Agent Zero API costs: main, utility and embedding models

Agent Zero, the open-source AI agent framework, costs $0 — its LICENSE file carries MIT terms, copyright Agent Zero, s.r.o., and agent-zero.ai advertises no pricing page, no plan and no paid hosted tier for it — so "Agent Zero pricing" means the model API bill underneath it. What makes that bill different from a single-model agent is that Agent Zero splits it across three model roles, configured separately and billed separately: a main model, a utility model, and an embedding model. Kunavo serves no embedding model, so only the first two roles are ever bought here, and the third stays on the local default or goes to a provider that sells that step. Getting one chat reply back therefore does not mean the integration is finished — it means one of three roles answered.

One disambiguation first, because it sits on the results page rather than off to the side. Agent Zero Token (A0T) is a cryptocurrency on Base, and it placed fourth on the single search-results draw this page has for "agent zero pricing", recorded September 17, 2026, with a Coinbase price page at seventh. One draw is not a ranking. The token is genuinely the project's own — CoinGecko's official-links block points at agent-zero.ai and at the project's GitHub repository — which is what makes it easy to mistake for a price list. It is not one: no token quote is this software's price, this page gives no investment or trading guidance, and the token appears below only as a stated condition on a free endpoint.

Four things called "Agent Zero" that are not this software's price

Only the first row below was observed on the search-results draw for "agent zero pricing"; the other three are name collisions you meet while researching, not things that page returned.

Sources checked September 21, 2026: the LICENSE file (MIT, copyright Agent Zero, s.r.o.), the PyPI project page, and the LLM API terms, which state that the service "is a free, bonus feature for the Agent Zero community" and "is not a commercial or paid offering", available to token holders who stake their tokens, with quota "determined dynamically and is not guaranteed". The staking threshold and the resulting quota are not published anywhere this page could read, so neither is stated here. On the software itself, the GitHub API for the agent0ai/agent-zero repository reports it as not archived and not disabled with 19,227 stars, and the latest release as v2.12, published September 9, 2026.

Three roles, three separate bills

Agent Zero's Model Presets guide states that every setup contains a main, a utility and an embedding model. That is not a v1 leftover: conf/model_providers.yaml on main has exactly two top-level sections, chat: with 29 providers and embedding: with 16, and both the main and utility roles draw from the chat list. The installation guide adds that the built-in Browser has no separate default model slot of its own.

The default is worth understanding before you move it. In models.py, an embedding call whose provider is huggingface and whose model name starts with sentence-transformers/ short-circuits to a local SentenceTransformer wrapper — no HTTP request and no key — and the curated Default preset pins that role to sentence-transformers/all-MiniLM-L6-v2. The installation guide also states that changing the embedding_llm "will re-index all of A0's memory", which makes that slot a migration rather than a setting — plan it as one, and keep a backup of the memory directory before you touch it. For the wider question of how retrieval quality and storage interact, RAG implementation covers the same decision outside this framework.

Agent Zero custom API: where the endpoint and the key go

On the search draw recorded September 17, 2026, the query "agent zero custom api" put the project's api-integration guide third, and that document answers a different question — its actual title is "Adding External APIs Without Code" and it is about teaching the agent to call third-party APIs, not about choosing your LLM provider. Here is the provider path instead, read from the repository on September 21, 2026.

1. Pick the provider. Its id is other; conf/model_providers.yaml names it Other OpenAI compatible and lists it in both the chat: and embedding: sections. The installation guide spells the same thing OpenAI Compatible, so expect both labels in the interface.
2. Put the key in External Services. The guide's wording is to "add the API key under External Services -> Other OpenAI-compatible API keys, then select OpenAI Compatible as the provider in model settings".
3. Set the API URL on each role you are moving. It is a per-role field, stored as api_base in the preset file, and the guide describes it as the "URL of the API endpoint for the chat model".
4. Type the model id by hand, exactly as your gateway's catalog spells it, and keep the prefix convention straight: native providers take the bare id, gateways usually want a vendor prefix.

