Mithril

"Mithril! All folk desired it. It could be beaten like copper, and polished like glass; and the Dwarves could make of it a metal, light and yet harder than tempered steel." — Gandalf

A multi-model orchestration engine. Combine any mix of LLM providers (Gemini, OpenAI, Anthropic, Groq, local GGUF) into a single Ollama-compatible API endpoint. Configure who does what in a YAML file, then point any AI tool at it.

---

What It Does

You define a fellowship — a team of AI models working together:

# .mithril/fellowship.yaml
name: "my-team"
controller:
  provider: local          # Free GGUF model routes requests
  model: qwen-1.5b

agents:
  - name: coder
    provider: gemini
    model: gemini-2.5-flash
    when: "coding tasks"
    tools: ["*"]

  - name: reviewer
    provider: openai
    model: gpt-4o
    when: "code review requested"
    tools: ["read_psi", "grep_files"]

Then you start the engine:

mithril serve

Now any Ollama-compatible client sees your fellowship as a model:

# From Junie, OpenCode, Open WebUI, or any Ollama client:
curl http://localhost:16180/api/tags
# → {"models": [{"name": "my-team:latest", "details": {"family": "mithril-fellowship"}}]}

---

Use Cases

Use Case	How
Backend for Junie	Point Junie at http://localhost:16180, select your fellowship as the model
Backend for OpenCode	Same — Ollama API compatible
Backend for Open WebUI	Add as Ollama connection
Backend for LangChain / LlamaIndex	Use OpenAI API at http://localhost:16180/v1/chat/completions
Backend for Jupyter / Python	pip install mithril-cli — run directly in notebooks and data workflows
MCP server for Claude Desktop	mithril mcp-stdio
Standalone CLI	mithril chat — built-in terminal interface
Docker service for teams	docker compose up — shared orchestration backend
Telegram bot	mithril telegram — chat via Telegram with same fellowship

---

Architecture

graph TB
    subgraph "Clients (any Ollama/OpenAI consumer)"
        J[Junie]
        O[OpenCode]
        W[Open WebUI]
        L[LangChain]
        C[Claude Desktop]
        T[Telegram]
        CLI[Mithril CLI]
    end

    subgraph "Mithril Engine"
        API[API Layer<br/>Ollama + OpenAI + MCP]
        ORCH[Orchestrator<br/>GGUF Classifier → Agent Routing]
        TOOLS[24 Built-in Tools<br/>File, Git, Web, Code, Terminal]
    end

    subgraph "Cloud API Providers"
        G[Gemini]
        GPT[OpenAI]
        A[Anthropic]
        GR[Groq]
    end

    subgraph "Local"
        LOCAL[Local GGUF]
    end

    subgraph "CLI Providers"
        K[Kiro]
        JN[Junie]
        COP[Copilot]
        ANY[Any CLI]
    end

    J -->|Ollama API| API
    O -->|Ollama API| API
    W -->|Ollama API| API
    L -->|OpenAI API| API
    C -->|MCP stdio| API
    T -->|Internal| API
    CLI -->|Internal| API

    API --> ORCH
    ORCH --> G
    ORCH --> GPT
    ORCH --> A
    ORCH --> GR
    ORCH --> LOCAL
    ORCH --> K
    ORCH --> JN
    ORCH --> COP
    ORCH --> ANY
    ORCH --> TOOLS

---

Installation

One-liner (Linux & macOS)

Downloads the universal zero-dependency static binary and automatically configures your shell PATH:

curl -fsSL https://raw.githubusercontent.com/GiacomoSaccaggi/mithril/main/install.sh | bash

Python / Jupyter / Conda (pip)

Ideal for Jupyter notebooks, Google Colab, SageMaker, cloud VMs, and Python data science stacks:

pip install mithril-cli

Inside a Jupyter notebook cell:

!pip install mithril-cli
!mithril --version

Homebrew (macOS & Linux)

brew install GiacomoSaccaggi/tap/mithril
# or:
brew tap giacomosaccaggi/tap && brew install mithril

Standalone Pre-built Binaries

Download the standalone binary for your architecture from GitHub Releases:

Platform	Architecture	Archive
Linux (Universal Static MUSL)	x86_64 / amd64	mithril-linux-x64.tar.gz
Linux (Universal Static MUSL)	ARM64 / aarch64	mithril-linux-arm64.tar.gz
macOS	Apple Silicon (arm64)	mithril-macos-arm64.tar.gz
macOS	Intel (x64)	mithril-macos-x64.tar.gz
Windows	x86_64	mithril-windows-x64.zip

Docker

docker run -d -p 16180:16180 ghcr.io/giacomosaccaggi/mithril:latest

Or via Docker Compose:

git clone https://github.com/GiacomoSaccaggi/mithril.git
cd mithril
docker compose up -d
# API available at http://localhost:16180

Build from source

git clone https://github.com/GiacomoSaccaggi/mithril.git
cd mithril && cargo build --release

---

Quick Start

1. Configure providers

# API keys — stored encrypted with Argon2id + AES-256-GCM
mithril config set gemini "AIza..."
mithril config set openai "sk-..."

