anima · docs
ConceptsBrain
Overview
Get started
Concepts
Reference
Operate

Attested inference inside a TEE.#

The brain runs entirely on 0G Compute via the @0glabs/0g-serving-broker SDK. TeeML mode means every inference is signed by the provider's attestation key. If a request hits a tampered host, the broker's signature check fails and the call returns an error. There is no centralized fallback. Fail loud is the policy.

The provider model#

0G Compute hosts a catalog of OpenAI-compatible providers. Each provider runs one model in a TDX enclave. The catalog is live: broker.inference.listService() returns the current set with per-token pricing and the provider's EOA address. At anima init you pick from this catalog. The choice writes to brain.provider (provider EOA) and brain.model (model string) in anima.config.ts. Switch later with anima model.

The default flagship is whatever model 0G Compute features at the top. GLM-5 was first-class through Q1. Qwen3.6 took over. There is no hardcoded default in anima; the wizard pulls live every time.

Source: packages/core/src/brain/og-compute.ts.

How a turn happens#

OGComputeBrain.infer({ event, history }) (packages/core/src/brain/og-compute.ts) builds the request:

  1. Compose the frozen prefix (system prompt + memory index + identity + persona + skill index + tool list + env).
  2. Append the conversation history.
  3. Append the current event as a user message.
  4. Estimate token count. If above compaction.threshold * contextWindow (default 0.5 * 1,000,000), fold older turns via a separate compaction call before the main inference.
  5. Call broker.inference.getRequestHeaders(providerAddress, messageText) to get attested headers.
  6. POST to ${endpoint}/chat/completions with the OpenAI-compatible payload plus the attested headers.
  7. Parse the response. If tool_calls are present, dispatch each one through the tool registry and feed results back. Repeat until the model emits no more tool calls.
  8. Return the final turn to routeLoop.

Default max_tokens is 4096 (DEFAULT_MAX_OUTPUT_TOKENS in og-compute.ts). The default context window is 1,000,000. Both are configurable under brain in anima.config.ts.

The ledger#

The serving-broker maintains a per-agent ledger that prepays for inference. broker.ledger.addLedger opens it (at init), broker.ledger.depositFund tops up, broker.ledger.refund pulls funds back. Each provider has a sub-account; tokens get locked when you transact with that provider for the first time. The locks have a refund window per provider; anima ledger retrieve starts the window, and a second call after the window completes the refund.

anima balance shows the full position in one read-only call: main ledger total, per-provider available, per-provider locked, plus EOA balance and sandbox billing reserve. Use this before topping up so you know what is already locked versus available.

Source: packages/core/src/brain/ledger.ts, packages/cli/src/commands/balance.ts.

Auto-topup#

AutoTopupManager (packages/core/src/economy/auto-topup.ts) is opt-in via economy.autoTopup in config. When enabled, the manager polls every 5 minutes. If a per-provider locked envelope drops below the threshold, it auto-deposits more from the agent EOA, capped by a configured ceiling.

A 10-minute cooldown was added in v0.21.14 to kill the "insufficient wallet" spam loop that happened when the agent EOA ran dry mid-poll. Operator can override the polling interval via config.

Source: packages/core/src/economy/auto-topup.ts.

Vision#

The vision.analyze tool routes screenshots and image files through a separate broker pool. The mainnet vision provider lives at 0x4415ef5CBb415347bb18493af7cE01f225Fc0868 running qwen/qwen3-vl-30b-a3b-instruct. Same TeeML attestation. Same ledger model.

browser.vision is a convenience: screenshot the active agent-browser tab plus route through the same provider in one tool call.

Source: packages/core/src/brain/broker-pool.ts.

Compaction#

When estimated token count breaches compaction.threshold * contextWindow, the brain folds the oldest portion of conversation history into a summary. The summary is generated via a separate broker call with SUMMARY_SYSTEM_PROMPT and max_tokens: 1024. Older turns are replaced by the summary in subsequent inferences. The frozen prefix never compacts.

Source: packages/core/src/brain/og-compute.ts.

Fail loud#

There is no fallback to OpenAI or Anthropic. There is no key-storage layer for non-0G providers. If 0G Compute is degraded, the agent halts and the operator sees the error. The brain queue persists, so when 0G recovers the queued events resume.

Post-MVP, a user-local relay limb is on the roadmap: the user runs anima-limb paired to their agent, the limb holds the external API key locally, and the agent calls a delegate.task tool whose execution is the limb proxying to OpenAI. Keys stay on the user's machine. That is the only acceptable escape hatch given the sovereignty thesis.

Read Tools next.

Source: packages/core/src/brain/og-compute.ts, packages/core/src/brain/ledger.ts.

Source: packages/core/src/brain

Brain · anima docs