▓ PAREL
Console →
PAREL(1)Parel ManualPAREL(1)

DEPLOY AN AI AGENT
IN ONE CONFIG FILE.

NAME
parel — agent runtime control plane.
DESCRIPTION
Any model. Any sandbox. Any tool. Composed as plugins, scheduled by a tiny kernel, run serverless on Cloudflare. The kernel only dispatches. Everything else is a plugin.
$ parel deploy →man parel
[▓]agent.yaml
●YAML · 18 lines
$ parel deploy▓ ready in 1.2s
§01Platformfeatures.4

A tiny kernel.
An infinite surface.

▓ 01 · PROVIDER NEUTRAL
model:anthropic | openai | google | mistral
sandbox:e2b | local | modal | fly
memory:rolling-summary | redis | postgres
# → swap any cell, no app code changes

Any model. Any sandbox.

Swap Claude for GPT, E2B for local — without touching agent code. Plugins implement a tiny interface; the kernel handles the rest.

▓ 02 · PLUGIN ALL THE WAY
[ kernel ]
model
tools
memory
budget
policy
sandbox
steering
channel

The kernel only dispatches.

Models, sandboxes, tools, memory, budget caps, security policies — every capability is a plugin. Nothing is bundled, everything is composable.

▓ 03 · SERVERLESS BY DEFAULT
concurrent agents · last 60s↑312 peak

Cloudflare edge, globally.

Runs on Durable Objects with automatic checkpointing. Crash recovery by platform. Scale from one user to ten thousand on the same agent.yaml.

▓ 04 · ONE CONFIG
agent.yaml · main ← feat/budget+3 −1
model: claude-sonnet-4-20250514
− budget: $5/day
+ budget: $10/day
+ schedule: always-on
+ memory: rolling-summary
sandbox: e2b

Git-native, diffable, gradual.

Version control, code review, gradual rollouts. agent.yaml is the contract between your team and production.

§02Architecturecontrol plane / data plane

Push a config.
Get a running agent.

LIVE TRACE
ARCH(7)control plane → data planeARCH(7)
agent.yamlyour repo
name: research-agent
model: claude
tools: [bash, file]
▾
▓ parel kernelpolicy hooks
scheduler
router
budget
policy
hooks
trace
▾
agent runtimeCF Workers
w-01running
w-02running
w-03idle
1DECLARE

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

2DEPLOY

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

3RUN

Messages dispatch through the kernel. Logs, traces, and costs stream back in real time.

§03Examplesexamples.3

Deploy from one file.
No runtime plumbing.

Define agent.yaml, authenticate once, and deploy. PAREL handles model adapters, sandbox lifecycle, session state, and policy hooks.

▶
DURABLE EXECUTION
Workflow checkpoints let turns resume after platform restarts.
▶
LIFECYCLE HOOKS
Intercept any stage: before/after model calls, tool execution, context building.
▶
PLUGIN SDK
Expose model, sandbox, memory, and guardrail capabilities through one plugin contract.
# Example agent definition.
name: research-agent
runtime: parel/v1
model:
provider: anthropic
name: claude-sonnet-4-20250514
temperature: 0.2
sandbox: e2b
memory: rolling-summary
tools:
- bash
- file_read
- file_write
schedule:
cron: "*/15 * * * *"
guards:
budget: $25/day
maxSteps: 40
§04Ecosystemplugins.7

First-party plugins.
Yours next.

Author a plugin →
▓ MODEL2
  • anthropic
    Claude family
    built-in
  • openai
    GPT family
    built-in
▓ SANDBOX2
  • e2b
    isolated cloud VMs
    built-in
  • local
    dev only, no isolation
    built-in
▓ MEMORY1
  • rolling-summary
    LLM-compressed context
    built-in
▓ GUARDS2
  • budget-cap
    daily/request caps
    built-in
  • security-basic
    command and secret filter
    built-in
§05Get Startedinstall.sh

Ship your first
controlled agent.

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

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