agent infrastructure / online

The platformto build agents.

Design the behavior. Axon runs the loop, secures the tools, remembers the work, and ships the same agent anywhere.

»axonruntime

agentsportable behavior modulestyped capability promptsreusable context cognetsswappable cognition benchesreproducible evidence

portable by design policy enforced cloud or self-hosted

portable behavior

Begin with an agent. Make it yours.

Install complete agents into a folder you can inspect, run, change, and deploy.

official

@axon/release-engineerv1.4.0

Owns the path from a green build through deployment and release notes.

2.4k 186

verified

@cody/repo-operatorv0.8.4

Repository-native engineering agent with durable project context.

1.1k 94

@acme/support-triagev2.1.2

Classifies inbound support and routes escalations with full traceability.

864 71

composition, not lock-in

Assemble the stack. Keep the seams.

Each artifact has one job. Swap a model without rewriting the agent. Add a module without handing it the runtime. Change cognition through a versioned ABI. Measure the result independently.

selected artifacts

agent@you/research-operator

module@axon/arxiv + @axon/github

cognet@axon/zero

modelclaude-sonnet / qwen-local

bench@axon/repo-navigation

assembled agentresearch-operator

ready

capabilities18 tools

inferencelocal / cloud

policyrestricted

verification92.4 score

$ axon install @you/research-operator

from project to public primitive

If it works once, make it available everywhere.

Package agents and capabilities under your scope. Published versions are immutable, inspectable, and installable in one command—by you, your team, or the wider ecosystem.

$ axon prepare

→

$ axon publish

publish preview

@acme/release-engineer

v1.4.0

axon.config.tsincluded

src/included

prompts/included

.envexcluded

node_modules/reconstructed

policy visible before install immutable version

quality you can inspect

Claims become reproducible results.

#subjectnavigationtool userecoveryoverall

1 @cody/repo-operator92.4

2 @axon/release-engineer89.7

3 @acme/support-triage83.1

benchmark / @axon/repo-navigation24 cases · evidence attached · methodology public

open edges

The platform earns the workload.

Bring your own inference. Run local models. Build a standard container. Keep agent source and durable state in a shape that does not depend on one hosting provider.

Your inference

Axon Cloud, direct provider keys, or models on your hardware.

Your infrastructure

Managed deployment or a standard container for any platform.

Your source

The agent remains a readable folder, not a hosted configuration object.

Your authority

Policy stays with the agent and applies whichever route it takes.

build · compose · verify · run

The ecosystem for agents that do real work.

Start with a working runtime. Add what you need from the registry. Publish what should become reusable.

start building browse modules

agent infrastructure / online

The platform to build agents.

Design the behavior. Axon runs the loop, secures the tools, remembers the work, and ships the same agent anywhere.

start building

$bun install -g @arcforge/axon

portable by design policy enforced cloud or self-hosted

»control plane

all systems operational

agent / production

release-engineer

running

uptime99.98%

sessions1,284

policyenforced

cognetzero@0.1.0attached

inferenceanthropic/claude-sonnetaxon cloud

capabilitiesgithub · linear · filesystem12 tools

release-engineer.europe-west1.axon.runv0.8.4

artifactone folder

runtimemanaged

inferenceany model

infrastructureyour choice

the complete lifecycle

One system from first prompt to production.

Axon gives every agent a standard shape, then carries that shape through development, execution, distribution, and operations. Improvements to the runtime reach every agent without a rewrite.

Build

Tools, prompts, scripts, routes, and policy in a project you already know how to version.

axon init

Run

A managed cognitive loop with durable sessions, visible execution, and a real security boundary.

axon dev

Verify

Boot the real runtime in tests. Compare agents and models through reproducible benchmarks.

bun test

Ship

Publish for reuse, deploy to managed infrastructure, or build a portable container.

axon deploy

model independent

Inference is a route, not your architecture.

Keep the agent, tools, policy, and session model fixed. Change where intelligence comes from with one line—or switch live in the TUI.

Axon Cloud

Curated models, metered billing, and no provider keys to manage. Available the moment you log in.

engine: Axon({ model: "auto" })

live inference routes model palette *

your agentbehavior stays fixed

same runtime loopsame capability surfacesame session trace

secure by architecture

The model can request. Only the kernel can take.

Agent code runs in an isolated capsule. Filesystem, network, and process access cross a real policy boundary before they execute. The model is untrusted by design.

your resources

filesystem

processes

network

policy gateallow · deny · escalate

trusted kernel

mediate

record

persist

versioned ABIoutput() · stream() · run()

untrusted capsule

⠿

cognet + toolsisolated subprocess

ship without lock-in

Managed when you want it. Portable when you don’t.

Deploy to Axon Cloud for a URL, identity, durable data, scaling, schedules, and logs. Or build the same folder into a standard container and run it on your infrastructure.

