Agentic-Nets

A durable runtime for AI agents, people, and automated work.

Watch the product preview: personas, nets, and the live Studio in motion, before you read a single line about Petri nets.

What you can do with it

Agentic-Nets runs personas, teams, and automated processes that keep existing after the chat window closes:

• A specialist that remembers. A health coach that knows every week of your training, a domain expert that carries your product's real context, a support persona that can still explain the decision it made last month.
• A team with real handoffs. Product manager, architect, developer, reviewer, release guardian: each one a persona with its own tools, context, and authority, passing work through explicit states instead of one long prompt.
• Work that continues while you sleep. Put a lane on a cron or an interval and it keeps firing: collect sources overnight, digest a backlog every morning, probe an endpoint every five minutes and speak up only when it breaks.
• Deterministic steps beside AI steps. Seven transition kinds, so a shell command, an HTTP call, or a plain data transform lives in the same process as the reasoning, and the parts that never needed a model stop paying for one.
• Boundaries and approvals you can point at. Capability flags, tool allowlists, credential scopes, and approval gates decide what a persona may do before it does it, not after.
• A history that answers questions later. Every token, fire, tool call, and emission is event-sourced, so "what did this process actually do over the last three days, and where did it get stuck" is a query rather than a guess.
• Whole systems you can reuse. Publish a persona, a team, a context system, a tool, or a complete application to NetHub and install it into another workspace, dependencies bundled and credentials scrubbed.

How easy it has become

You do not have to draw a net to get one. Connect the MCP client you already use, Claude Code, Codex, Claude Desktop, or Cursor, and say what you want in plain language:

"Create a health coach that checks in with me every Monday."

"Start a safe product team for this repository."

"Give me a research analyst that reads these three sources overnight."

The client builds the places, transitions, schedules, and personas through MCP, tells you what it armed, and the system is running when it finishes answering. Watching it is just as direct: the bundled Studio shows the live net, its tokens, its event story, and every handoff as it happens. Changing it is another sentence, and the change is versioned and inspectable like everything else.

Observability you can act on, not just look at

Every committed change is an event, not a log line. The data engine records each tree mutation with its causal ids, and the master keeps a story journal of fires, outcomes, and errors on disk, one file per model per day. Three layers, each with a stated durability, and every tool tells you which one it is reading.

That turns the question most systems cannot answer into an ordinary query:

• "What did this transition do over the last three days, and where did it get stuck?" transition_history joins the master's stories with the structural mutations by exact fire id, and focus:"failures" returns the last correlated failure together with the engine's own binding diagnosis.
• "Who created this token, who changed it, who consumed it?" token_lineage walks it back through the retained event blocks.
• "What is happening right now?" events_wait long-polls the live event line with honest cursors, so a restart appears as a reset rather than a silent gap.

Retention is a setting rather than a side effect: the data engine keeps structural history per model (historyRetentionDays), and the story journal keeps narratives for its own configured window.

The point of keeping history is what you do with it. An agent can read it back and improve the system that produced it. The domain-neutral Model Steward reviews a whole model, its bottlenecks, failures, rework, wait time, and spend, then proposes changes that are approval-gated and versioned like any other change. Crystallization goes further: reasoning that has proven itself collapses into deterministic transitions, so a process gets cheaper, faster, and more predictable the longer it runs. A net that has been working for a month should not look like the one you started with.

One caveat the docs repeat rather than hide: retention windows are finite, and absence of evidence in a layer is not proof that something never ran. See docs/observability.

Command transitions and executors: real work on real machines

Map, HTTP, LLM, and agent lanes run on the master. A command lane is the one that leaves the building: it is dispatched to an executor, a small service that polls for work over an outbound connection only. Nothing has to reach in, so an executor can sit behind a firewall or NAT, on a build machine, on your laptop, or in another cloud, and still be part of the same net.

• Pick the machine per lane. action.executorId names one executor; "*" offers the work to every executor that polls and the first token reservation wins. Each executor carries its own model allowlist.
• Scoped credentials. Executors authenticate with a dedicated OAuth2 client whose JWT scope permits the polling protocol and nothing else.
• Typed results. Exit code, stdout, stderr, and duration come back as a token, so a failing build routes down an error path instead of disappearing into a log file.
• Anything your shell can do becomes a step: run the test suite, tag a release, call a CLI, or drive an installed Claude Code or Codex session headless as one lane inside a larger process.

