"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 backend. Combine any mix of LLM providers (Gemini, OpenAI, Anthropic, Groq, JEV, local GGUF) into a single Ollama-compatible API endpoint. Configure who does what in a YAML file, then point any AI tool at it.
Why backend-only? During the beta versions (pre-1.0), Mithril included a built-in terminal REPL, a full-screen TUI, and a Telegram bot. After extensive use, it became clear that tools like Junie and OpenCode are vastly superior as coding frontends — better UX, richer tool integration, and active development by dedicated teams. Starting with v1.0, Mithril focuses exclusively on what it does best: fellowship orchestration and multi-provider routing. You bring the frontend you love, Mithril is the engine behind it.
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 serveNow any Ollama-compatible client sees your fellowship as a model:
curl http://localhost:16180/api/tags
# → {"models": [{"name": "my-team:latest", "details": {"family": "mithril-fellowship"}}]}That's it. Point Junie, OpenCode, Open WebUI, LangChain, or any Ollama/OpenAI client at http://localhost:16180 and select your fellowship.
| 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 |
| Docker service for teams | docker compose up — shared orchestration backend |
graph TB
subgraph "Clients (any Ollama/OpenAI consumer)"
J[Junie]
O[OpenCode]
W[Open WebUI]
L[LangChain]
C[Claude Desktop]
end
subgraph "Mithril Engine"
API[API Layer<br/>Ollama + OpenAI + MCP + System 1]
ORCH[Orchestrator<br/>Zero-Shot Classifier (JEV / GGUF) → Agent Routing]
TOOLS[24 Built-in Tools<br/>File, Git, Web, Code, Terminal]
end
subgraph "Cloud API & Decision Providers"
G[Gemini]
GPT[OpenAI]
A[Anthropic]
GR[Groq]
JEV[JEV System 1]
end
subgraph "Local"
LOCAL[Local GGUF<br/>jevos-v4, qwen, llama]
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
API --> ORCH
ORCH --> G
ORCH --> GPT
ORCH --> A
ORCH --> GR
ORCH --> JEV
ORCH --> LOCAL
ORCH --> K
ORCH --> JN
ORCH --> COP
ORCH --> ANY
ORCH --> TOOLS
Downloads the universal zero-dependency static binary and automatically configures your shell PATH:
curl -fsSL https://raw.githubusercontent.com/GiacomoSaccaggi/mithril/main/install.sh | bashIdeal for Jupyter notebooks, Google Colab, SageMaker, cloud VMs, and Python data science stacks:
pip install mithril-clibrew install GiacomoSaccaggi/tap/mithrilDownload from GitHub Releases:
| Platform | Architecture | Archive |
|---|---|---|
| Linux | x86_64 / amd64 | mithril-linux-x64.tar.gz |
| Linux | 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 run -d -p 16180:16180 ghcr.io/giacomosaccaggi/mithril:latestOr via Docker Compose:
git clone https://github.com/GiacomoSaccaggi/mithril.git
cd mithril
docker compose up -dgit clone https://github.com/GiacomoSaccaggi/mithril.git
cd mithril && cargo build --release# 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-..."mithril fellowship init
# Creates .mithril/fellowship.yaml with sensible defaultsmithril serve
# → http://localhost:16180 (Ollama + OpenAI + MCP)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"}]
)Mithril reads API keys in this priority order:
- Environment variables (recommended for Docker):
MITHRIL_KEY_<PROVIDER> - 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_...
MITHRIL_KEY_JEV=... # Optional for JEV cloud endpointsNo secrets are stored in the Docker image. Mount .mithril/fellowship.yaml for your agent configuration.
A fellowship defines who does what:
name: "code-team"
description: "Multi-model coding assistant"
controller:
provider: local # Routes requests (free, fast)
model: jevos-v4 # Dedicated zero-shot decision router (~620MB, ~20-50ms)
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 userNEXT: reviewer+TASK: check auth.rs— delegate to another agent
Mithril supports four types of providers:
| Type | Examples | How It Works |
|---|---|---|
| Local GGUF | jevos-v4, 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) |
| Decision / System 1 | JEV (jevos-v4 via HTTP or local GGUF) |
Zero-shot agent selection and classification engine (POST /v1/systemone) |
| 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: jevos-v4 # Dedicated zero-shot decision router (~20-50ms)
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-14bCLI 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.
Note on the controller: The controller defaults to a local GGUF model (
jevos-v4orqwen-1.5b) which is free, fast (~20–100ms), and private.jevos-v4is purpose-trained for zero-shot candidate selection. You can also connect to a remote JEV service viaprovider: jev.
| 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 /api/embed |
Ollama | Embeddings |
POST /api/rerank |
Ollama | Reranking |
POST /v1/chat/completions |
OpenAI | Chat completion |
GET /v1/models |
OpenAI | List models |
POST /v1/systemone |
JEV | System 1 zero-shot decision & classification routing |
POST /mcp |
MCP | JSON-RPC tool calls |
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
- Credential encryption: API keys are encrypted at rest with Argon2id + AES-256-GCM
- Input redaction: Credentials and secrets in prompts are automatically masked before being sent to cloud providers
- Terminal sandbox: Blocks dangerous commands (
rm -rf /,sudo,curl | bash, etc.) - API token auth: Optional bearer token for the HTTP server (
mithril config set api_token <token>) - No telemetry: Zero data collection, zero phone-home
| Command | Purpose |
|---|---|
mithril serve |
Start the HTTP server (Ollama + OpenAI + MCP) |
mithril config |
Manage API keys and settings |
mithril fellowship |
Create and manage fellowship configurations |
mithril fellowships |
List all available fellowships |
mithril download-model |
Download GGUF models for local inference |
mithril scan |
Build the Palantír semantic index for the current directory |
mithril mcp-stdio |
Start MCP server over stdio (for Claude Desktop) |
mithril init |
Analyze codebase and generate project steering file |
MIT