Skip to main content

Open source, MIT licensed

Run your AI agent on its own canister

AgentVault packages your AI agent, deploys it to a durable Internet Computer canister, and keeps it running with its own identity, wallet, secrets, and memory. No server to babysit, no session to lose.

Not tied to one chatbot or one model. Bring Claude Code, OpenClaw, Goose, OpenCode or an agent of your own.

my-agent

npx agentvault@latest init my-agent --template default

cd my-agent

agentvault package ./

agentvault deploy --network ic

Canister Info

Agent deployed

Canister ID
vjk3q-wyaaa-aaaaa-aabqa-cai
Network
ic
Agent
my-agent
Deployed
2026-10-31T09:41:07.312Z
WASM Hash
3f9a1c0be27d4a58...
Example session. The canister ID and hash are placeholders.

Works with your stack

Bring any model, any agent

AgentVault is where your agent lives, not the model it talks to. Each layer gives you more than one choice.

  • Models

    The dashboard's inference API tries a Bittensor subnet, then Venice AI, then a model you run locally with Ollama. Each answer names the one that replied.

    • Bittensor
    • Venice AI
    • Ollama (local)
  • Agents

    Pick your agent's type in the dashboard, deploy it to its own canister and record an archive receipt on Arweave.

    • Claude Code
    • OpenClaw
    • Goose
    • OpenCode
    • Your own
  • MCP tools

    The built-in MCP server runs over stdio, so any MCP client that can launch a local server gets the same tools.

    • HyperVault
    • Wiki
    • Total Recall
  • Chains

    Give each agent wallets of its own, all managed from one CLI.

    • ICP
    • ckETH
    • Solana
    • Polkadot
    • Arweave

Get started

From zero to a running agent in three steps

Every agent gets its own canister, a persistent identity, wallet, and memory that keep running whether or not you're watching. No infrastructure to provision, no server to keep alive.

Wire status, health, and backup into your deployment pipeline. Fast recovery should be built in from day one, not bolted on later.

# Create a new project
agentvault init my-agent
cd my-agent

# Package and deploy locally
agentvault package ./
agentvault deploy --network local

# Check that it's alive
agentvault status
agentvault health

Connect a wallet and copy your deploy commands

Connect your wallet, tune your deployment settings, and get ready-to-run commands for the exact environment you're shipping to.

Step 1

Connect a wallet

No wallet connected yet

Connecting Plug Wallet sets the deploy network to ic. MetaMask and ArConnect set it to local. You can change it in step 2.

Step 2

Configure your deploy

Install

npx agentvault@latest init my-agent --template default

Deploy

cd my-agent && npx agentvault@latest package ./ && npx agentvault@latest deploy --network local

Back up and restore in two commands

Name your agent, choose where it comes back, and copy the exact commands. The backup is encrypted on your machine, and a restore installs it on a fresh canister.

Step 1

Describe your backup

Run the commands from the agent's project folder after agentvault package. After agentvault vault init (HashiCorp Vault or Bitwarden), store the passphrase once with agentvault vault put my-agent backup-passphrase; both commands read it from there, so it never lands in your shell history.

Step 2

Copy your commands

Export command

agentvault backup export my-agent --encrypted --wasm ./dist/my-agent.wasm --output ./my-agent.zip.enc

Restore command

agentvault backup restore --zip ./my-agent.zip.enc --agent my-agent --network local

AES-256-GCM · Passphrase from Vault · Optional Arweave manifest · Fresh canister on restore

Backup lifecycle

Three steps. Each one lists what it leaves behind.

  1. 01

    Export

    Stage the agent's state, wallet files, logs, dependencies and WASM, then encrypt the zip with AES-256-GCM.

    Output
    <agent>.zip.enc + <agent>.manifest.json
  2. 02

    Anchor (optional)

    Upload the manifest and the zip's SHA-256 hash to Arweave. The encrypted zip stays on your machine.

    Output
    Arweave manifest transaction
  3. 03

    Restore

    Decrypt the zip, refuse unsafe paths and symlinks, restore the state, and install the WASM on a fresh canister.

    Output
    new canister ID

How it works

Package, deploy, recover

  1. 01

    Package your agent

    Compile local agent code to WASM and generate the artifacts you deploy.