This is what makes a net more than an orchestrator of API calls. The same graph that reasons about a change can build it, test it, and ship it, with every step recorded in the history above. See docs/commands.

Tools: an HTTP call, a script, a container, or a whole net

Everything a net reaches outside itself lives in one durable tool catalog. The same catalog can describe very different things because each entry separates the contract (what may be called) from the binding (how it runs):

Two properties make this more than a plugin list. Lookup is local-first: a model's own catalog shadows the shared global one, so one workspace can ship its own variant of a shared script without disturbing anybody else. And capability flags decide which kinds a persona can even see: h for HTTP, s for scripts, d for containers, t for tool nets, each granted separately, so "may register a script" never quietly means "may spawn containers".

The ladder is the part worth stealing. Start with a shell one-liner in a command lane. Promote it to a registered script once it earns a name. Wrap a container when it needs its own runtime. Publish a tool net when the whole pattern becomes reusable, and other nets call it by name. The Forge agent can build that last step from a plain-language intent, inside the running system. See docs/tool-catalog.

Deployment options, easiest first

The same nets, personas, and packages run in every deployment. Pick the one that fits today and move later without rebuilding anything.

Desktop Lite is the easy one, and for an unusual reason: you never configure an LLM provider at all. The bundled server LLM stays switched off, and the model inside your connected MCP client does the reasoning through external fires. The subscription you already pay for is the only provider involved. There is no API key to paste, no Ollama process to feed, and no cloud account to create. Install the package, connect the client from the tray, ask for what you want. Token binding, scheduling, emission, accounting, and the event trail stay with the local runtime, so the work is still governed and still replayable. If you want a persona to keep going while no client is connected, it can drive an installed Claude Code or Codex CLI instead.

1. Download the latest release: one package per platform.
2. Open it. The tray launcher starts and health-checks the whole stack on your machine.
3. Connect your client: tray → Connect Claude Code / Connect Codex / Copy MCP URL + Token, one paste, then ask for the persona or team you want.

Checksums, platform notes, and the full workflow: Desktop Lite.

The runtime underneath

Agentic-Nets turns a process into a live, inspectable Petri net. Typed JSON tokens hold the current state; transitions run deterministic code, LLMs, personas, commands, or HTTP calls; and every mutation becomes part of an event-sourced history.

The platform solves the part that a prompt or chat transcript does not: how work keeps running, who may change what, where handoffs go, what happened after the model answered, and how the whole process can be reused. A developer agent, review team, health coach, research pipeline, or operations process all use the same primitives: explicit state, controlled transitions, durable context, permissions, evidence, and approvals.

Humans can operate that state through Studio or an installable Net Application. Agents use the same state and actions through MCP, the CLI, or runtime tools. The interface changes; the governed net underneath does not.

At a glance

The mental model

Model (isolated workspace)
└── Session (one installed or created runtime)
    └── Net (the executable graph)
        ├── Places + tokens (typed state)
        └── Transitions (work and routing)

Net Application = session runtime + manifest + optional compiled UI

The product in one loop

Design -> Run -> Observe -> Review -> Propose change -> Approve -> Version
       -> Verify -> Compare -> Crystallize

This makes Agentic-Nets useful well beyond software delivery. The reusable substrate is the same for research, operations, support, education, finance, health coaching, logistics, or another domain: explicit state, controlled transitions, evidence, permissions, and a feedback loop.

Domain-general does not mean domain-omniscient. A useful autonomous process still needs domain context, success criteria, trustworthy data and integrations, bounded tools and authority, human or policy approvals, and validation matched to its risk. Agentic-Nets provides the governed runtime and historical evidence; it does not silently self-rewrite or remove responsibility from the operator.

The current platform in five points

1. External execution / bring your own model. Mark an llm or agent transition as external, or apply that policy to a net, session, or model. Master leaves it alone; an MCP client leases the exact prepared prompt, supplies the answer with its own model, and hands it back to the normal emit-and-consume pipeline.
2. Installable agent teams. Agent Hub currently ships seven starting templates: Safe Product Team, Model Steward, Dev Crew, Research Analyst, Health Coach, Context Curator, and Crystallizer. Model profiles can compose the right resident agents and context systems automatically.
3. NetHub packages whole systems. Publish and install ten artifact kinds: nets, sessions, applications, models, agent teams, context systems, tool nets, individual tools, catalogs, and blobs. Dependencies are bundled and credentials are scrubbed. Applications combine an ordinary session runtime with an optional verified UI surface; see the developer guide and the complete Persona Kanban tutorial.
4. Governance is enforced at runtime. Ten positional capability flags (rwxhludcts), named capability profiles, tool allowlists, resource scopes, Vault-backed credentials, a fleet-wide LLM freeze, and an automatic spend breaker bound what an agent can do.
5. One MCP server exposes the platform. @agenticnets/mcp provides a focused curated workflow surface plus an optional native catalog used by in-net agents. Desktop Lite defaults to curated; server installs retain both for compatibility.