$axon deploy

$axon build

deployment / release-engineerrunning

release-engineer.axon.run

regioneurope-west1

runtimeaxon 0.3.6

storagedurable

scaling0 → 3

14:32:08deploy:revision:ready184ms14:32:09agent:boot:complete612ms14:32:11session:attachedremote

build the agent, not the plumbing

Your behavior. One maintained platform.

Start locally in the terminal. Keep every model and infrastructure option open. Deploy when the agent is ready.

read the docs explore agents

see the complete lifecycle in docs

registry online

Everything agents need, in one ecosystem.

Build from open components. Combine agents, capabilities, cognition, models, and benchmarks on one portable runtime.

explore the registry start from zero

»axonruntime

agentsportable behavior modulestyped capability promptsreusable context cognetsswappable cognition benchesreproducible evidence

search agents, modules, cognets, models, benches.../

portable behavior

Begin with an agent. Make it yours.

Install complete agents into a folder you can inspect, run, change, and deploy.

@axon/dungeon-masterv0.1.3

A solo, rules-light tabletop RPG Dungeon Master — plays a live adventure with you in the terminal, with real dice and a fresh world every game.

@axon/fragcheckv0.1.3

A Linux gaming-readiness specialist — inspects your rig and tells you, straight, how ready it is to game, with a scored report and ranked fixes.

@axon/voxunversioned

A published Axon agent.

composition, not lock-in

Assemble the stack. Keep the seams.

Each artifact has one job. Swap a model without rewriting the agent. Add a module without handing it the runtime. Change cognition through a versioned ABI. Measure the result independently.

selected artifacts

agent@you/research-operator

module@axon/arxiv + @axon/github

cognet@axon/zero

modelclaude-sonnet / qwen-local

bench@axon/repo-navigation

assembled agentresearch-operator

ready

capabilities18 tools

inferencelocal / cloud

policyrestricted

verification92.4 score

$ axon install @you/research-operator

from project to public primitive

If it works once, make it available everywhere.

Package agents and capabilities under your scope. Published versions are immutable, inspectable, and installable in one command—by you, your team, or the wider ecosystem.

$ axon prepare

→

$ axon publish

publish preview

@acme/release-engineer

v1.4.0

axon.config.tsincluded

src/included

prompts/included

.envexcluded

node_modules/reconstructed

policy visible before install immutable version

quality you can inspect

Claims become reproducible results.

#subjectnavigationtool userecoveryoverall

1 @cody/repo-operator92.4

2 @axon/release-engineer89.7

3 @acme/support-triage83.1

benchmark / @axon/repo-navigation24 cases · evidence attached · methodology public

open edges

The platform earns the workload.

Bring your own inference. Run local models. Build a standard container. Keep agent source and durable state in a shape that does not depend on one hosting provider.

Your inference

Axon Cloud, direct provider keys, or models on your hardware.

Your infrastructure

Managed deployment or a standard container for any platform.

Your source

The agent remains a readable folder, not a hosted configuration object.

Your authority

Policy stays with the agent and applies whichever route it takes.

build · compose · verify · run

The ecosystem for agents that do real work.

Start with a working runtime. Add what you need from the registry. Publish what should become reusable.

start building browse modules

explore the registry

a software-native agent framework

Build agents like software.

An Axon agent is a folder: behavior, capabilities, policy, and state in source you can understand. Run it locally, test it for real, then ship the same artifact.

build your first agent

$axon init my-agent

my-agentready

axon.config.tsidentity / engine / policy

export default defineAgent({  engine: Axon({ model: "auto" }),  modules: ["@axon/github"],  policy: {    fs: { read: ["./src/**"] },    process: { allow: ["git *"] },  },})

declares authorityone source of truth

one folderone runtimeany destination

the agent is the folder

Everything important is visible.

No graph editor, hidden orchestration layer, or proprietary project format. The source describes what the agent is, what it can do, and where its authority ends.

my-agent/8 resources

boot.vueidentity

src/tools/capabilities

src/scripts/workflows

axon.config.tsauthority

data/memory

boot.vue

Give it a stable identity.

The base prompt is rendered from source on boot. It does not drift as conversations grow, and edits hot-reload into future work.

src/tools/

Export a function. It becomes a tool.

TypeScript signatures and JSDoc become the model’s typed capability surface. Your scripts call the same functions directly.

src/scripts/

Turn expertise into repeatable workflows.

Scripts gather context, call tools, invoke the loop, and run identically from the terminal, TUI, HTTP, or another script.

axon.config.ts

Models and permissions stay explicit.

Choose inference, install modules, and declare filesystem, network, and process policy at the project boundary.

data/

State lives beside the agent.

Session traces, deliberate knowledge, and working data persist through named threads, restarts, and cloud redeploys.

the runtime underneath

You write the behavior. Axon owns the loop.

Every invocation follows one maintained path. The model cannot skip tool results, silently escape policy, or leave execution outside the session record.

