends are available on PATH.
Code Quality – Fixed process/buffer leaks, added vectorp guards for API responses, replaced busy-waits, consolidated duplicate code, improved error handling across the package.

What’s New (v5.1.0)

Unguarded tool mode and MiniMax-M2.7 cloud model.

Unguarded Mode – /unguarded slash command (or transient G) bypasses ALL safety prompts, letting tools run to completion without interruption. No confirmation for unsafe tools, no large-result warnings, safe-mode overridden. Header line turns red and shows ☠ as a clear visual warning.
MiniMax-M2.7 Cloud Model – Added minimax-m2.7:cloud to ollama-buddy-cloud-models.

What’s New (v5.0.1)

Auto-scroll toggle for following streaming output.

Auto-Scroll Toggle – C-c g toggles ollama-buddy-auto-scroll, jumping to the end of the buffer when enabled. Header line shows ↓ when active. Works when toggled mid-stream.

What’s New (v5.0.0)

Tool calling for all providers — OpenAI, Claude, and Gemini API types all support function calling through ollama-buddy-provider-create.

Tool Calling for External Providers – All three API types (OpenAI-compatible, Claude, Gemini) now support tool calling. Tools are automatically converted from the internal format to each provider’s native schema at send time. This covers OpenAI, DeepSeek, Grok, OpenRouter, Copilot, Codestral, LM Studio, Anthropic Claude, and Google Gemini.

What’s New (v4.2.0)

Tool-augmented search — LLMs can now autonomously search the web and query local documents during tool calling.

=web_search= Tool – Models can autonomously search the web for current information during tool calling conversations. Uses the Ollama web search API; returns plain-text results (titles, URLs, snippets) directly as tool context.
=rag_search= Tool – Models can autonomously search local RAG indexes for relevant document chunks during tool calling. Searches all indexes by default, or a specific one by name.
Emacs 29.1 Minimum – Org folding now uses the org-fold text-property API directly, fixing thinking block folding issues on modern Emacs.
Thinking Fold Robustness – TAB to peek at a thinking block mid-stream now stays open instead of being immediately re-folded.
Welcome Screen – Extra commands under a folded ** More Commands heading; project path is a clickable dired link.
Model Management – Running models listed individually with per-model Unload buttons; [x]/[ ] checkbox removed.
=C-c E= – Toggle tool auto-execute from the chat buffer.