Net Applications: deploy a UI without rebuilding Studio

Release status: The SDK, generic Studio host, and certified examples are currently on main under Unreleased. Use matching application-capable runtime and Studio builds; the v2.47.0 Desktop packages predate this finalized contract.

A Net Application is not an Angular module imported into the closed GUI. It is an independently built, versioned NetHub artifact with three parts:

• an ordinary session runtime containing nets, inscriptions, permitted initial tokens, and dependencies;
• a manifest declaring semantic store roles, schema-checked actions, permissions, agent instructions, and UI metadata;
• an optional self-contained browser module that registers one custom element.

flowchart LR
    source["Angular UI + session runtime + manifest"] --> package["Versioned application package"]
    package --> hub["NetHub"]
    hub -->|"install into a model"| instance["Application instance"]
    instance --> runtime["Session, nets, and event-sourced tokens"]
    instance --> descriptor["Installed UI descriptor"]
    descriptor --> studio["Studio's generic application host"]
    human["Human"] --> studio
    studio --> actions["Declared roles and actions"]
    agent["Persona or MCP agent"] --> actions
    actions --> runtime

Publishing adds the reusable definition to NetHub. Installing it verifies the package and UI hashes, imports the runtime into the selected model, and records the installed descriptor. Studio discovers that descriptor, loads the verified module, creates its custom element, and injects a constrained runtime bridge. There is no private route change, source-code import, npm build, or Studio image rebuild per application.

The UI is only a projection and controller. The event-sourced net is the source of truth, so a human in Studio, an MCP-connected agent, and a resident Persona agent can see and operate the same work through the same declared contract.

The complete open-source example is Persona Kanban. One installation gives humans a five-column board while connected and resident agents can discover, create, claim, comment on, review, and complete the same tasks. Its canonical card state and append-only activity facts live in ordinary net places—there is no Kanban-specific backend or database.

Start with:

cd agentic-net-apps
npm install
npm run build
npm run pack:kanban
npm run test:kanban-package

• Net Application SDK workspace
• Developer guide: architecture, contract, packaging, deployment, and security
• Persona Kanban: complete source-to-Desktop tutorial
• Application certification levels and real-stack gates

Executable UI currently uses a same-origin trusted-element model. Treat it as a trusted beta: install only reviewed packages from trusted publishers. Public marketplace execution still requires publisher signatures and stronger UI isolation; the certification specification tracks that boundary explicitly.

One complete worked example is the Safe Product Team: Product Manager, Architect, Developer, Reviewer, Release Guardian, and Chronicle over a deterministic backlog/review backbone. Repository policy is explicit, release effects are approval-gated, every stage writes structured status, and meaningful milestones appear in Protocol. Invoke the MCP prompt start-safe-product-team. It is an example, not a domain boundary: the domain-neutral Model Steward reviews any model's nets, running processes, event evidence, risks, bottlenecks, and optimization opportunities without modifying them.

One net, two execution locations: an external MCP agent hands structured findings to a master-run LLM lane.

If prompt-based agents feel powerful but structurally weak, this is the missing layer:

• Agents live in nets. Context is structured state, not a fragile chat session.
• Nets talk to nets. Teams, tools, approvals, memory, and pipelines become explicit handoffs.
• Everything stays inspectable. Tokens, tool calls, events, and emissions remain queryable and replayable.
• The same model scales up. Build one guarded developer agent or a whole product runtime with the same primitives.

Start here

Licensing note. Agentic-Nets is a hybrid stack. Public components in this repository are licensed under BSL 1.1 and convert to Apache 2.0 on 2030-02-22. The orchestration core ships as closed-source Docker Hub images and desktop release assets under the Proprietary EULA. See licensing before production use.

