Open-source infrastructure for AI agents · early access

Define agents once.
Run them anywhere.
Prove what they did.

A universal manifest format for AI agents, an autonomous sprint orchestrator, a verified composition engine, and an agent-friendly CLI standard. Four independent open-source tools we're building in the open and using ourselves.

Pre-1.0 software. APIs may change. We're looking for early adopters and design partners willing to break things with us.

Watch the demo See the projects

4

Open-source projects

EUPL-1.2

EU public licence

370+

Composable stages

6

LLM runtimes supported

What we build

Four independent open-source tools that solve real problems for teams using AI agents

📝

AgentSpec

A universal file format for AI agents

Every agent framework today has its own configuration format. Change frameworks and you rewrite everything. AgentSpec defines an agent in a single portable .agent file that runs on Claude, Gemini, Codex, Aider, opencode, or local Ollama. The resolver picks the best available runtime from your environment — your agent definition stays the same.

pip install agentspec-alpibru and you're ready to go.

🔁

Caloron-Noether

An autonomous sprint orchestrator

Give it a goal. Caloron-Noether plans the tasks, assigns agents, generates code, opens pull requests, reviews them, runs the tests, merges, and learns from each sprint to do better in the next one. Real PRs, real tests passing, real Git history — proven end-to-end on a charging-window optimizer demo. Heading into our first internal pilots now.

Uses Noether for components that benefit from verifiable composition.

🧮

Noether

A verified composition engine

When you need provable, reproducible computation: every operation is content-addressed, type-checked before execution, and identical across machines and time. Use Noether for any data pipeline or workflow that demands auditability and verifiable behavior, not just AI agents.

Compiles to standalone binaries or browser WASM. Self-hostable as an HTTP service.

📡

ACLI

A standard for agent-friendly CLIs

If AI agents are going to use your CLI, it needs to teach itself to them at runtime. ACLI is the spec for CLIs that agents can discover and use without human-written documentation — structured help, machine-readable command trees, semantic exit codes, JSON envelopes. SDKs in Python, Rust, TypeScript, Go, Java and more.

Used by AgentSpec, Caloron-Noether, and Noether — every CLI we ship is ACLI-compliant.

🔗 Built to integrate. Every CLI we publish is ACLI-compliant so agents can discover it. AgentSpec and Caloron-Noether use Noether internally for components that benefit from verified composition. But each tool stands on its own — adopt one, several, or all four depending on what you need.

See it in action

One sprint, recorded end-to-end: agents build a charging-window optimizer, 14 tests passing, no human edits

Terminal recording: Caloron-Noether running an autonomous sprint that builds a charging window optimizer with SoC validation. Shows PO Agent planning, two worker agents implementing code and tests, two pull requests merged, 14 pytest tests passing, and the retro KPIs.

What you're watching

  1. The PO Agent reads the goal and plans the sprint into discrete tasks
  2. An agent implements the charging window optimizer with SoC validation, opening PR #3 against a real Git repo
  3. The PR is reviewed and merged
  4. A second agent writes the pytest tests for it — opens PR #4, merged
  5. The full test suite runs: 14 passed in 0.04s
  6. The retro collects sprint KPIs and feedback the next sprint will use to improve

Result: working production-ready code, real PR history, agent profile updated for sprint #2. Watch the full recording →

An .agent file

This is all you write. AgentSpec figures out the rest from your environment.

# researcher.agent
apiVersion: agent/v1
name: deep-researcher
version: 1.0.0

model:
  capability: reasoning-high
  preferred:
    - claude/claude-sonnet-4-6
    - gemini/gemini-2.5-pro
    - local/llama3:70b

skills:
  - web-search
  - cite-sources

behavior:
  traits:
    - cite-everything
    - never-guess

trust:
  filesystem: none
  network: allowed
  exec: none

Run it anywhere

# Resolves to claude-code if you have ANTHROPIC_API_KEY
$ agentspec run researcher.agent

# Same file — now resolves to gemini-cli
$ unset ANTHROPIC_API_KEY
$ export GOOGLE_API_KEY=...
$ agentspec run researcher.agent

# Same file — now runs locally on Ollama
$ unset GOOGLE_API_KEY
$ ollama pull llama3:70b
$ agentspec run researcher.agent

Same agent definition. Three different runtimes. The resolver figures out what's available and picks the best option from your preference list.

Projects

Open-source infrastructure for the AI agent era

01

AgentSpec

Universal agent manifest standard with resolver and signed portfolios.

Write a .agent file describing what you want — the resolver figures out which runtime, model, and tools to use based on what's installed. Agents accumulate signed portfolios across sprints, building verifiable CVs. Inheritance with hardcoded trust-restrict invariant. Works with Claude, Gemini, Codex, Aider, opencode, and Ollama.

Agent Manifest Resolver Signed CVs Multi-Runtime
EUPL-1.2 Open Source
Python 3.11+ pip install agentspec-alpibru
GitHub → Docs →
02

Noether

Verified composition platform for AI agents — type-safe, content-addressed.