Recent Changes

Version	Summary
7.6.0	Dynamic cloud model discovery — `ollama-buddy-cloud-models` auto-fetched from `ollama.com/api/tags`; `/sync` force-refresh; `?` opens transient.
7.5.0	OpenCode Go subscription integration — single `n:` prefix exposes all 14 models with one shared API key; `ollama-buddy-provider-create` now accepts function forms for `:api-type` / `:endpoint` for per-model dispatch.
7.4.1	Response post-processing hook (`ollama-buddy-response-post-process-functions`) with numbered-list normaliser for dense cloud models; thinking-block point-anchor fix; Emacs Blog Reviewer user prompt.
7.4.0	Directory-scale project annotation — `/annotate-dir` batches source files through a structured-output schema, auto-applies thread-shaped annotations to `.simply-annotations.el`, context-aware, project-relative keys.
7.3.0	Direct agent launch — external tools invoked as standalone executables; standalone launch from transient/dired; detached terminal sessions; Oh My Pi (`omp`) as first direct agent.
7.2.0	Small model launch warning — agents prepend 20K+ tokens of system prompt; models below 7B now trigger a confirmation warning explaining potential performance impact.
7.1.4	Expanded `ollama launch` agent registry — Cline, Droid, OpenCode, OpenClaw, Pi now detected alongside Claude Code and Codex.
7.1.0	Thinking toggle (`/think`, `✦=/=✧` indicator) — control reasoning output; Ollama offline detection with clear messages.
7.0.0	Plan mode — structured task planning with org-mode TODO integration; tool calling progress feedback; enhanced `write_file` tool output.
6.0.0	`/launch` slash command — launch models in external terminals with frontends (e.g. Claude Code); auto-detects terminal; Launch button in Model Management; comprehensive code quality fixes.
5.1.0	Unguarded tool mode (`☠` indicator, red header line) — bypass all safety prompts; MiniMax-M2.7 cloud model.
5.0.1	Auto-scroll toggle (`C-c g`, `↓` indicator); works mid-stream.
5.0.0	Tool calling for all providers: OpenAI, Claude, and Gemini API types support function calling; HTTP error detection; Gemini native `systemInstruction`.
4.2.0	Tool-augmented search: `web_search` and `rag_search` built-in tools; Emacs 29.1 minimum; thinking fold robustness; welcome screen folded commands; model management running models section.
4.1.1	Recommended Models browser (`C-c L`), model letter ordering matches display, session load moved to `C-c f`.
4.1.0	Response property drawers (`:TOKENS:`, `:RATE:`, etc.), org file links for attachments, SVG context bar, removed token stats toggle.
4.0.0	Generic provider system (`ollama-buddy-provider-create`) — register any OpenAI/Claude/Gemini-compatible API with one call. Existing `require` wrappers still work but are deprecated.
3.7.1	Code quality, async model info, error cleanup, `/cd` directory switching, removed Fabric/Awesome.
3.7.0	RAG pause/resume with periodic checkpointing; new built-in tools (`search_files`, `get_region`, `get_buffer_info`, `describe_symbol`, `eval_elisp`).
3.6.0	RAG incremental updates (re-index only changed files) and cancellation support.
3.5.1	Curl backend refactor — unified request pipeline, shared stream processor, full feature parity with network-process backend (46% code reduction).
3.5.0	Project Init (experimental) — `/init` generates a cached project summary that auto-loads in future sessions; a step towards code-oriented LLM workflows.
3.4.1	Thinking Peek and Rewind — `TAB` peek at thinking mid-stream; `C-u C-u C-c C-c` to rewind conversation from any prompt heading.
3.4.0	Skills — Local Org-mode based AI personas and task-specific instructions; `/skill` slash command and `@skill(name)` inline syntax.
3.3.0	Smart Project Integration — `project.el` support for contextual file attachment and `/project` slash command.
3.2.2	Main transient menu reorganization (`C-c O`) — 4x2 compact grid with streamlined sub-menus and top-level History/Sessions.
3.2.1	Benchmark all models (`C-c u`) — one-command multishot across all models with per-model timeout; builds persistent token stats.
3.1.0	Inline ghost-text code completions (`ollama-buddy-completion`) — simple, manual Copilot-style completions with FIM context, thinking model support, fence stripping.
3.0.0	Generic OpenAI-compatible chat provider (`ollama-buddy-openai-compat`) — connect to LM Studio, llama.cpp, vLLM, Jan, or any OpenAI-API server.
2.9.2	`:destination` property for command definitions — per-command routing to `chat` or `in-buffer`, overriding the global toggle; all built-in commands and presets pre-annotated.
2.9.1	Welcome screen tips — `ollama-buddy-tips.el` shows a random usage tip on every fresh chat buffer; disable with `(setq ollama-buddy-show-tips nil)`.
2.9	Configurable RAG embedding backend — `ollama-buddy-rag-embedding-base-url` and `ollama-buddy-rag-embedding-api-style` let you point RAG at any `/v1/embeddings` server (e.g. LM Studio) instead of Ollama’s `/api/embed`.
2.8.1	In-buffer replace polish: inline diff view (`C-c d`, word-level smerge highlighting), automatic clean-output tone, code-fence stripping, `C-c W` toggle.
2.8.0	Initial in-buffer replace mode — stream rewrites directly into the source buffer; `C-c C-c` accept / `C-c C-k` reject.

See CHANGELOG.org for full history.

Installation

MELPA Simple

(use-package ollama-buddy
  :ensure t
  :bind
  ("C-c o" . ollama-buddy-role-transient-menu)
  ("C-c O" . ollama-buddy-transient-menu))

On first launch, press C-c O → I to install the bundled presets and user prompts for the full experience (see Installing Presets and User Prompts).