# Or via environment variables (for Docker/CI):
export MITHRIL_KEY_GEMINI="AIza..."
export MITHRIL_KEY_OPENAI="sk-..."

2. Create a fellowship

mithril fellowship init
# Creates .mithril/fellowship.yaml with sensible defaults

3. Start the engine

mithril serve
# → http://localhost:16180 (Ollama + OpenAI + MCP)

4. Connect your tools

Junie / OpenCode / Open WebUI:

• Ollama URL: http://localhost:16180
• Model: select your fellowship name from the list

LangChain / custom:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:16180/v1", api_key="unused")
response = client.chat.completions.create(
    model="my-team",
    messages=[{"role": "user", "content": "Review this code"}]
)

---

Credentials in Docker

Mithril reads API keys in this priority order:

1. Environment variables (recommended for Docker): MITHRIL_KEY_<PROVIDER>
2. Encrypted config file: ~/.mithril/config.yaml (used by CLI)

# Docker Compose — set in .env file or environment:
MITHRIL_KEY_GEMINI=AIza...
MITHRIL_KEY_OPENAI=sk-...
MITHRIL_KEY_ANTHROPIC=sk-ant-...
MITHRIL_KEY_GROQ=gsk_...

No secrets are stored in the Docker image. Mount .mithril/fellowship.yaml for your agent configuration.

---

Fellowship Configuration

A fellowship defines who does what:

name: "code-team"
description: "Multi-model coding assistant"

controller:
  provider: local         # Routes requests (free, fast)
  model: qwen-1.5b
  context_window: 2       # Messages the router sees

agents:
  - name: worker
    provider: gemini
    model: gemini-2.5-flash
    role: "Fast coder — implements features"
    when: "any coding task"
    can_call: [reviewer]
    tools: ["*"]           # All 24 tools

  - name: reviewer
    provider: openai
    model: gpt-4o
    role: "Senior reviewer — catches bugs"
    when: "review requested or complex logic"
    can_call: []
    tools: [read_psi, grep_files, git_diff]

Agents communicate via the NEXT/TASK protocol:

• NEXT: DONE — task complete, return to user
• NEXT: reviewer + TASK: check auth.rs — delegate to another agent

---

The CLI (Optional)

Mithril includes a full-featured terminal interface:

mithril chat              # Interactive REPL with Tab completion
mithril chat --tui        # Full-screen TUI with panels
mithril exec "fix bug"    # Non-interactive (for CI/scripts)

Features: @file expansion, #agent routing, /commands, Plan/Build modes, undo/redo, session persistence, custom commands, hooks.

See docs/CLI.md for details.

---

API Endpoints

Endpoint	Protocol	Use
GET /health	—	Health check
GET /api/tags	Ollama	List models (includes fellowships)
POST /api/chat	Ollama	Chat completion
POST /api/generate	Ollama	Text generation
POST /v1/chat/completions	OpenAI	Chat completion
GET /v1/models	OpenAI	List models
POST /mcp	MCP	JSON-RPC tool calls

---

Provider Types

Mithril supports three types of providers:

Type	Examples	How It Works
Local GGUF	qwen-1.5b, qwen-14b, llama-8b	Direct inference via llama.cpp (free, private, fast for routing)
Cloud API	Gemini, OpenAI, Anthropic, Groq	HTTP calls to cloud LLM endpoints (pay-per-token)
CLI Tools	Kiro, Junie, Copilot, any CLI	Subprocess calls to local CLI tools that have their own model access

# .mithril/fellowship.yaml
name: "my-team"

controller:
  provider: local          # Local GGUF (free, used for routing)
  model: qwen-1.5b

agents:
  # Cloud API provider
  - name: coder
    provider: gemini
    model: gemini-2.5-flash

  # CLI provider (uses kiro-cli with its own auth)
  - name: reviewer
    provider: kiro
    model: claude-opus-4.6

  # GitHub Copilot CLI (2000 credits/month)
  - name: specialist
    provider: copilot
    model: gpt-5.4

  # Local GGUF (free, private, offline)
  - name: local-coder
    provider: local
    model: qwen-14b

CLI providers are useful when you have access to tools like Kiro, Junie, or GitHub Copilot with their own authentication and model access. Mithril orchestrates them as part of your fellowship without needing separate API keys. Each CLI tool has its own credit budget — use them strategically for complex tasks while Gemini handles the bulk.

Note on the controller: The controller (classifier/router) defaults to a local GGUF model which is free, fast (~100ms), and private. You can technically use any provider as controller (e.g., provider: gemini), but it's not worth the cost unless you have a very large agent structure where precise routing justifies paying per-classification.

---

24 Built-in Tools

File: read_file, write_file, edit_file, delete_file, apply_patch Terminal: run_terminal (sandboxed) Discovery: list_files, grep_files, find_file, file_stats, glob_files Git: git_status, git_log, git_diff, git_blame, git_branch Web: web_search, fetch_page Code: search_symbols, document_outline Knowledge: lore_write, lore_read Interaction: todo_write, question

---

Documentation

Document	Contents
docs/ARCHITECTURE.md	System design and module map
docs/TOOLS.md	All 24 tools with parameters
docs/CLI.md	Terminal commands and features
docs/API.md	HTTP endpoints reference
docs/PROVIDERS.md	Provider configuration
docs/SECURITY.md	Security model
docs/SESSION.md	Session persistence
docs/ENGINE.md	GGUF inference engine
CONTRIBUTING.md	Join the Fellowship

---

License

MIT