Run your agent on your model
Claude Code, codex and opencode, on a model you own. Run what you own instead of renting it.
Terminal launcher for coding agents
Save on your AI bill*. Free to run your own models†.
Send the work that does not need a premium model to models you run yourself or to cheaper tiers, and keep hosted models, with failover, for the rest.
14-day trial · 30-day refund · never bricked: falls back to free
$29 is for the first 100 customers, then $39/year.
* Savings equal the share of your tokens you route to models you run or to cheaper tiers, times your rate. They depend on your workload, model quality and hardware, so no percentage is promised; see the calculator for your own figure. † No per-token fee on models you run yourself; hardware, power and your time are your cost. Oaica does not host models or provide inference.
What it is for
Four jobs, one launcher.
Cut or cap the token bill
Route the work you can off the meter. Only what stays hosted is billed per token.
Keep proprietary code home
Prompts go only where you point them. No telemetry, so your code stays on your machine.
Mix local and hosted
Use both together, with failover that moves a live reply when a leg goes down.
The two parts
Oaica Code launches it.
Oaica Engine runs it.
Two pieces, one installer. You do not have to pick between them.
Oaica Code
The terminal launcher, a fork of Ollama's client. One config puts your agent on the model you choose, splits the primary and subagent tiers across different backends, and fails over when a leg goes down. Free to install, no account.
Oaica Engine
Our own inference engine, written in Rust and CUDA and developed here, not inherited. It serves the .oqm models and offloads experts to system RAM when the GPU is small, so a 35B model runs on an 8 GB card from a desktop install.
Specifications of the shipped CLI — not usage statistics, not benchmarks
- 15+Agent CLIs, and more added
- 44Provider presets
- 6Route policies
- 4Model tiers
- 0Telemetry requests
- 20hUpdate check, at most once per 20 hours
The idea
One meter, or two paths.
Every token you send to a hosted model is billed. oaica sits on loopback and lets you decide which requests go to a model you run.
Cut the bill
Route the work you can to a model you own. Only what stays hosted stays on the meter.
- Four model tiers in one launch
- Six route policies
Token savings
- Probes the real context window, so prompts are not truncated and retried
- Titles, topic detection and tool-output digestion go to the background tier
- Failover keeps the stream, so a lost turn is not re-sent
- The oversize tier takes over only when the window fills
Keep your code
Prompts go only to the endpoint you point it at. No telemetry, and your provider key never leaves the machine.
Never locked in
15 agent CLIs and 44 provider presets behind one config. If you stop paying, it falls back to free and never bricks.
Savings calculator
Do the arithmetic on your own numbers.
Enter what you use and what you pay. Nothing here is a benchmark or a promise.
Tokens off the metered bill, per month
24M hosted16M routed to you
- Dollars moved off your bill
- $96.00
- Licence per month, annual plan
- $3.25
- Net after licence
- $92.75
- Net on the $7 monthly plan
- $89.00
- Background and tool-heavy tokens
- 10M
If 40% of your tokens run on a model you host, your hosted bill is 40% lower. Enter yours.
Illustrative: arithmetic on the numbers you enter (moved = spend x share), not a benchmark or a promise. Savings depend on how much work you route to models you run, their quality for your tasks, and your hardware and power costs. Oaica does not host models or provide inference.
14-day trial · 30-day refund · never bricked: falls back to free
Pricing
One flat licence. The model is yours.
Free
$0
Never expires. Install with no account.
- Local models and
oaica serve - Your own endpoints and API keys
- Subagent tier and a saved plan
AnnualFirst 100
$29 first year
$39 list price; $29 for the first 100 customers. Renews at $39/year; cancel the renewal any time and access runs to the paid-through date.
- The full paid set
- Three devices per licence
- 14-day trial first, no card
- 30-day refund on the first payment
Monthly
$7 per month
The same product as annual. Switch to annual at renewal.
- The full paid set
- Month to month
- 14-day trial first, no card
- 30-day refund on the first payment
Bundle
Oaica Code + Oaica Engine, one installer
The licence covers both. The installer wires the engine in as a leg, so your subagent tier can run on a model you own — and nothing on that leg is billed per token.