With Remote Providers (v4.0+ recommended)

(use-package ollama-buddy
  :ensure t
  :bind
  ("C-c o" . ollama-buddy-role-transient-menu)
  ("C-c O" . ollama-buddy-transient-menu)
  :config
  (require 'ollama-buddy-provider)
  (ollama-buddy-provider-create
   :name "OpenAI"
   :prefix "a:"
   :api-key (lambda () (auth-source-pick-first-password :host "ollama-buddy-openai" :user "apikey"))
   :endpoint "https://api.openai.com/v1/chat/completions"
   :default-model "gpt-4o"
   :models-endpoint "https://api.openai.com/v1/models"
   :models-filter (lambda (id) (string-match-p "gpt" id)))
  ;; add other providers similarly …
  )

With Remote Providers (legacy, still works)

(use-package ollama-buddy
  :ensure t
  :bind
  ("C-c o" . ollama-buddy-role-transient-menu)
  ("C-c O" . ollama-buddy-transient-menu)
  :custom
  (ollama-buddy-openai-api-key
   (auth-source-pick-first-password :host "ollama-buddy-openai" :user "apikey"))
  ;; add other providers similarly …
  :config
  (require 'ollama-buddy-openai nil t)
  ;; other cloud providers …
  )

With a Local OpenAI-Compatible Server (LM Studio, llama.cpp, vLLM…)

No API key needed for local servers. Use ollama-buddy-provider-create with :models-endpoint to auto-discover models:

(use-package ollama-buddy
  :ensure t
  :bind
  ("C-c o" . ollama-buddy-role-transient-menu)
  ("C-c O" . ollama-buddy-transient-menu)
  :config
  (require 'ollama-buddy-provider)
  (ollama-buddy-provider-create
   :name "LM Studio"
   :prefix "l:"
   :endpoint "http://localhost:1234/v1/chat/completions"
   :models-endpoint "http://localhost:1234/v1/models"))

Optionally also route RAG embeddings through the same server:

(setq ollama-buddy-rag-embedding-base-url "http://localhost:1234")
(setq ollama-buddy-rag-embedding-api-style 'openai)
(setq ollama-buddy-rag-embedding-model "nomic-embed-text")

Manual

git clone https://github.com/captainflasmr/ollama-buddy.git

(add-to-list 'load-path "path/to/ollama-buddy")
(require 'ollama-buddy)
(global-set-key (kbd "C-c o") #'ollama-buddy-role-transient-menu)
(global-set-key (kbd "C-c O") #'ollama-buddy-transient-menu)

Installing Presets and User Prompts

Role presets and user prompts ship with the package but need to be copied into your Emacs configuration directory to be editable. If they are missing, the chat welcome screen will show a reminder and the transient menu will offer an Install Extras entry.

From Emacs (recommended)

Press C-c O → I (or M-x ollama-buddy-install-extras). This copies the bundled directories from the installed package, or downloads them from GitHub if the package was installed without them.

Manual (shell)

cd ~/.emacs.d

# extract out the ollama-buddy-presets and ollama-buddy-user-prompts
curl -sL https://api.github.com/repos/captainflasmr/ollama-buddy/tarball/main | \
    tar -xzf - --transform=’s|[^/]*/||’ \
        --wildcards \
        ‘*/ollama-buddy-presets/*’ \
        ‘*/ollama-buddy-user-prompts/*’

The two directories installed are:

=ollama-buddy-presets= – Pre-defined role presets (developer, writer, tutor, etc.) as .el files.
=ollama-buddy-user-prompts= – Bundled system prompts organized by category (coding, writing, emacs, skills, etc.).

Exporting Skills for AI Agent Tools

User prompts can be exported as named skills for Claude Code, opencode, and other compatible agent tools. Each skill becomes a directory under ~/.claude/skills/ with a SKILL.md file containing YAML frontmatter, description, trigger phrases, and the prompt body as markdown.

M-x ollama-buddy-user-prompts-export-skills – Export all prompts as skills (with confirmation). With C-u prefix, prompts for custom directories.
M-x ollama-buddy-user-prompts-export-skill – Export a single prompt selected via completion.

