akaion-ai/annona

---

Rome imported its grain. Sicily, Africa, Egypt — the city could not feed itself, and everyone knew what that meant: whoever controlled the ships controlled Rome. So the Republic, and then the Empire, refused to leave it to the market. The cura annonae was a permanent office with a prefect at its head, and its job was to decide where the grain came from, which route it took, which granary held it, and who received it — and to keep the record of all four.

Your organisation is now in that position with compute. Every agent you deploy sends your material somewhere, and “somewhere” is currently decided by whichever provider a developer typed into a config file eighteen months ago.

Annona is the office. It is a daemon you install where your data already lives, and for every single step of every agent run it decides where that step is allowed to execute, enforces the decision, and writes it down.

That refusal is the product. A gateway in the same situation would have quietly failed over to the frontier API and returned a good answer.

The argument nobody was having

You have been told to pick one of three architectures: run models on-prem (private, capped by your hardware), call frontier APIs (excellent, and your material leaves), or run your own weights in your own cloud (a fine compromise that costs you an MLOps team).

The industry argues about which column wins. That argument is the mistake.

The right column is a property of the request, not of the company.

Summarising a public tender is not the same problem as reasoning over a client’s medical file, and the second does not become safe because procurement signed a DPA. One organisation needs all three columns, chosen per step, ten thousand times a day, by something that can prove afterwards what it chose.

Nothing in the stack does that:

The claim, stated so you can falsify it. Annona is the first open-source project in which the placement of every inference and every tool call is a policy decision, enforced by the runtime and verifiable after the fact. Find a project that does this and we will say so here.

How it works

Annona is deliberately not the thing that decides what to do. Planning and memory stay in the control plane — ours, or yours over three HTTP endpoints. The component that touches your data is the one you can read, and it is small enough to read. That is the entire trust argument.

Four rules give the picture its teeth:

Fail closed. No permitted substrate is available → the step is held. Not downgraded, not rerouted, not “best effort”.

Failover may cost latency, money or model quality. Never jurisdiction. This is the one sentence that separates a kernel from a gateway with a fallback list.

Contamination is monotone. Once a transcript has touched restricted material it stays restricted. Tools executing locally is not the same as data staying local — the transcript is the leak, and it travels with the next inference.

Outbound only. No listening port on the internet. Nothing to firewall, no inbound rule to request from a customer’s IT department — historically the step where sovereign deployments die.

The same plan, two different verdicts

Ask it to compare a client file against your case law:

Now ask the same question about a public tender document. Steps 1 and 3 are placed on the frontier model, the answer is better, and it costs a tenth as much.

Same code. Same plan. Different policy verdict. That is the product.

What that looks like

Every answer carries where it ran and under which rule. Nothing is inferred from a log afterwards — the decision is made before the step runs, and shown with it:

The same question under a policy that permits only the smaller model. Same code, same plan, different placement — and the answer says so rather than leaving you to wonder which model produced it:

And with nothing the policy permits available, the outcome the whole design is for. Not a fallback, not a degraded answer from somewhere else:

The Perimeter view is the policy as the runtime reads it, which is not always what somebody thinks they wrote:

What it can actually do

The kernel decides placement; these are the things it places. All of it predates the perimeter and none of it was dropped — what changed is that a tool the policy does not name no longer runs.

The shipped policy enables the three read-only tools. shell and browser are left off on purpose and the template says why: a shell has no path argument, so enabling it is an all-or-nothing decision, and a browser reaches the network, which is an egress this policy cannot yet classify.

Skills, with a jurisdiction

A skill is what Anthropic’s Agent Skills are — a folder with a SKILL.md, front matter plus prose, loaded only when the model asks for it. Same format on purpose: a skill written for one runtime still works in the other.

Annona adds one field, and it is the only interesting part:

---
name: image-report
description: Read one or more images and produce a structured, factual report.
requires: [vision]
tools: [document_reader, explorer]
pins: local          # ← the run is confined the moment this is loaded
---

pins: local is enforced by the kernel, not by a sentence at the bottom of the instruction asking nicely. The moment the model loads that skill the working set is restricted, and no later turn can be placed on a frontier API — whatever the prompt looks like by then. That is the difference between a prompt library and a capability system.