Assemble

Render identity, history, prompts, tools, and live state.

→

Invoke

Send the same context through any configured engine.

→

Act

Execute TypeScript inside the policy-controlled capsule.

→

Witness

Commit results and continue until the agent has seen them.

application boundaryconst { stream } = axon.stream ({ prompt, thread: "project-alpha" }) runtime managed

work with the live agent

The terminal is the development environment.

Talk to local, deployed, and repository agents through the same interface. Inspect tool calls, move between sessions, run scripts, switch providers, and approve escalations without leaving the work.

Axon TUI connected to an agent — one interface · local, repository, and cloud agents

Axon Fleet event log open in VS Code — every step, as it happens · beside the code it touched

stay in your editor

The whole fleet, beside your code.

Fleet brings every agent into VS Code. Watch the event log as work happens, open a session next to the file it touched, and approve escalations without breaking the loop you are already in.

treat behavior as code

Test the thing that actually runs.

The test harness boots the real kernel, capsule, tools, and policy. A deterministic mock replaces inference—not the runtime— so failures remain fast, free, and reproducible.

tests/behavior.test.tsbun:test

const runtime = await Axon({
  blueprint: { config: {
    engine: Mock({
      "review the file": [
        run(`fs.read("src/index.ts")`),
        "The file looks correct."
      ]
    })
  } }
})

const result = await runtime.axon.request(
  "review the file"
)

tests

Real runtime. Deterministic model.

12 passed

benchmarks

Vary models, agents, and configuration.

4 × 6 matrix

session trace

Every context, action, and result.

jsonl

my-agent/

the same source

axon deployaxon build

managedAxon Cloud

URL · identity · storage · scale

portableContainer image

your account · your infrastructure

source you own · infrastructure you choose

One folder. A complete agent.

Begin with a working agent, add only what makes it yours, and let the platform carry it from your terminal to production.

get started read the architecture

see more about the TUI

agent infrastructure / online

The platform to build agents.

Design the behavior. Axon runs the loop, secures the tools, remembers the work, and ships the same agent anywhere.

start building

$bun install -g @arcforge/axon

portable by design policy enforced cloud or self-hosted

»control plane

all systems operational

agent / production

release-engineer

running

uptime99.98%

sessions1,284

policyenforced

cognetzero@0.1.0attached

inferenceanthropic/claude-sonnetaxon cloud

capabilitiesgithub · linear · filesystem12 tools

release-engineer.europe-west1.axon.runv0.8.4

artifactone folder

runtimemanaged

inferenceany model

infrastructureyour choice

the complete lifecycle

One system from first prompt to production.

Axon gives every agent a standard shape, then carries that shape through development, execution, distribution, and operations. Improvements to the runtime reach every agent without a rewrite.

Build

Tools, prompts, scripts, routes, and policy in a project you already know how to version.

axon init

Run

A managed cognitive loop with durable sessions, visible execution, and a real security boundary.

axon dev

Verify

Boot the real runtime in tests. Compare agents and models through reproducible benchmarks.

bun test

Ship

Publish for reuse, deploy to managed infrastructure, or build a portable container.

axon deploy

model independent

Inference is a route, not your architecture.

Keep the agent, tools, policy, and session model fixed. Change where intelligence comes from with one line—or switch live in the TUI.

Axon Cloud

Curated models, metered billing, and no provider keys to manage. Available the moment you log in.

engine: Axon({ model: "auto" })

live inference routes model palette *

your agentbehavior stays fixed

same runtime loopsame capability surfacesame session trace

secure by architecture

The model can request. Only the kernel can take.

Agent code runs in an isolated capsule. Filesystem, network, and process access cross a real policy boundary before they execute. The model is untrusted by design.

your resources

filesystem

processes

network

policy gateallow · deny · escalate

trusted kernel

mediate

record

persist

versioned ABIoutput() · stream() · run()

untrusted capsule

⠿

cognet + toolsisolated subprocess

ship without lock-in

Managed when you want it. Portable when you don’t.

Deploy to Axon Cloud for a URL, identity, durable data, scaling, schedules, and logs. Or build the same folder into a standard container and run it on your infrastructure.

$axon deploy

$axon build

deployment / release-engineerrunning

release-engineer.axon.run

regioneurope-west1

runtimeaxon 0.3.6

storagedurable

scaling0 → 3

14:32:08deploy:revision:ready184ms14:32:09agent:boot:complete612ms14:32:11session:attachedremote

build the agent, not the plumbing

Your behavior. One maintained platform.

Start locally in the terminal. Keep every model and infrastructure option open. Deploy when the agent is ready.

read the docs explore agents

see deployment options in docs

build the agent, not the plumbing

Your behavior. One maintained platform.

Start locally in the terminal. Keep every model and infrastructure option open. Deploy when the agent is ready.

start building explore the registry