Configure targets:

;; Primary target (default ~/.claude/skills/)
(setq ollama-buddy-skills-export-directory "~/.claude/skills/")

;; Optional secondary target (e.g. project-local .agents/skills/)
(setq ollama-buddy-skills-export-agents-directory
      (expand-file-name ".agents/skills/" (project-root (project-current t))))

My current configuration

(use-package ollama-buddy
  :demand t
  :bind
  ("C-c o" . ollama-buddy-role-transient-menu)
  ("C-c O" . ollama-buddy-transient-menu)
  :config
  ;; overall default model
  (setq ollama-buddy-default-model "glm-4.7:cloud")

  ;; ;; deprecated external LLM
  ;; (setq ollama-buddy-openai-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-openai" :user "apikey"))
  ;; (setq ollama-buddy-claude-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-claude" :user "apikey"))
  ;; (setq ollama-buddy-gemini-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-gemini" :user "apikey"))
  ;; (setq ollama-buddy-grok-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-grok" :user "apikey"))
  ;; (setq ollama-buddy-codestral-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-codestral" :user "apikey"))
  ;; (setq ollama-buddy-openrouter-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-openrouter" :user "apikey"))
  ;; (setq ollama-buddy-deepseek-api-key
  ;;       (auth-source-pick-first-password :host "ollama-buddy-deepseek" :user "apikey"))
  ;; (require 'ollama-buddy-openai nil t)
  ;; (require 'ollama-buddy-claude nil t)
  ;; (require 'ollama-buddy-gemini nil t)
  ;; (require 'ollama-buddy-grok nil t)
  ;; (require 'ollama-buddy-codestral nil t)
  ;; (require 'ollama-buddy-copilot nil t)
  ;; (require 'ollama-buddy-openrouter nil t)
  ;; (require 'ollama-buddy-deepseek nil t)
  
  ;; cloud / web-search keys (not provider-managed)
  (setq ollama-buddy-cloud-api-key
        (auth-source-pick-first-password :host "ollama-buddy-cloud" :user "apikey"))
  (setq ollama-buddy-web-search-api-key
        (auth-source-pick-first-password :host "ollama-buddy-web-search" :user "apikey"))
  (setq ollama-buddy-cloud-session-token "<session-token>"

  ;; Generic provider registration (replaces individual require files)
  (require 'ollama-buddy-provider)

  (ollama-buddy-provider-create
   :name "OpenAI" :prefix "a:"
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-openai" :user "apikey"))
   :endpoint "https://api.openai.com/v1/chat/completions"
   :models-endpoint "https://api.openai.com/v1/models"
   :models-filter (lambda (id) (string-match-p "\\(gpt\\|o[0-9]\\)" id)))

  (ollama-buddy-provider-create
   :name "Claude" :prefix "c:" :api-type 'claude
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-claude" :user "apikey"))
   :endpoint "https://api.anthropic.com/v1/messages"
   :models-endpoint "https://api.anthropic.com/v1/models")

  (ollama-buddy-provider-create
   :name "Gemini" :prefix "g:" :api-type 'gemini
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-gemini" :user "apikey"))
   :endpoint "https://generativelanguage.googleapis.com/v1beta/models/%s:generateContent"
   :models-endpoint "https://generativelanguage.googleapis.com/v1/models"
   :models-filter (lambda (id) (string-match-p "gemini" id)))

  (ollama-buddy-provider-create
   :name "Grok" :prefix "k:"
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-grok" :user "apikey"))
   :endpoint "https://api.x.ai/v1/chat/completions"
   :models-endpoint "https://api.x.ai/v1/models")

  (ollama-buddy-provider-create
   :name "Codestral" :prefix "s:"
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-codestral" :user "apikey"))
   :endpoint "https://api.mistral.ai/v1/chat/completions"
   :models '("codestral-latest"))

  (ollama-buddy-provider-create
   :name "OpenRouter" :prefix "r:"
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-openrouter" :user "apikey"))
   :endpoint "https://openrouter.ai/api/v1/chat/completions"
   :models-endpoint "https://openrouter.ai/api/v1/models"
   :extra-headers '(("HTTP-Referer" . "https://github.com/captainflasmr/ollama-buddy")
                    ("X-Title" . "ollama-buddy")))