Quick local run

This starts the lightweight local stack from Docker Hub. Use the longer install section if you want monitoring, local public-service builds, Ollama cloud-model login details, or troubleshooting notes.

git clone https://github.com/alexejsailer/agentic-nets.git
cd agentic-nets/deployment

cp .env.template .env
# Optional for master-run llm/agent lanes: choose one server provider.
# LLM_PROVIDER=claude + ANTHROPIC_API_KEY=...
# LLM_PROVIDER=openai + OPENAI_API_KEY=...
# LLM_PROVIDER=ollama for the bundled Ollama container
#
# You can instead leave AI lanes stopped/external and let a connected MCP
# model execute them. See "Connect over MCP" below.

docker compose -f docker-compose.hub-only.no-monitoring.yml up -d

cat data/gateway/jwt/admin-secret
open http://localhost:4200

Desktop Lite (macOS + Windows + Linux) — no Docker or server LLM

This is the fastest local persona/team creator: install one package, connect Codex, Claude Code, or another MCP client, and choose a small starter pattern, invoke start-safe-product-team for the product-delivery example, or ask for one developer, domain expert, reviewer, or another specialist. Agentic-Nets gives those personas durable context, task/result places, deterministic tools, schedules, review hand-offs, readable Protocol reporting, and an event-sourced audit trail. The bundled server LLM is disabled by default; interactive reasoning comes from the connected client. A persona can also run unattended through an installed Claude Code/Codex CLI. No Docker daemon, Java or Node installation, API key, or Ollama process is required.

Desktop Lite is loopback-only and is not the recommended production deployment. Use the Docker/server deployment for remote access, clustering, monitoring, or production lifecycle controls.

Download the latest release. Every release page carries one package per platform:

Verify your download against the release's SHA256SUMS.txt and its Ed25519 signature SHA256SUMS.txt.sig (both attached to every release).

Installation notes:

• macOS: AgenticNetOS-<version>-macos-<arch>.dmg — open, accept the license, drag to Applications. Current builds are unsigned, so macOS refuses the first open: right-click the app → "Open", or allow it under System Settings → Privacy & Security → "Open Anyway".
• Debian/Ubuntu: sudo apt install ./AgenticNetOS-<version>-linux-<arch>.deb, then run /opt/agenticnetos/bin/AgenticNetOS. On a server without a desktop it runs headless (no tray, same services).
• Fedora/RHEL: the matching .rpm.
• Windows: run AgenticNetOS-<version>-windows-x64.msi. Current builds are unsigned, so SmartScreen may warn on first run ("More info" then "Run anyway"). Upgrades install over the old version in place; the msiexec-level upgrade path (previous release installed, data preserved, app serving afterwards) is tested in CI before every release.

Build the installer yourself: clone this repository and run agentic-net-desktop/scripts/build.sh (macOS/Linux) or agentic-net-desktop\scripts\build-windows.ps1. The primary requirements are a JDK 21+ and Node.js 22; see the Desktop Lite guide for the small platform packaging prerequisites. Closed node/master/GUI binaries come from matching checksum-verified release assets, with Docker Hub images only as a fallback.

Then: use "Connect Codex (copy config)" or "Connect Claude Code (copy command)" in the tray and ask the client to read agenticnets://docs/starter-patterns, then choose the smallest matching example. Use agenticnets://docs/safe-product-team for the worked product-delivery team or agenticnets://docs/personas for a custom specialist/domain. With no server provider, master skips provider-backed AI lanes rather than failing them, and the connected model can serve them. For unattended personas, use an agent transition with llmMode:"bash" and binary:"claude"|"codex"; for one-shot headless work use a command lane with the prompt on stdin. Master still owns token binding, emission, accounting, and the event trail. Data lives in ~/.agenticos/ and survives updates. The bundled command executor is eligible for every model and activates that model on demand after its first command lane, so newly created domains need no executor configuration. Full workflow and limitations: Desktop Lite.

Updating: quit the app, install the new package over the old one (macOS: drag-replace in Applications; Debian/Ubuntu: sudo apt install ./<new>.deb), relaunch. All data and settings live in ~/.agenticos/ and survive updates — the app itself is stateless. The tray notifies you when a new release is out.

Everything binds to localhost by default. The bundled node, master and gui binaries are covered by the Proprietary EULA; the rest of the bundle is built from this repository under BSL 1.1.

