ExAthena

Provider-agnostic agent loop for Elixir. Drop-in replacement for the Claude Code SDK that runs on Ollama, OpenAI-compatible endpoints (OpenAI, OpenRouter, LM Studio, vLLM, Groq, Together, llama.cpp server…), Google Gemini, or Anthropic Claude itself — with the same tools, hooks, permissions, and streaming semantics across every provider.

Status (v0.12): two front-ends land on top of the agent loop — a full-screen terminal TUI (mix athena.chat, built on ex_ratatui) with split message/details panes, real-time thinking, a live git diff Changes tab, mouse support, and a stop button; and a Phoenix LiveView web UI (mix athena.web) with session recall, fork, and a diff viewer. This release also adds a first-class llama.cpp provider and streams thinking/reasoning deltas as loop events. The TUI/web deps (ex_ratatui, phoenix, phoenix_live_view, bandit) are now optional, so the core library stays lean. See the v0.12.0 changelog for the full list.

The operational harness it builds on — file-based memory (AGENTS.md/CLAUDE.md), Claude Code-style skills, a five-stage compaction pipeline, 14 hook events, five permission modes, custom agents with git-worktree isolation, and append-only session storage with checkpointing — is documented under The operational harness.

Why

If you're using claude_code today and want to switch to a local Ollama model — or route per-task to OpenAI-compatible endpoints, or try Groq behind the same Elixir code — you don't want to rewrite every orchestrator. ExAthena is that abstraction layer. Pick a provider, run the same call, get back the same shape.

Install

The one-liner (Igniter auto-installs + writes sensible config):

mix igniter.install ex_athena

Or manually — add to mix.exs:

def deps do
  [
    {:ex_athena, "~> 0.12"},
    # optional — only needed for the Claude provider:
    {:claude_code, "~> 0.36"},
    # optional — only needed for the TUI (`mix athena.chat`):
    {:ex_ratatui, "~> 0.10"},
    # optional — only needed for the web UI (`mix athena.web`):
    {:phoenix, "~> 1.7"},
    {:phoenix_live_view, "~> 1.0"},
    {:bandit, "~> 1.5"}
  ]
end

The core agent loop has no Phoenix/TUI dependency; add the optional deps above only if you want mix athena.chat or mix athena.web. Then run mix ex_athena.install once to wire up defaults, or configure manually (see Configuration).

Quick start

# config/config.exs
config :ex_athena, default_provider: :ollama
config :ex_athena, :ollama, base_url: "http://localhost:11434", model: "llama3.1"

# anywhere
{:ok, response} = ExAthena.query("Tell me a joke")
IO.puts(response.text)

# streaming
ExAthena.stream("Explain quantum computing", fn event ->
  case event.type do
    :text_delta -> IO.write(event.data)
    :stop -> IO.puts("\n[done]")
    _ -> :ok
  end
end)

Swap the provider by changing one option:

ExAthena.query("hi", provider: :openai_compatible, model: "gpt-4o-mini")
ExAthena.query("hi", provider: :claude, model: "claude-opus-4-5")
ExAthena.query("hi", provider: :ollama, model: "qwen2.5-coder")
ExAthena.query("hi", provider: :gemini, model: "gemini-2.5-flash")

Attach images with the images: shorthand — same API across every provider:

png = File.read!("diagram.png")
{:ok, response} = ExAthena.query("Describe this diagram",
  provider: :ollama,
  model: "llava",
  images: [%{data: png, media_type: "image/png"}]
)

See the Multimodal guide for inline images, URL references, and per-provider notes.

Try it: mix athena.chat

Drop into an interactive chat REPL against a local Ollama model:

ollama serve &              # if not already running
ollama pull llama3.1        # any model you like

mix athena.chat
mix athena.chat --model qwen2.5-coder:14b --mode plan_and_solve

Tokens stream in real time. A pinned status line at the bottom tracks the current model, runner mode, iteration count, token usage, and cost. Slash commands switch state without restarting:

Defaults: :ollama provider, the model in config :ex_athena, :ollama, :model, :react runner, every builtin tool, permission_mode: :default.

Try it: mix athena.web

A browser-based chat UI with the same agent loop, accessible from any device on your network:

mix athena.web                  # http://0.0.0.0:4000
mix athena.web --port 8080      # custom port

The sidebar lets you switch provider, model, and mode without restarting. Features:

The web UI is a Phoenix LiveView application that requires no separate server process — mix athena.web starts everything in one command. Sessions are serialized with :erlang.term_to_binary and survive restarts. The JS bundle is served directly from the installed phoenix and phoenix_live_view hex packages, so there is no npm or esbuild step.