# One Model Preset, three independent blocks. Each carries its own
# provider, model name and api_base, so the three roles can point at
# three different endpoints — and its own rl_* rate-limit fields.
# Schema read from agent0ai/a0-presets
# on September 21, 2026; the values below are an illustration.
chat:
  provider: other        # the UI calls it 'Other OpenAI compatible'
  name: <the id your gateway's own catalog lists>
  api_base: "https://api.kunavo.com/v1"
  ctx_length: 200000
utility:
  provider: other
  name: <a cheaper id from the same catalog>
  api_base: "https://api.kunavo.com/v1"
  ctx_length: 128000
embedding:
  provider: huggingface  # shipped default: local, CPU, no HTTP call
  name: sentence-transformers/all-MiniLM-L6-v2
  api_base: ""

Under the hood, models.py rewrites the provider immediately before the call — the comment in the source reads remap other to openai for litellm — so a custom endpoint must speak the OpenAI Chat Completions shape, and every chat provider in the catalog carries a0_api_mode: chat. LiteLLM is the transport for all of it, pinned in requirements.txt as litellm==1.88.1 with the comment "CVE-2026-42271 fix: patched floor is 1.83.7". If you would rather not type the credential into the interface, the environment route is read in this order:

# models.py get_api_key() reads these three spellings in this order,
# for provider id `other`. A comma-separated value is round-robined
# across calls. Read from main on September 21, 2026.
API_KEY_OTHER=your-key
# OTHER_API_KEY=your-key
# OTHER_API_TOKEN=your-key

One Docker trap worth repeating from the installation guide: inside the container, localhost and 127.0.0.1 in an API base URL mean the Agent Zero container, not your host, so a model server on the host needs http://host.docker.internal:<port> or the bridge address. That applies to local runtimes, not to a public HTTPS endpoint.

Two defects in the official setup documentation are worth knowing before you follow it literally, because both were confirmed on September 21, 2026. First, the installation guide's A0 Launcher download links are hardcoded to v0.9 and the text points at "the A0 Launcher v0.9 release", while the launcher's own releases page shows v1.7, published September 9, 2026 — eight releases newer. Use the releases page rather than the pinned download URLs. Second, the guide's "Model Naming by Provider" table puts an Anthropic model id in the example cell of its OpenAI row; the surrounding prose has the distinction right. Take the general rule from the prose, not from that table row: a native provider takes the bare model id, and a gateway usually wants a vendor prefix.

Four limits a custom endpoint runs into

The caching gate is the one most likely to change a budget, so here is exactly what the source says. _supports_cache_control_markers returns true only when the api_base contains openrouter.ai or anthropic.com, or when the normalised provider is in a fixed set that includes anthropic, bedrock, gemini, minimax, openrouter and vertex_ai. The other provider normalises to openai, which is not in that set. The second path, _is_openai_prompt_cache_provider, returns false for any api_base that is not api.openai.com or openai.azure.com. So for a base URL that carries none of those four host names, and whose provider normalises to openai, Agent Zero applies neither Anthropic-style breakpoints nor OpenAI cache keys. Scope that carefully: it is a statement about what Agent Zero sends, not a claim that a cache discount is unavailable.

On the two bug reports, state them as what they are. Issue #1769 is a user report, open on September 21, 2026, titled "LiteLLM calls exit(-9) when a model in model_cost has a litellm_provider that differs from the provider's actual litellm_provider mapping"; its own scope line says it affects "any provider that resells or proxies models from multiple origins", which describes a gateway. Nobody reproduced it for this page, and whether it triggers for any specific model id is untested. PR #1437 reports that LiteLLM 1.80.11 and later send encoding_format: null when it is unset and that strict validators reject that with a 422, plus an un-truncated memory-recall query that can overflow an embedding model's context. It is open and unmerged; its parent issue #1436 is closed, and how that issue was resolved could not be determined, so neither "fixed" nor "live" is claimed here.