Decompose problems into immutable, content-addressed stages. Two stages with the same hash are provably identical computations forever. The composition engine type-checks graphs before execution, supports multi-language stages (Rust, Python, JS, Bash), compiles to standalone binaries or browser WASM, and serves any composition as an HTTP API.

Composition Engine Content-Addressed Type-Safe Rust
SHA-256 Stage IDs
WASM + Native Build targets
GitHub → Docs →
03

Caloron-Noether

Autonomous sprint orchestrator — agents that ship code via Git.

Give it a goal. The PO Agent generates a DAG, HR Agent assigns skills, agents execute via sandboxed runtimes, supervisor monitors for stalls, retro produces signed memories that improve the next sprint. Real PRs, real reviews, real merges, real tests passing. Proven end-to-end with both Python orchestration and Noether-backed verification.

Sprint Loop DAG Orchestration Self-Healing Sandboxed
End-to-End PO → PRs → Retro
6 Frameworks Claude, Gemini, ...
GitHub → Docs →
04

ACLI

An agent-friendly CLI standard with multi-language SDKs.

ACLI specifies how command-line tools should expose themselves so AI agents can discover and use them at runtime — without pre-loaded schemas, external registries, or hand-written documentation. Structured help, machine-readable command trees via introspect, semantic exit codes, JSON output envelopes. SDKs in Python, Rust, TypeScript, Go, .NET, R, and Java enforce compliance automatically. Used by every CLI we ship.

CLI Standard Multi-Language SDK Self-Describing JSON Envelopes
7 SDKs Python, Rust, TS, Go, ...
EUPL-1.2 Open Source
GitHub → Docs →
05

Noether-Cloud Commercial · Beta

Hosted registry and infrastructure for the agent ecosystem.

HTTP registry server (Rust + Postgres), semantic search, and a curated stage catalog. The same infrastructure can be deployed in your VPC for private agent ecosystems. Public registry available at registry.alpibru.com. Enterprise features (SSO, audit logs, multi-tenant isolation, managed Kubernetes) are on the roadmap — talk to us if you need them.

Hosted Registry Postgres Semantic Search Self-Hostable
Beta Free tier available
Self-Host or hosted by us
Talk to us →

Design principles

Three opinions that shape what we build and what we won't build

Your agent definitions shouldn't be locked to one LLM vendor. AgentSpec resolves the same .agent file against whatever runtime is installed in your environment — Claude, Gemini, Codex, Aider, opencode, or local Ollama.

🌐

Vendor-neutral by default

6 runtimes supported

If agents start shipping code, "which agent built this and on what evidence?" needs to be answerable. We sign every memory and portfolio entry with Ed25519 — agent CVs that anyone can verify against the supervisor's public key.

🔐

Provable, not promised

Ed25519 signatures, content-addressed

Reuse over reinvention. Noether stages are content-addressed and type-checked before execution — two stages with the same hash are provably identical forever. Compose them instead of asking an LLM to re-invent the wheel.

🤖

Composition over generation

Powered by Noether

Where we are right now

Pre-1.0. Honest about what works, what's coming, what isn't built yet.

✓ What works today

  • Install all four projects from PyPI / GitHub Releases
  • Write and validate .agent manifests; the resolver picks runtimes correctly across 6 LLM CLIs
  • End-to-end autonomous sprint demoed on a charging-window optimizer (recorded above)
  • Push and pull .agent manifests against the public registry at registry.alpibru.com
  • Signed agent portfolios accumulate across sprints and verify against the supervisor's public key
  • Every CLI we ship is ACLI-compliant — agents discover commands via introspect

~ What's in progress

  • First internal pilots at our own companies (OTA + electromobility) — generating the case studies
  • Caloron CLI — solid tests and docs, needs more validation in real projects beyond the demo
  • Public registry: stable but still in beta, rate limits and SSO not yet configured
  • Real semantic embeddings on the registry (mock today)
  • Narrated video walkthrough (the asciinema cast is for engineers; we owe a 5-minute version)

✗ Not built yet

  • Enterprise features in Noether-Cloud (SSO, audit logs, multi-tenant isolation, SLAs)
  • Web UI for browsing the agent registry — CLI and API only for now
  • VS Code extension for .agent file editing and validation
  • External security audit (we welcome scrutiny — open the PR)
  • Hosted agent execution (you bring your own runtime today)

We're using all four projects ourselves. We'd rather have ten engineers find real bugs than a thousand stars on GitHub. If you try something and it breaks, please open an issue — that's the contract for now.

Looking for design partners

If you're experimenting with AI agents — building tooling, running them in production, or just curious — we'd like to talk. Bug reports, integration questions, harsh feedback, or "I'd use this if it did X" — all welcome.

💬

Get in Touch

Drop us a message — we typically reply within a day

🐙

Open Source

github.com/alpibrusl
📚

Documentation

agentspec docs