Note: phoenix, phoenix_live_view, and bandit are optional dependencies. Add them to your mix.exs (see Install) before running mix athena.web.

Providers

Pass a custom module that implements ExAthena.Provider directly if you have an endpoint that doesn't fit the above.

Configuration

config :ex_athena,
  default_provider: :ollama

config :ex_athena, :ollama,
  base_url: "http://localhost:11434",
  model: "llama3.1"

config :ex_athena, :openai_compatible,
  base_url: "https://api.openai.com/v1",
  api_key: System.get_env("OPENAI_API_KEY"),
  model: "gpt-4o-mini"

config :ex_athena, :claude,
  api_key: System.get_env("ANTHROPIC_API_KEY"),
  model: "claude-opus-4-5"

Resolution is tiered — per-call opts always beat app env:

ExAthena.query("…",
  provider: :openai_compatible,          # overrides default_provider
  base_url: "https://openrouter.ai/api/v1",  # overrides :openai_compatible, base_url
  api_key: System.get_env("OPENROUTER_API_KEY"),
  model: "anthropic/claude-opus-4.1")

Tool calls

ExAthena.ToolCalls handles both protocols and auto-falls-back between them:

• Native — OpenAI tool_calls arrays and Anthropic tool_use blocks. Parsed into canonical ExAthena.Messages.ToolCall structs.
• TextTagged — ~~~tool_call {json} fenced blocks embedded in assistant prose, for models without native support.

The agent loop (Phase 2) will pick the protocol based on the provider's declared capabilities, and fall back when the model gets it wrong.

The operational harness (v0.4)

The "1.6% reasoning, 98.4% harness" upgrade — built around the Claude Code paper's observation that production agent value comes from the operational scaffolding, not the loop itself.

File-based context. Drop an AGENTS.md (or CLAUDE.md) at the project root and ex_athena prepends it as user-context on every turn. Drop a SKILL.md with YAML frontmatter under .exathena/skills/<name>/ and its description joins the system-prompt catalog at ~50 tokens; the body loads only when the model writes [skill: <name>]. See the memory + skills guide.

Five-stage compaction. The default Compactor.Pipeline runs cheapest-first: budget reduction (truncate huge tool results) → snip (drop stale ones) → microcompact (collapse runs of identical calls) → context collapse (read-time-only projection) → LLM summary. When a provider returns "context too long" the pipeline forces every stage and retries.

14-event hook surface. Every transition in the loop is observable + interceptable: SessionStart/End, UserPromptSubmit, ChatParams, Stop, StopFailure, all the *ToolUse* variants, PermissionRequest/Denied, Subagent*, three compaction events, Notification. Hooks can {:inject, msg} to add context or {:transform, prompt} to rewrite the user's message. See hooks reference.

Five permission modes. Add :accept_edits (auto-allow file edits, still prompt for bash) and :trusted (skip prompts; with optional respect_denylist: false for full YOLO) on top of the existing :plan / :default / :bypass_permissions. The denylist always wins, including in bypass — that's locked in a doctest. See permissions.

Subagents v2. Define custom agents in .exathena/agents/<name>.md with frontmatter (tools, permissions, mode, isolation); spawn by name via agent: "explore". Optional git-worktree isolation creates an isolated checkout per subagent (with safety checks + graceful fallback). Sidechain transcripts persist the full subagent conversation to disk so the parent only spends tokens on the final text. Three builtin definitions ship: general, explore, plan. See agents + subagents.

Storage + checkpoint. ExAthena.Sessions.Store is an append-only event log behaviour with two stores: in-memory (default) and ETS-buffered JSONL with periodic flush. Sessions emit :user_message / :assistant_message / :tool_result events; Session.resume/2 rebuilds the message history from any store. File-checkpoint snapshots fire before every Edit / Write, and Checkpoint.rewind/3 restores files + truncates the session log to a chosen UUID. See sessions + checkpoints.

Guides

• Getting started
• Providers
• Multimodal (vision)
• Gemini setup
• Tool calls
• The agent loop
• Tools (incl. tool-result split)
• Memory + skills — v0.4
• Compaction pipeline — v0.4
• Hooks reference — v0.4
• Permissions — v0.4
• Agents + subagents — v0.4
• Sessions + checkpoints — v0.4

License

Apache 2.0 — see LICENSE.