A worked per-role cost estimate

These are illustrative token arithmetic, not measured task costs and not a bill ceiling. Assume one week of steady use: 6 million uncached input tokens and 300,000 output tokens on the main role, and 1.5 million input and 150,000 output tokens on the utility role. That split is an assumption for illustration — nobody metered an Agent Zero loop for this page. Rates are live Kunavo catalog prices per million tokens.

Two readings, and they point in opposite directions. First, under these assumptions the main-role choice spans $10.14 a week between the dearest and cheapest rows above, while the utility-role choice spans $0.73. Optimising the second slot is not where the money is unless your workload is unusually background-heavy — and the installation guide warns against going too small there anyway, since very small models (it names 4B) often fail at reliable memory extraction. Second, the two roles combine independently: GPT-5.6 Terra with GPT-5.6 Luna estimates at $5.63 a week against $16.50 for GPT-5.6 Sol with Claude Haiku 4.5. The third role adds no API charge at all while it stays on the local CPU default.

For calibration against the project's own choices, Agent Zero downloads three curated presets at first start when no saved collection exists. Default pairs a GPT-5.6 Terra chat model with a Google flash-lite utility model; Power pairs GPT-5.6 Sol with GPT-5.6 Luna; Efficiency pairs a GLM chat model with a DeepSeek utility model. Only Default spells the embedding role out, and it pins the local CPU model; Efficiency and Power carry no embedding block at all in that file. Those refs are written in the OpenRouter style that file uses, so none of them is a Kunavo id as written — use the ids Kunavo's own catalog lists. That file was last pushed July 17, 2026, so its model choices may lag the market.

Scale by your own weeks before treating any of this as a budget. Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges and external tools sit outside this example. The minimum Kunavo top-up is $10 in prepaid credit — a funding minimum, not a task fee and not a subscription. See billing details.

Best API for Agent Zero: which route wins when

One first-party detail makes the role-by-role argument concrete rather than theoretical. The Cline Pass plugin exists partly because a subscription API that only streams breaks the utility calls specifically — the plugin injects a no-op streaming callback so background work such as chat naming, compaction and memory consolidation keeps working. Roles fail separately, which is the whole reason a single successful chat reply is not an integration test. Agent Zero also ships an LLM Cost Tracker plugin that groups usage by call class — main chat, subordinate agents and utility among them — and ships with no prices at all, telling you to "Fill in rates from your provider's official pricing page". Fill it in per role, or the dashboard bills every class at whatever single rate you typed.

Set it up, then verify all three roles

Kunavo has not runtime-tested Agent Zero against its endpoint, and publishes no Agent Zero integration page; everything above was read from Agent Zero's own source and documentation on September 21, 2026. So verify by role rather than by first reply: send one chat message for the main role, then trigger background work — start a new conversation so automatic naming fires, or let a session run long enough to compact — and confirm the utility model was the one that answered. Leave the embedding role on the local default unless you have decided to migrate it, because that change re-indexes memory. Keep a working route available throughout, run one bounded task, and read the charge your account recorded against the cost tracker's own estimate.

If you are routing the two paid roles at one endpoint, the quickstart has the base URL and a first request, and creating a Kunavo account takes the $10 prepaid minimum with no subscription. For the wider decisions: the OpenAI-compatible API reference covers the endpoint shape Agent Zero speaks, Agent Zero vs OpenClaw compares the two harnesses on execution and memory, LiteLLM alternatives covers the transport layer every Agent Zero call passes through, and AI cost optimization covers measuring a real task instead of a rate card.

FAQ

Agent Zero repository state, latest release, issue and pull-request status, provider catalog, models.py, the LiteLLM transport, requirements, the curated preset file, the installation and troubleshooting guides, the LLM API terms and the PyPI collision were all checked September 21, 2026. Search-result positions come from one draw recorded September 17, 2026, and are a single draw, not a ranking. Kunavo token rates are read from the live catalog; every dollar figure on this page is illustrative token arithmetic on a stated assumption rather than a measured task cost.