Skip to main content
Glama

Ultron

One command line and one MCP server in front of every AI you can reach: cloud model APIs, subscription CLIs, local models and coding-agent harnesses. You hand Ultron a job; it picks the best available model or agent for that kind of job, falls through when one is down or out of quota, checks the result, and learns from the outcome.

Ultron grew out of AI Fleet (a task-card dispatcher) and two brain-structured rewrites of it (ai-neural-network and neural-fleet, merged 2026-11-14). The Python package is still named ai_neural_network; the commands are ultron and brain (same program).

Quick start

git clone https://github.com/d123d/ultron.git
cd ultron && uv venv && uv pip install -e ".[dev]"
uv run ultron slots        # what you can reach: providers, models, harnesses
uv run ultron jobs         # kinds of work and the models each one uses
uv run ultron job quick "Say hello"

Keys are read from ~/.secrets/.env (or the environment) by the names in src/ai_neural_network/workers.yaml. No key belongs in this repo: CI runs gitleaks over the full history on every push, and run logs are redacted. A provider with no key simply stays unavailable and its fallbacks answer.

Related MCP server: DZ23 Subagents Universal MCP

The three ideas

Provider slots. One slot per provider: endpoint or CLI login, concurrency, key state, and every model it serves (ultron slots --refresh pulls the live /models lists). Any listed model can be called without configuring it: ultron ask "..." --tier "nvidia:<model id>,". Local harnesses (Codex, OpenCode, Pi, Antigravity, Copilot, Claude Code, Muse, OpenHuman) have slots too.

Job slots. config/jobs.yaml maps a kind of work to an ordered list of models or to an agent harness. ultron job <kind> "<prompt>":

Kind

What runs

quick, bulk, extract, draft

fast free text models

reason, plan, research

strongest free reasoning and long-context models

verify

three providers answer, majority wins

vision

image QC (--image)

private

local model only; nothing leaves the machine

leader, best

subscription models

code, review, investigate, local-code

an agent harness works in --cwd, checked by --done-when

unattended

handed to OpenHuman's card queue

Board and daemon. ultron board add --text "<goal>" --cwd <repo> --done-when "<check>" queues agent work; --dep chains cards so they run one at a time; --hold parks one. ultron daemon (a logon task on Windows) takes one ready card a minute, runs it on the first free harness, fails over to the next harness on quota or crash, verifies with the card's checks and closes or reopens it. Cards left in progress by a dead daemon are reopened on restart.

Details and the list of defects found in live use: docs/SLOTS.md.

Under the hood: the network

Workers and model tiers are neurons in five regions (sensory, memory, association, prefrontal, motor) joined by weighted excitatory and inhibitory synapses. A task is a stimulus: it spreads, recruits a firing population, the population does the work, checks verify it, and the outcome rewires the weights (Hebbian learning with STDP-lite, neuromodulation by cost, learned inhibition, sleep consolidation). See docs/ARCHITECTURE.md and docs/BRAIN-MAP.md.

ultron status
ultron excite "fix the parser" --capability code --trace   # who would fire
ultron fire "summarise notes" --capability text            # propagate, run, verify, learn
ultron leader elect                                        # which executive model is available
ultron doctor                                              # health and quota of every worker and tier

Shipped seed weights are config/weights.json; what a machine learns is saved in its state directory, never in the repo.

Skills and tools

Worker prompts carry the working rules (smallest complete change, terse output, never fake a check) and, per task topic, the paths of matching skills from the machine's skill library plus its tool registry: 3D worlds and assets, animation, cinema, fiction, audio, images, documents, code. Topics are in config/skills.yaml; install skills with skills.sh. A few skills ship in skills/. See docs/SKILLS.md.

Providers

Direct tiers for OpenRouter, NVIDIA NIM, Ollama Cloud, Google, OpenAI, Anthropic, Mistral, Cohere, Groq, Cerebras, Together, Fireworks, DeepInfra, SambaNova, DeepSeek, xAI, Perplexity, Moonshot, MiniMax, Hugging Face, OpenCode Go; adapters for Azure, Bedrock, Vertex and Cloudflare; subscription routes through the claude, codex, agy, copilot and muse CLIs; local models through any OpenAI-compatible server. Matrix: docs/PROVIDERS.md; vendor quirks: docs/PROVIDER-BEST-PRACTICES.md.

For other agents

ACCESS.md: clone, MCP config and the worker contract. MCP tools include brain_job, brain_slots, brain_ask, brain_run, brain_fire, brain_doctor, brain_leader.

Honest limits

  • Built and tested on one Windows 11 workstation; CI runs the tests on Linux. Paths in workers.yaml, config/skills.yaml and the board settings point at that machine: edit them.

  • Job-slot model order is hand-set from live measurements, not learned yet.

  • Health state is safe across threads, not across processes.

MIT licence.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      F
      maintenance
      Enables AI assistants to intelligently select and switch between different AI models (OpenAI, Anthropic, etc.) within the same conversation based on task requirements. Provides a unified interface for accessing multiple AI providers through a single MCP tool.
      1
      107 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables users to delegate text and code tasks across multiple AI models with automatic failover, maintain shared mission memory and checkpoints across MCP harnesses, and coordinate parallel role-based subagents. It also supports model discovery, provider inventory, health checks, and stdio/HTTP tool access while leaving patches and tests to the harness.
      24
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Lets MCP clients route prompts and tasks to user-selected AI providers and models for coding, reasoning, planning, agentic analysis, image generation, and deep research, with configurable fallback order.
      13 npm
      MIT