The shipped set is deliberately generic in name and specific in effect:

annona skills                      # installed · allowed · usable here
annona skills --show image-report  # read the instruction yourself

Claude’s skills work here

The format is identical, so anything from anthropics/skills — or anything you already wrote for Claude Code — installs unchanged:

annona skills-install ~/Downloads/pdf          # a folder
annona skills-install pdf                      # or by name, from ~/.claude/skills
annona skills                                  # see what landed

Two things happen on the way in, and both are the point:

An imported skill is pinned to the perimeter. A skill is an instruction your agent will follow — a supply-chain dependency that happens to be prose. One you did not write runs inside your walls until you read it and say otherwise with --trust. Provenance goes into the front matter; the body is copied byte for byte, because silently editing somebody’s instruction would be its own kind of supply-chain problem.

Installed is not enabled. The policy still has to name it. Copying a file into a directory is not a decision about what your agents may do.

Skills that bundle scripts/ are copied whole and the install says so plainly: those scripts will not run unless your policy allows the shell tool, which by default it does not. The instructions still work; the automation in them does not.

Skills are default-deny like tools: one that the policy does not name is never offered, and asking for a disabled skill is answered exactly like asking for one that does not exist. Your own live in ~/.annona/skills/ and override the shipped ones by name, so a practice can encode its house style without forking anything.

image-report produces observations for a professional to interpret. It is not a diagnostic device, it does not conclude, and the instruction says so to the model as well as to you — software intended for diagnosis is regulated as a medical device, and calling it something else does not change that.

Built on the Italian open-source stack

Annona is a kernel, not a monolith: it decides where work runs and delegates everything else. Two of the three things it delegates to are Italian open-source projects, and that is a choice rather than a coincidence — a sovereignty story told with somebody else’s entire stack is a slide, not an argument.

Neither is vendored. datapizza is a dependency with an ADR explaining the choice; rizzo-pii is reached over its documented HTTP contract by an adapter in runner/capability/redactors/, attributed in NOTICE, and the layering means the decision layer cannot import it even by accident.

The division of labour is clean, and it is the reason the two fit:

Together they add a fourth answer to the three the perimeter already had. When a step is too sensitive for every available substrate, instead of only holding it, the policy can ask for redaction:

redaction:
  provider: rizzo-pii
  endpoint: http://127.0.0.1:5005
  labels:                       # its 22 categories → your three classes
    CF: restricted
    PIVA: restricted
    CATASTO: restricted
    FULLNAME: internal
  on_error: hold                # a redactor that is down stops the step

rules:
  - match: { class: restricted }
    allow: [local-gpu]
    on_unavailable: redact      # replace the identifiers rather than refuse

Il Sig. Mario Rossi, C.F. RSSMRA85T10A562S, chiede una proroga.
  ↓  locally, by rizzo-pii
Il Sig. [FULLNAME_1], C.F. [CF_1], chiede una proroga.
  ↓  reclassified public → cleared → frontier model
Il Sig. [FULLNAME_1] ha tempo fino al 15 marzo.
  ↓  re-identified here, from a mapping that never left
Il Sig. Mario Rossi ha tempo fino al 15 marzo.

Three properties make this a control rather than a hope, and each is a test:

• the redacted text is reclassified from scratch. A redactor that missed an identifier produces text that merely looks safe; if anything is still there, the step is held. The redactor is the instrument, the perimeter is the authority.
• a redactor outage holds the step by default, because a control that disappears under stress is not a control.
• the mapping never leaves the process and never enters the ledger — the record says two identifiers of kind CF and FULLNAME were replaced, and nothing more.

Pseudonymous is not anonymous. Under GDPR, replacing a name with a stable token leaves personal data: the mapping exists and re-identification is possible by design. This reduces exposure, it does not remove the need for a lawful basis, and Annona says so rather than implying otherwise.

Any redactor satisfying the same small protocol can be wired in its place — the policy names a provider, and the decision layer never learns which one.

Get it

The desktop app

A tagged release builds all four targets and publishes a GitHub Release with the bundles attached. They are unsigned during beta, and on macOS that is not a dialog you can click through: since macOS 15 an un-notarised app that carries a quarantine flag is reported as damaged and moved to the Trash. Clear the flag on the disk image before opening it —

xattr -dr com.apple.quarantine ~/Downloads/Annona_*.dmg

— and drag it in as usual. On Windows, SmartScreen → More info → Run anyway. Full detail, and the two routes that avoid this entirely, in Install.

The app is the same daemon with a window around it — the CLI, the policy and the ledger are identical.

