PAREL
PAREL(1)Parel ManualPAREL(1)

DEPLOY AN AI AGENT
IN ONE CONFIG FILE.

NAME
parel — run AI agents without running servers.
DESCRIPTION
Describe an agent in one agent.yaml and deploy it in minutes. Conversations keep their state between messages — nothing to host, no queues to wire. Works with Anthropic, OpenAI, and compatible model APIs; sandboxes, memory, and policies snap in as plugins.
[▓]agent.yaml
YAML · 18 lines
$ parel deploy ready
§01Platformfeatures.4

Everything an agent needs.
In one file.

01 · PROVIDER LAYER
provider:anthropic | openai | openai-responsescompatible:openai-compatible | anthropic-compatiblesandbox:e2bruntime:memory | budget | security
# current built-ins, extended through plugins

OpenAI, Anthropic, and compatible endpoints.

Select OpenAI, OpenAI Responses, Anthropic, or compatible API endpoints through agent.yaml; run on E2B sandboxes without changing agent code.

02 · COMPOSABLE RUNTIME
model
tools
memory
budget
policy
sandbox
steering
channel
[ kernel ]

Sandboxes, memory, and policies are plugins.

Models are selected through the model provider layer; sandboxes, tools, memory, budget caps, and security policies are runtime plugins.

03 · SERVERLESS STATE
active sessions · illustrativeautoscale ready

Agent state, not agent servers.

Sessions can sleep, resume, and wake from API calls, channels, schedules, or callbacks. PAREL keeps identity, state, policy, and execution history durable.

04 · AGENT-FRIENDLY DX
agent.yaml · main ← feat/budget+4 −1
model: claude-sonnet-5
− plugins: []
+ plugins:
+ - memory-rolling-summary
+ - budget-cap:
+ max_usd: 25
runtime.maxSteps: 40

A spec agents can edit.

agent.yaml is textual, diffable, schema-validated, and scriptable. Humans and coding agents can review, patch, validate, and deploy runtime intent safely.

§02Architecturecontrol plane / data plane

Push a config.
Get a running agent.

LIVE TRACE
ARCH(7)control plane → data planeARCH(7)
agent.yamlyour repo
agent: research-agent
provider: anthropic
plugin: sandbox-e2b
▓ parel kernelpolicy hooks
scheduler
router
budget
policy
hooks
trace
agent runtimeserverless
w-01running
w-02running
w-03idle
1DECLARE

Define your agent in agent.yaml — model, tools, memory, channels, budget. Store it in git.

2DEPLOY

parel deploy ships the spec to the control plane. The runtime validates policy and prepares durable session state.

3RUN

Events wake sessions through the kernel. Logs, traces, costs, and execution history stream back in real time.

§03Examplesexamples.3

Deploy from one file.
No agent server plumbing.

Define agent.yaml, authenticate once, and deploy. PAREL handles model provider dispatch, sandbox lifecycle, durable session state, channels, and policy hooks.

DURABLE SESSIONS
Message an agent now, come back next week — the conversation and its state are still there. No process to keep alive.
VERSIONED DEPLOYS
Every deploy creates an immutable version. Ship, compare, and roll back from the console.
PLUGIN SDK
Expose sandbox, memory, tools, channels, and guardrails through one plugin contract.
# Example agent definition.
version: "1"
agent:
name: research-agent
model:
provider: anthropic
model: claude-sonnet-5
plugins:
- system-static:
prompt: "You are a research assistant."
- sandbox-e2b
- memory-rolling-summary
- security-basic
- budget-cap:
max_usd: 25
runtime:
maxSteps: 40
§04Ecosystemproviders.5 / plugins.7

Model providers.
Runtime plugins.

Author a plugin →
MODEL PROVIDERS5
  • anthropic
    Claude family
    provider
  • openai
    Chat Completions
    provider
  • openai-responses
    Responses API
    provider
  • openai-compatible
    OpenAI-compatible endpoints
    provider
  • anthropic-compatible
    Anthropic-compatible endpoints
    provider
TOOLS2
  • sandbox-e2b
    isolated cloud VMs
    plugin
  • subagent
    child-agent delegation
    plugin
CONTEXT3
  • memory-rolling-summary
    LLM-compressed context
    plugin
  • system-static
    static system prompt
    plugin
  • steering-immediate
    mid-turn steering
    plugin
GUARDS2
  • budget-cap
    cost/turn caps
    plugin
  • security-basic
    command and secret filter
    plugin
§05Get Startedinstall.sh

Ship your first
serverless agent.

Create a workspace, connect the CLI, and deploy from a versioned spec. Keep models, sandboxes, tools, channels, memory, and policies behind the same runtime contract.

quickstart3 steps
$ npm install -g @parel/cli
$ parel login
$ parel deploy agent.yaml
# API key from Console / Settings# Node.js >= 22