  (ollama-buddy-provider-create
   :name "DeepSeek" :prefix "d:"
   :api-key (lambda () (auth-source-pick-first-password
                        :host "ollama-buddy-deepseek" :user "apikey"))
   :endpoint "https://api.deepseek.com/chat/completions"
   :models '("deepseek-chat" "deepseek-reasoner"))

  (require 'ollama-buddy-completion)

  (setq ollama-buddy-completion-model "qwen3-coder-next:cloud")

  ;; setup default custom menu for preferred models
  (ollama-buddy-update-menu-entry 'refactor-code     :model "minimax-m2.1:cloud")
  (ollama-buddy-update-menu-entry 'git-commit        :model "glm-4.7:cloud")
  (ollama-buddy-update-menu-entry 'describe-code     :model "minimax-m2.1:cloud")
  (ollama-buddy-update-menu-entry 'dictionary-lookup :model "minimax-m2.1:cloud")
  (ollama-buddy-update-menu-entry 'synonym           :model "minimax-m2.1:cloud")
  (ollama-buddy-update-menu-entry 'proofread         :model "minimax-m2.1:cloud")

  ;; dired integration
  (with-eval-after-load 'dired
    (define-key dired-mode-map (kbd "C-c C-a") #'ollama-buddy-dired-attach-marked-files)))

Key Bindings (Chat Buffer)

Key	Action
C-c C-c	Send prompt
C-c RET	Send prompt (alt)
C-c C-k	Cancel request
C-c m	Change model
C-c l	Pull new model
C-c O	Main transient menu
C-c ?	Open Info manual
M-p / M-n	Browse prompt history
C-c M	Manage models
C-c L	Recommended models
C-c i	Show model info
C-c U	Multishot
C-c N	New session
C-c f	Load session
C-c S	Save session
C-c Z	Sessions directory
C-c J	Toggle history
C-c X	Clear history
C-c H	Edit history
C-c C-a	Attach file
C-c C-d	Detach file
C-c a	Attachment menu
C-c 0	Clear all attachments
C-c C-s	Show system prompt
C-c C-r	Reset system prompt
C-c y	System prompts menu
C-c s	User prompts menu
C-c ~	Select response tone
C-c SPC	Toggle tool calling
C-c G	Toggle unguarded mode
C-c E	Toggle tool auto-execute
C-c Q	List registered tools
C-c P	Project menu
C-c p	Parameter menu
C-c t	Display parameters
C-c +	Settings menu
C-c r	RAG menu
C-c /	Web search menu
C-c x	Toggle streaming
C-c e	Switch backend (network/curl)
C-c !	Toggle airplane mode
C-c g	Toggle auto-scroll
C-c A	Auth menu (Cloud, Copilot)
C-c K	Exit/kill session

ollama API Support Tables

Core API Endpoints

API Endpoint	Method	Purpose	Supported	Notes
`/api/chat`	POST	Send chat messages to a model	Yes	Core feature with streaming, tools, vision
`/api/generate`	POST	Generate text without chat context	Partial	Used for inline completions and model unloading
`/api/tags`	GET	List available local models	Yes	Used for model management
`/api/show`	POST	Show model information	Yes	Model info display
`/api/pull`	POST	Pull a model from Ollama library	Yes	Async pull with progress
`/api/push`	POST	Push a model to the Ollama library	No	Not planned
`/api/delete`	DELETE	Delete a model	Yes	Model management
`/api/copy`	POST	Copy a model	Yes	Model management
`/api/create`	POST	Create a model from Modelfile/GGUF	Yes	GGUF import supported
`/api/ps`	GET	List running models	Yes	Shows loaded models in management buffer
`/api/embed`	POST	Generate embeddings from text	Yes	RAG module via `ollama-buddy-rag.el` (v2.5.0+)
`/api/version`	GET	Get Ollama version	Yes	Displayed on Model Management page
`HEAD /api/blobs/:digest`	HEAD	Check if blob exists	No	Not needed for current functionality
`POST /api/blobs/:digest`	POST	Push a blob	No	Not needed for current functionality
Cancel requests	Custom	Terminate ongoing request	Yes	`C-c C-k` in chat buffer