On first launch it asks the three questions the policy answers, rather than defaulting silently and leaving you to find policy.yaml later:

Each option says what it means rather than what it sets, because “adds a substrate capped at public” describes YAML and “the hosted provider can only ever see material classified public” is what somebody is agreeing to. annona setup asks the same three in a terminal, from the same profiles.

From source, in sixty seconds

git clone [email protected]:akaion-ai/annona.git
cd annona
make setup          # venv + dependencies
make demo           # a real agentic run — no credentials, no network

make demo is the fastest way to understand this repository. It drives a real agentic loop — real tool execution against real files, real policy checks — from a scripted backend, so it needs no API key and opens no socket. It shows one task the policy permits and one it refuses:

1 · a task the policy permits
  1. ok      explorer         {'operation': 'map', 'path': '…/documents'}
  2. ok      document_reader  {'path': '…/reports/q1_report.txt'}
     answer  Q1 2026: 142 pratiche aperte, 98 chiuse, 412.000 EUR…

2 · a task the policy refuses
  1. denied  filesystem       {'operation': 'read', 'path': '~/.ssh/id_rsa'}
     → {'error': 'Permission denied for tool: filesystem'}

Nothing left the process. The same run is a CI gate on every push (python -m runner.demo --check), so that claim cannot rot.

Then:

annona setup        # the policy, in three questions — or --yes for the defaults
annona doctor       # check this machine can actually run a step
make run            # daemon + local UI on 127.0.0.1:7070

No account required, ever. annona doctor is the one to run when something is wrong: it names what is missing, including the case a liveness probe cannot see — a runtime that is up with the model your policy names not pulled.

As an appliance, with a real model

make up                                                # kernel + Ollama, with this host's GPU
docker compose exec annona annona doctor               # the machine, and the model that fits it
docker compose exec ollama ollama pull qwen2.5:14b     # or what doctor said fits
docker compose exec annona annona setup --yes --endpoint http://ollama:11434 --model qwen2.5:14b
make verify                                            # the acceptance run

make verify plants a canary in a client file, lets a real agent read it, and checks the nine things a customer’s auditor would:

  pass  the local runtime answers
  pass  the model called the tool
  pass  reading a client file made the run restricted
  pass  no payload reached the frontier substrate
  pass  leak rate is zero
  pass  every inference was placed on-prem
  pass  the ledger chain verifies
  pass  the run produced an answer
  pass  with the GPU down, restricted work is held (not rerouted)

The last check is the commercial one. Deployment, sizing and the DGX Spark specifics are in deploy/README.md.

Operating it

annona policy show             # the policy as the runtime understands it
annona policy test restricted  # where would restricted work go, right now?
annona substrates              # what is registered, where, and whether it is up
annona why step_7f3a           # reconstruct one decision from the ledger
annona verify                  # check the chain, offline, contacting nobody
annona audit --held            # every refusal, with its reason

The policy is a file you own

# ~/.annona/policy.yaml
classes:
  restricted:                      # never leaves the walls
    paths:    ["/mnt/pratiche/**", "~/clienti/**"]
    patterns: ['[A-Z]{6}\d{2}[A-Z]\d{2}[A-Z]\d{3}[A-Z]']    # codice fiscale
  internal:
    paths:    ["~/Documents/**"]
  public:
    default: true

rules:
  - match: { class: restricted }
    allow: [local-gpu]
    on_unavailable: hold           # the whole point. No silent downgrade.
  - match: { class: internal }
    allow: [local-gpu, eu-cluster]
    on_unavailable: queue
  - match: { class: public }
    allow: [local-gpu, eu-cluster, frontier]
    prefer: cost

A DPO can read that in one sitting, which is the design constraint. Full schema, placement algorithm and the state machine behind it: docs/design/hld.md.

You can edit it in your editor, or in the app under Perimeter → Edit. Being editable from a window is a real widening of what that surface can do, so three things hold it up: the replacement is parsed before it is written (a policy that does not load would stop enforcement, and this must not be able to cause that), the previous file is always copied aside, and every change is appended to the same hash-chained ledger as the decisions.

That last one is what makes it defensible rather than merely convenient — widen the perimeter, run something, narrow it back, and the record reads:

A pairing token does not grant this. Pairing lets a web app run steps here; an origin that could also rewrite the policy would hold every permission the perimeter exists to withhold, so the write routes refuse anything but this machine.