See it running in production

Not slideware — these are live systems, each one Agentic-Nets running a real harness end to end:

• forum.agentic-nets.com — a real product forum. A feature request posted here is picked up by a virtual team net, triaged, built, tested, deployed, and reported back on the thread — automatically, with every lifecycle milestone posted as it happens.
• gitanalytics.agentic-nets.com — the actual product that team is building: a live git-commit-analytics service whose new endpoints are shipped by agents, not people.
• The safe-teams net — the virtual agile team that connects the two: PM, Architect, Developer, QA, DevOps, and RTE agents coordinating through a single net — intake → design → code (a real command transition runs the coding CLI on an executor) → QA gate → deploy → status. One harness turns a forum post into a shipped, verified feature.
• agentic-nets.com — full documentation, the concept chapters, and the product tour.

Public read-only live demo

You can open the live Studio in monitor mode and watch the safe-teams net working without admin access:

1. Open agentic-nets.com/#/monitor.

2. Paste this public read-only demo token into the login form:

   07a9af1d663f899f79f08ca56050a977d41472e34cc0dd0f74abe046446f78f9
   

3. Keep Read-only access (no writes) enabled and log in.

The monitor view is scoped to the safe-teams model. It lets guests inspect the live net, token counts, event story, console, agenda, and current handoffs. Read-only sessions cannot edit nets, fire transitions, deploy changes, or use write-capable assistant personas.

You can still ask questions from the monitor. The chat is pinned to the Domain Expert (read-only) persona, which can explain what is happening in the visible system. The normal Universal Assistant, Workflow Builder, Persona, and other write-capable personas are reserved for authenticated Studio use.

What you can model with it

• Virtual developers with explicit permissions, memory, and execution boundaries.
• Virtual agile teams where planner, builder, reviewer, tester, and releaser agents coordinate through nets.
• Smart development tools that behave like reusable nets instead of throwaway prompts.
• Development pipelines that generate code, run checks, gate releases, and keep a durable audit trail.
• Product-level systems where backlog, QA, docs, incidents, and operations communicate as structured nets.
• Industry-specific operating models in software, finance, support, operations, research, healthcare, logistics, or any other domain that can be expressed as communicating nets.

Who this is for

• Builders who want agents to operate inside explicit state machines instead of loose prompt loops.
• Teams that need remote execution, approvals, secrets, and audit trails around autonomous work.
• Product engineers turning one-off agent workflows into reusable internal systems.
• Researchers and tool builders exploring Petri nets as a runtime model for agent coordination.

Probably not for you if

• You only need a one-off chat wrapper or a single scripted LLM call.
• You require every runtime component to be permissively licensed today.
• You need certified production software for regulated environments without doing your own validation.

Net of nets

One net can contain one or many agents. Many nets can also work together as a larger runtime: one can guard, one can gather, one can synthesize, one can execute, and all of them can exchange structured state through explicit flows instead of hidden prompt handoffs.

Example net

This simple crawler net shows the model in practice: places hold the state, http fetches, an agent transition categorizes content, a command transition runs remote work through an executor, and map plus pass transitions route results through the graph.

How behavior is modeled

Agentic-Nets uses seven transition types: pass, map, http, llm, agent, command, and link.

• agent transitions are the core runtime primitive. They can mimic almost any agent behavior or mode, but inside a governed net with explicit inputs, outputs, permissions, and memory boundaries.
• agent transitions can also adapt the net itself. If an agent has sufficient rights, it can read tokens in the net, create additional places and transitions, and extend the structure on demand instead of staying confined to a fixed graph.
• command transitions connect the net to remote execution. They define which executor can run a command remotely and bring the result back into the net as structured state.
• Deterministic and non-deterministic transitions coexist. Fixed logic can stay fixed, while open-ended reasoning stays open-ended, in the same runtime and on the same graph.
• This is what makes the model powerful across domains. A net can combine several cooperating agents with deterministic control flow, verification, remote execution, and cross-net communication.

Everything a harness needs — assembled by prompting

A production agent harness is all the scaffolding around the model: the tools it can call, the control flow between steps, the memory it keeps, where it runs, who is allowed to do what, and how you see what happened afterwards. Most teams hand-write that harness in code and re-write it for every new agent. Agentic-Nets gives you every one of those pieces as a first-class primitive you build by describing it — and each piece is inspectable, reusable, and governed by default.

Drive it your way — visual, conversational, or API-first

You are not tied to one model vendor or one interface.

In the Studio (the GUI). Watch every net, token, tool call, and event as it happens, and set the whole thing up by clicking — create models, sessions, nets, places, transitions, and inscriptions, deploy them, and adapt anything live. The Studio ships **several built-in ass