Chat API Parameters (`/api/chat`)

Parameter	Supported	Notes
`model`	Yes	Per-command or global via `ollama-buddy-default-model`
`messages`	Yes	Full conversation history support
`tools`	Yes	Tool calling via `ollama-buddy-tools.el` (v2.0.0+)
`think`	Yes	Reasoning model support with visibility toggle
`format`	Partial	JSON mode not yet exposed
`options`	Yes	Full implementation of all model parameters
`stream`	Yes	Toggleable with `ollama-buddy-toggle-streaming`
`keep_alive`	Yes	Set via `ollama-buddy-set-keepalive` or `ollama-buddy-keepalive`

Generate API Parameters (`/api/generate`)

Used by ollama-buddy-completion.el for inline ghost-text completions and for model unloading.

Parameter	Supported	Notes
`model`	Yes	Set via `ollama-buddy-completion-model` or current model
`prompt`	Yes	Buffer content before point (FIM prefix)
`suffix`	Yes	Buffer content after point (FIM suffix)
`images`	No	Vision uses chat endpoint instead
`format`	No	Not currently exposed
`options`	Yes	`num_predict` and `temperature` for completions
`system`	Yes	Completion system prompt (code-only output)
`template`	No	Uses model defaults
`stream`	Yes	Streaming enabled by default
`raw`	No	Not currently needed
`keep_alive`	Yes	Used for model unloading (`keep_alive: 0`)
`context`	N/A	Deprecated - using messages array instead

Model Options Parameters

Parameter Group	Parameters	Supported
Temperature Controls	`temperature`, `top_k`, `top_p`, `min_p`, `typical_p`	Yes
Repetition Controls	`repeat_last_n`, `repeat_penalty`, `presence_penalty`, `frequency_penalty`	Yes
Advanced Sampling	`mirostat`, `mirostat_tau`, `mirostat_eta`, `penalize_newline`, `stop`	Yes
Resource Management	`num_keep`, `seed`, `num_predict`, `num_ctx`, `num_batch`	Yes
Hardware Optimization	`numa`, `num_gpu`, `main_gpu`, `low_vram`, `vocab_only`, `use_mmap`, `use_mlock`, `num_thread`	Yes

Additional API Features

Feature	Supported	Notes
Streaming responses	Yes	Default mode, toggleable
Tool calling	Yes	Built-in tools + custom tool registration (v2.0.0+)
Vision/Images	Yes	Base64 image encoding for multimodal models
Structured outputs (JSON)	No	Not yet exposed
Reproducible outputs (seed)	Yes	Via parameters
Load/Unload models	Yes	Via model management buffer
Cloud models	Yes	Auto-fetched from `ollama.com/api/tags` (v7.6+)
Thinking/Reasoning models	Yes	`think` parameter with visibility toggle

Roadmap

Future Considerations

Structured Outputs (JSON Schema)

The format parameter supports JSON schema for structured responses.

Implementation considerations:

Define output schemas per command
Integration with org-mode tables
Structured data extraction workflows

Anthropic Messages API Compatibility

Ollama v0.14+ implements /v1/messages endpoint compatible with Anthropic’s API format.

Implementation considerations:

Option to route Claude-format requests through local Ollama
Could simplify ollama-buddy-claude.el or provide fallback
Support for tool_use and tool_result content blocks

Image Generation Support

Ollama v0.14+ supports local image generation with models like Z-Image Turbo (6B) and FLUX.2 Klein.

