Does omp support camelStream?

Yes. omp accepts custom OpenAI-compatible providers in models.yml alongside its built-in ones, and camelStream is a standard chat-completions endpoint. The two blocks on this page are the whole integration.

Which omp roles should point at Stream?

Start with default, which handles your normal turns. If you want the whole toolbox on one endpoint, point plan, slow, smol, and task at camel-stream/auto too. There's no longer a cost reason to split roles across providers, so split only when you want different behavior.

What happens when the task role fans out subagents?

Subagents on the same key share the stream's queue, one generation at a time, so a wide fan-out completes in sequence rather than in parallel. If you fan out often, give subagent roles a stream of their own.

Do hashline edits still matter on a flat rate?

For the bill, no. For the agent, yes. Leaner edits leave more of the 260K-token window for code and tool output, and shorter generations finish sooner. The efficiency stops saving money and keeps saving time.

Can Stream sit in a fallback chain?

Yes, on either end. Running Stream as the primary with a metered provider behind it is the usual arrangement. The reverse works too while you evaluate.

What context window should the model entry declare?

Set contextWindow to 262144. Stream guarantees a 260K-token window on every request, and you get more when the model serving you supports it.

CAMELSTREAM · AGENT GUIDE

Point every omp role at unlimited frontier intelligence.

omp grew out of the minimalist pi harness by wiring in everything: an LSP client, a real debugger, subagents, and ten model roles to route them. camelStream gives those roles one flat answer. Frontier models, $5 a month per stream, however long the session runs.

Get your Stream API key Explore camelStream

Setup

Two YAML files.

Get a Stream key

Create the key in the Stream console. omp reads it straight from models.yml rather than the environment, so paste it there and keep that file out of any public dotfiles.

Register the provider

Add the camel-stream block to ~/.omp/agent/models.yml with auto as the model ID and a 262144 context window.

Route your roles

Set default to camel-stream/auto in config.yml. In a running session, /model cycles through configured models, which makes it easy to compare before you commit.

omp documentation

~/.omp/agent/models.yml

# ~/.omp/agent/models.yml
providers:
  camel-stream:
    baseUrl: https://stream.camelai.com/v1
    api: openai-completions
    apiKey: qaml_live_...
    models:
      - id: auto
        name: Auto
        contextWindow: 262144

# ~/.omp/agent/config.yml
modelRoles:
  default: camel-stream/auto

Base URL

stream.camelai.com/v1

Model ID

camel-stream/auto

Protocol

openai-completions

Auth

apiKey in models.yml

Why omp on Stream

Maximalist harness, minimalist bill.

A lot of omp's design answers token prices. Hashline edits exist to make every change cheaper, and roles like smol and tiny exist so subagent chatter never runs on your expensive model. Stream removes that half of the math. A role pointed at camel-stream/auto costs the same whether it's planning, stepping through a debugger, or fanning tasks across worktrees. The efficiency features keep earning their place in your context window. They just stop mattering to the bill.

Ten roles, one endpoint

default, plan, slow, task, and the rest can all resolve to the same stream. Your routing table goes back to being about behavior instead of about cost.

Debug sessions that take their time

A DAP session is stack traces, variable dumps, and repeated stepping, and all of it lands in context. On a flat rate the agent can actually finish the investigation.

A calm fallback chain

omp's retry.fallbackChains exist for 429s and quota walls. A busy stream queues instead of erroring, which makes it a steady primary with your old providers waiting behind it.

The deal

What every stream includes.

Frontier models only

You never pick a model; keeping the fleet smart is our job. Every model we serve scores 70% or higher on Terminal-Bench 2.1, or 50 or higher on the Artificial Analysis Intelligence Index, and always at its newest public release.

260K context, guaranteed

Every request gets at least a 260K-token window, and more when the model serving you supports it. Beyond the guaranteed window we compact the middle of the conversation first; your original task and latest turns stay intact.

One generation per stream

Unlimited works because each stream runs one generation at a time. Extra requests wait in a visible queue instead of raising your bill; add streams for parallel capacity.

$5 a month, flat

No token metering, no overage charges, cancel anytime. Sourcing efficiently across the fleet is what keeps the price flat.

The fleet changes as new models launch. The floor doesn't. The full spec sheet lives on the Stream page, and speed numbers are published targets, not a service-level agreement — the fine print is in the terms.

Troubleshooting

Check these first.

Most connection issues come down to the endpoint, API mode, model ID, or key. Start here before anything else.

Set api to openai-completions. The baseUrl ends at /v1, and omp appends the chat-completions path itself.

Write role targets as camel-stream/auto. A bare auto won't resolve without its provider prefix.

apiKey in models.yml is the literal key. If you see auth errors, look for quoting problems or an environment reference that never expanded.

If a YAML edit doesn't seem to take, restart omp. /model shows which model the session is actually using.

Give subagent-heavy work its own stream if you run wide fan-outs. One stream runs one generation at a time.

Frequently asked questions

omp, answered.

More guides

Point another agent at Stream.

Hermes Agent

Built for the long loop.

Read the guide OpenCode

A provider you can check in.

Read the guide OpenClaw

An always-on agent needs an always-on price.

Read the guide

Give omp the whole stream.

Unlimited frontier intelligence over the API format omp already speaks, for $5 a month per stream.

Get your API key