Optional: send it work from Agents Studio

A machine you own can take jobs from Agents Studio: someone in your organisation picks the machine, writes the task and closes the browser. The machine dials out for work, runs it under its own policy on its own models, and sends back only what that policy allows. Nothing talks to Studio until you enroll, and either side can end it.

Four jobs sent from Studio to a MacBook on the Akaion organisation, run on its qwen2.5:7b: a public question comes back; a letter carrying a codice fiscale stays on the machine (Studio never receives the answer); one cancelled while running; one sent with the local model switched off, which the policy held rather than reroute:

annona link enroll ann_enr_… --name dgx1   # the code from Studio → Machines → Add machine
annona link serve                           # or as a service; outbound HTTPS only
make verify-link ORG=                 # seven checks against the real Studio, then revoke

The whole story — what crosses the wire, the release ceiling, running it detached on a DGX Spark — is in Link a machine to Agents Studio.

Where it runs

Three topologies, one binary, one release. They differ in configuration — which substrates are registered and what the policy permits — never in code. A fork per deployment is how sovereignty claims rot, so there isn’t one.

Detached means detached: no account, no remote host, no outbound connection at all. It is the configuration we expect a fork to start from.

On a DGX Spark the appliance runs Annona and vLLM under one compose file (--profile vllm), with the daemon unprivileged and only vLLM touching the GPU. Two things appliance vendors usually skip, stated up front: every image must be linux/arm64 + CUDA 13 — x86 images silently do not run on a GB10, so the release matrix builds and tests both — and the real ceiling is memory bandwidth, not the 128 GB. An AMD Ryzen AI Max (Strix Halo) box takes the same compose file with the AMD overlay; make up picks the right one, and Hardware says which model fits and what has actually run where. See deploy/README.md and HLD §7.2, including what GPU attestation does not buy you on that hardware.

What is built, and what is not

This project publishes its gaps, because a perimeter you cannot verify is a slogan.

Built and under test. Classification (paths, symlink targets, content patterns, and paths named in a prompt), the monotone working set, a default-deny tool gate, the placement engine with its conformance matrix, substrate health with a circuit breaker, failover that cannot widen the permitted set, locally produced briefs that are reclassified before they may cross, a hash-chained ledger with verify / why / audit, an arm64 + amd64 container image, and a nine-check acceptance run for a new appliance.

$ make contracts
L0 kernel does not depend on outer layers                    KEPT
L3 agent depends only inward on L0                           KEPT
L3 agent and L1 capability do not know about each other      KEPT
L0 kernel imports no provider SDK                            KEPT
L3 agent loop imports no provider SDK                        KEPT
L2 policy depends only inward on L0                          KEPT
L2 placement depends on policy and the kernel, never outward KEPT
L2 audit depends on nothing but the kernel                   KEPT
L2 policy, placement and audit cannot reach an L1 adapter    KEPT
L2 imports no provider SDK                                   KEPT

The last four are why “failover cannot widen the permitted set” is structural: a shortcut from the decision layer to an adapter is not a code review away, it is a build failure.

Not built. Stated as plainly as the rest:

Each has a metric and a target rather than a promise — HLD §9 and docs/research/index.md, where negative results get published too.

The trust boundary

Annona is open source on purpose. It is the component that sees your material, so it must be the component you can audit — and replace.

Point it at your own backend by implementing three endpoints — verify a token, receive pushed notes, optionally serve inference — or set ai.provider: local and skip the third entirely. Details in the trust boundary section of the HLD.

Documentation

Contributing

make check            # lint · types · contracts · tests — exactly what CI runs
make docs-serve       # documentation at 127.0.0.1:8000
make                  # list every target

Two things worth knowing before your first PR:

• Do not mock a vendor SDK in tests. Use the echo backend and drive the loop through its ports — see tests/test_agent_loop_unified.py.
• If you touch the trust boundary — permissions, placement, egress, the ledger — say in the PR description what a reviewer should check to convince themselves the boundary still holds. Not “I tested it”: what to look at.

CONTRIBUTING.md · SECURITY.md · CHANGELOG.md

The Firebase Web API key in auth.py is the project’s documented public default; Firebase Web SDK keys are designed to be client-visible.

---

---

Deploying this in an organisation? labs.akaion.com — the kernel is open and always will be. What a company usually wants alongside it is the policy written against their own folders, the hardware, and somebody who answers when it breaks.