    View the workflow
  2. 02

    Deploy to a canister

    Ship to a local replica or ICP mainnet. Upgrade an existing canister, and list an agent’s deployment history.

    Open deployment docs
  3. 03

    Keep it recoverable

    Run health checks, roll a deployment back, and export an encrypted backup you can restore to a fresh canister.

    Open the backup guide

What the platform includes

  • Multi-chain wallets

    Keep ICP, ckETH, Solana, Polkadot and Arweave wallets in one CLI, with encrypted local key custody. ckETH, Solana and Polkadot transfers can be signed; ICP and Arweave signing is not implemented yet.

    Explore wallets
  • Security controls

    Approval requests with a signer quorum and an optional TOTP check, both tracked locally (they do not block deploys and are not cryptographic multisig), plus canister write guards: authorized principals, freeze and kill switches.

    Review security
  • Operations tooling

    Monitoring, promotion, rollback, and task visibility for managing long-lived agents.

    Open monitoring
  • Architecture

    Understand canister internals, module boundaries, and ICP integration primitives before you scale out.

    Read the architecture

The agent stack

AgentVault is the runtime; the smallchat suite covers the rest of the stack

Long-running agents need more than a place to execute. They need a way to pick the next tool, a memory that survives, and a way to keep that memory inside a token budget. The smallchat suite covers those layers and just shipped 1.0. AgentVault ships on its own schedule: today it has its own dispatcher modeled on smallchat and an MCP client for Polytician 3.0, and Stenographer and Short-hand are not integrated yet.

  • AgentVault

    Runtime

    Packages your agent to WASM and runs it on an Internet Computer canister, with a durable identity, multi-chain wallet, secrets vault, and versioned memory that survive independent of any browser tab or host process.

  • smallchat

    Tool dispatch

    www.smallchat.dev
    Release
    @smallchat/core 1.0.0
    Release
    SmallChat Swift package 1.0.0

    Semantic tool dispatch. It resolves an intent to at most one tool, checks the arguments against that tool's JSON Schema, and records a replayable proof. The same artifact, embedder and runtime state give the same choice.

    In AgentVault: Not a dependency. AgentVault's orchestration library has its own opt-in dispatcher modeled on smallchat's selector design: exact selector lookup with class fallback and no embeddings, followed by policy checks, rate limits and MFA gating.

  • Polytician

    Semantic memory

    polytician.smallchat.dev
    Release
    polytician 3.0.0

    A local-first MCP server for semantic memory. Each concept can be stored as, and converted between, a 384-dimension vector, markdown and structured ThoughtForm JSON.

    In AgentVault: AgentVault's orchestrate --polytician-entry option calls a Polytician 3.0 MCP server over stdio, adds the concepts most relevant to the task to Claude's prompt, and saves each session's result as a concept in the agent's own namespace. The polytician push-all, pull and archive commands use Polytician's vault_* tools, which AgentVault turns on by passing Polytician its webapp's URL and token. Pushes reach the memory_repo canister through webapp routes that sign with an identity the repo owner has authorized.

  • Stenographer

    Conversation memory

    stenographer.smallchat.dev
    Release
    @stenographer/core 1.0.0

    An MCP server that indexes agent conversation logs for GraphRAG search and keeps a hash-chained truth ledger. One agent alone settles nothing. A person signs, or two or more agent sessions agree within 15 minutes, each citing its own checkable evidence. Agents never override or rule, and a contested claim goes to a person.

    In AgentVault: Not integrated yet. The engineering guide sketches how it could plug in.

  • Short-hand

    Context compaction

    short-hand.smallchat.dev
    Release
    @shorthand/core 1.0.0

    Progressive, LSM-tree-style compaction of conversation history into a token-budgeted context frame. Corrections are tracked as tombstones, so overridden facts do not quietly come back.

    In AgentVault: Not integrated yet. The engineering guide proposes it in place of the Polytician enricher's character-count truncation.

Next step

Backups and recovery

Export an encrypted backup, restore it to a fresh canister, and manage backup files from the CLI.

Open the backup guide

Deep dive

Security and hardening

Review production hardening, key custody controls, and incident response strategy.

Read the security guide