- Oaica Code — the launcher, every paid tier
- Oaica Engine — included, no extra charge
.oqmmodel packs — optional, priced per pack, yours to keep
$29 first year
Engine included. Packs sold separately or run your own.
14 days
The whole paid set, free. No card taken.
30 days
Refund on the first payment.
Never bricks
If payment lapses, it falls back to the free tier.
3 devices
Per licence.
Teams, per seat per year: 1 $39 · 5+ $32 · 20+ $27 · 50+ $22 · 100+ $17 · 250+ $12. Questions: [email protected]
Free and paid, line by line
| Capability | Free | Paid |
|---|---|---|
Local models and oaica serve | Included | Included |
| Your own endpoints and API keys | Included | Included |
Subagent tier, --sonnet-model | Included | Included |
A saved --plan | Included | Included |
| Route policies that pin routing | Included | Included |
| Your own Claude login | Included | Included |
Background tier, --haiku-model | Not included | Included |
Context escalation, --oversize | Not included | Included |
Weighted routing, --shard | Not included | Included |
Guided wizard, --wizard | Not included | Included |
| Policies that pick a leg automatically | Not included | Included |
| Vendored paid APIs | Not included | Included |
How it works
Three commands to a running agent.
One config covers all 15 agents.
Install
The installer checks the archive against SHA256SUMS before extracting.
macOS / Linux$curl -fsSL https://oaica.com/install.sh | bash>irm https://oaica.com/install.ps1 | iexPoint at a model
Pull a small model and serve it, or register an endpoint you already have.
Local$oaica pull qwen2.5-0.5b && oaica serve qwen2.5-0.5b$oaica remote add mine --base-url https://your-endpoint/v1 --api-key-env MY_API_KEYLaunch
The wizard offers to save a plan. oaica doctor checks every leg any time.
$oaica launch claude --plan work$ oaica launch claude
? primary model qwen3-coder-30b (local)
? route policy local-first
? save as plan work
✓ launching Claude Code → http://127.0.0.1:11434
illustrative
? primary model qwen3-coder-30b (local)
? route policy local-first
? save as plan work
✓ launching Claude Code → http://127.0.0.1:11434
illustrative
VERSION.txt (no body, no identifiers; off with OAICA_NO_UPDATE_CHECK=1), and, on activated installs only, a licence revalidation sending the stored key and activation id. No telemetry, crash reports or analytics exist to switch off.- Tier 1
Primary
Free--model
The main model for the session. - Tier 2
Subagent
Free--sonnet-model
Execution. - Tier 3
Background
Paid--haiku-model
Titles and topic detection. - Tier 4
Oversize
Paid--oversize
Takes over when the context window fills.
Six route policies
- local-first (default)
- remote-first
- auto
- local-only
- remote-only
- weighted
--shard) are part of the paid set.- claude
- chatgpt
- hermes
- openclaw
- opencode
- hermes-desktop
- codex
- copilot
- omp
- cline
- droid
- dsh
- pi
- pool
- qwen
Local Ollama server
llama.cpp via
oaica serveAny OpenAI-compatible endpoint
A machine you control
Questions people ask first
Does it see my code or prompts?
No. They go only to the endpoint you selected. There is no telemetry, crash reporting or analytics. Two small network calls exist, an update check and, on activated installs, a licence check; both are shown in the data-flow figure above.
What if I stop paying?
It drops to the free tier and never bricks. Local models, your own endpoints and a saved plan keep working.
Is $29 a subscription?
Yes. $29 for the first year for the first 100 customers, then $39/year, which is also the list price. Cancel the renewal any time; access runs to the paid-through date. A $7 monthly plan covers the same product.
Which agents work?
Fifteen agent CLIs ship in one binary, including claude, codex and opencode. Forty-four provider presets are compiled in, and any OpenAI-compatible endpoint can be added.
Does it work offline?
Yes. The licence verifies offline against an Ed25519-signed claim with the key pinned in the binary, and local models run offline. A remote leg needs a route to its endpoint.
How do refunds work?
Ask within 30 days of your first payment. The 14-day trial needs no card. Full wording: terms and refunds.
Put your coding agent on a flat licence.
$29 for your first year, then $39. First 100 customers. 14-day trial, no card. Ask within 30 days of paying and you get your money back.
14-day trial · 30-day refund · never bricked: falls back to free. Try free, no card