Implementation considerations:

Detect image generation models
Handle image output (save to file, display in buffer)
Support generation parameters (width, height, steps, seed)
Integration with org-mode inline images

Documentation

Manual – Full reference manual
Full README – Detailed README, tutorials, transient menu guide, and role setup
Changelog – Version-by-version changes
Ollama Buddy Blog – Blog articles for Ollama Buddy

Name		Name	Last commit message	Last commit date
Latest commit History 1,035 Commits
.agents/skills		.agents/skills
archive		archive
docs		docs
img		img
ollama-buddy-presets		ollama-buddy-presets
ollama-buddy-user-prompts		ollama-buddy-user-prompts
pi-extension		pi-extension
tests		tests
.gitignore		.gitignore
ACP.org		ACP.org
BUGS.org		BUGS.org
CHANGELOG.org		CHANGELOG.org
LICENSE		LICENSE
Makefile		Makefile
README-full.org		README-full.org
README-videos.org		README-videos.org
README.org		README.org
ollama-buddy-annotate.el		ollama-buddy-annotate.el
ollama-buddy-claude.el		ollama-buddy-claude.el
ollama-buddy-codestral.el		ollama-buddy-codestral.el
ollama-buddy-completion.el		ollama-buddy-completion.el
ollama-buddy-copilot.el		ollama-buddy-copilot.el
ollama-buddy-core.el		ollama-buddy-core.el
ollama-buddy-curl.el		ollama-buddy-curl.el
ollama-buddy-deepseek.el		ollama-buddy-deepseek.el
ollama-buddy-gemini.el		ollama-buddy-gemini.el
ollama-buddy-grok.el		ollama-buddy-grok.el
ollama-buddy-openai-compat.el		ollama-buddy-openai-compat.el
ollama-buddy-openai.el		ollama-buddy-openai.el
ollama-buddy-opencode.el		ollama-buddy-opencode.el
ollama-buddy-openrouter.el		ollama-buddy-openrouter.el
ollama-buddy-plan.el		ollama-buddy-plan.el
ollama-buddy-project.el		ollama-buddy-project.el
ollama-buddy-provider.el		ollama-buddy-provider.el
ollama-buddy-rag.el		ollama-buddy-rag.el
ollama-buddy-remote.el		ollama-buddy-remote.el
ollama-buddy-rewrite.el		ollama-buddy-rewrite.el
ollama-buddy-tips.el		ollama-buddy-tips.el
ollama-buddy-tools.el		ollama-buddy-tools.el
ollama-buddy-transient.el		ollama-buddy-transient.el
ollama-buddy-user-prompts.el		ollama-buddy-user-prompts.el
ollama-buddy-web-search.el		ollama-buddy-web-search.el
ollama-buddy.el		ollama-buddy.el

Folders and files

Latest commit

History

Repository files navigation

Ollama Buddy: Local LLM Integration for Emacs

Ollama Buddy

Want More Details?

Demos

In-buffer replacement demo

Tool Calling Demo

Advanced Tool Calling Demo

Quick Demo

Features

Alternative LLM Feature Comparison Matrix

Updates

What’s New (v7.4.2)

What’s New (v7.4.1)

What’s New (v7.4.0)

What’s New (v7.3.0)

What’s New (v7.2.0)

What’s New (v7.1.4)

What’s New (v7.1.3)

What’s New (v7.1.2)

What’s New (v7.1.1)

What’s New (v7.1.0)

What’s New (v7.0.0)

What’s New (v6.0.0)

What’s New (v5.1.0)

What’s New (v5.0.1)

What’s New (v5.0.0)

What’s New (v4.2.0)

Recent Changes

Installation

MELPA Simple

With Remote Providers (v4.0+ recommended)

With Remote Providers (legacy, still works)

With a Local OpenAI-Compatible Server (LM Studio, llama.cpp, vLLM…)

Manual

Installing Presets and User Prompts

From Emacs (recommended)

Manual (shell)

Exporting Skills for AI Agent Tools

My current configuration

Key Bindings (Chat Buffer)

ollama API Support Tables

Core API Endpoints

Chat API Parameters (/api/chat)

Generate API Parameters (/api/generate)

Model Options Parameters

Additional API Features

Roadmap

Future Considerations

Structured Outputs (JSON Schema)

Anthropic Messages API Compatibility

Image Generation Support

Documentation

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Chat API Parameters (`/api/chat`)

Generate API Parameters (`/api/generate`)

Packages