Use your ChatGPT subscription

LumiBot can connect to your own ChatGPT account through OpenAI’s supported Sign in with ChatGPT flow. Choose this connection explicitly for an agent:

python -m lumibot.components.agents.chatgpt_auth login
python -m lumibot.components.agents.chatgpt_auth status
python -m lumibot.components.agents.chatgpt_auth models

researcher = self.agents.create(
    name="researcher",
    model="chatgpt/gpt-6-luna",  # Choose a slug returned by the models command.
    reasoning_effort="medium",
    system_prompt="Research evidence available at the strategy's current time.",
    allow_trading=False,
)

No OpenAI API key is needed for this connection. It does not fall back to an API key or BotSpot’s managed inference. Existing explicit provider models retain their own authentication. Broker, price-data and research permissions are separate from model authentication.

Model-limit discovery uses OpenAI model metadata without starting another login flow. The agent keeps its selected chatgpt/ connection for inference; missing model-limit metadata does not invent a capacity or change the funding source. Parallel tool calls retain their exact completed provider arguments even when the stream supplies no incremental argument events.

Choose your ChatGPT usage policy

Your ChatGPT connection can use included plan allowance and, when you allow it, your existing ChatGPT credit balance. Both use the same OAuth connection; neither requires an API key. Reaching zero remaining plan allowance does not mean inference is unavailable when credit use is enabled.

Review ChatGPT Usage before running evaluations. If you want included allowance only, turn off Allow other apps to use credits after reaching your usage limit. This setting applies across connected apps. If existing credits are authorized, leave it enabled. LumiBot does not change this account-wide setting, and its public inference API does not provide a documented way to verify it.

Exhausted or revoked plan permission produces a visible error. It does not create successful eval freshness or silently select another funding source. No automatic purchase, reset or credit reload is performed by LumiBot.

Profiles and sign-out

Profiles live outside the repository in ~/.config/lumibot-chatgpt by default. On POSIX systems credential files are owner-only; keep the containing directory private too. On Windows, keep the user-profile directory’s access controls restricted to your account. Never commit profiles, share them with other users, or copy credentials into a public issue or workflow artifact.

POSIX credential reads reject symlinks, hard links and unsafe file ownership or permissions. Authentication and run-usage locks recheck their paths before each acquisition, including containing directories and ancestors. Writable ancestors are rejected except root-owned sticky system temporary directories. Legacy owner-controlled readable locks remain usable; new locks are owner-only.

Login reconnects the active registration. Use login --profile <profile-id> to reconnect another existing profile or login --new-profile to create a separate registration for another account. Returning registrations validate the selected account rather than silently switching identity. In-flight processes pin their selected profile; changing the active selection does not move running agents to another account.

python -m lumibot.components.agents.chatgpt_auth logout

Sign-out revokes the renewable grant when possible and clears local tokens even if remote revocation cannot be confirmed. The registration and account identity remain so reconnection can reuse them. Refresh rotation is serialized across local processes.

Headless and self-hosted use

login --no-browser starts the same bounded loopback listener and writes a private pending-browser.json link. Forward the host’s loopback callback port and open that link on your trusted browser. Preserve the registration’s stable host identity and protect any transferred profile. Follow OpenAI’s self-hosted VM instructions. Do not extract Codex credentials or use private ChatGPT endpoints.

LUMIBOT_CHATGPT_HOME selects a private configuration directory. LUMIBOT_CHATGPT_PROFILE selects an existing profile ID. LUMIBOT_CHATGPT_PROFILE_PATH explicitly selects a protected profile file from your own supported SIWC registration. It is useful for an isolated developer evaluator; it does not bundle or distribute that registration.

Commercially hosted customer inference needs an approved product integration and each customer’s consent. A developer’s local subscription is not a shared funding account for customer bots. BotSpot’s research MCP connection is also a separate grant; signing into ChatGPT does not authorize deployment or trading.

Bounded test-worker credentials

A host running an authorized test can select a nonrenewable run profile through the existing LUMIBOT_CHATGPT_PROFILE_PATH setting. This does not transfer a credential, create a remote worker, change an agent’s selected model, or enable subscription inference for ordinary provider models. The agent must still use an explicit chatgpt/<model slug> connection.

The private run profile contains credential_kind: "run", a unique run_id, status: "connected", the granted scopes including chatgpt.tokens.use.direct, one authorized OAuth access_token, an absolute Unix expires_at, one model model slug, and positive integer limits for calls, input_tokens and output_tokens. It may be valid for at most one hour when loaded for inference. It must not contain a refresh token or an API key. The trusted host prepares it from its authorized connection; do not copy a renewable account profile into a test artifact or public workflow.

The adapter reserves the selected model catalog’s full input/output envelope before each actual HTTP request, including tool continuations. Unknown model limits fail before the request. A smaller caller output limit cannot reduce the reservation because this subscription endpoint does not accept that limit. Completed Responses usage settles the reservation; interrupted streams, missing usage and failed requests keep it. Output counts already include reasoning.

An owner-only sibling .usage.jsonl file records reservations and usage, without prompts, tool arguments or access tokens. Concurrent processes using that profile share its lock and limits. Restarting cannot reset use or enlarge the original policy. Expired credentials refuse new requests, but existing usage can still be recorded and inspected. Run profiles never refresh, switch accounts, or fall back to API billing.

These are client-side limits, not a newly scoped or independently revocable OpenAI credential. The underlying bearer token retains its original permission and provider expiration. Protect it through the host’s existing secret channel. The controlling test runner must reserve each worker’s whole allowance in its own cumulative ledger before provisioning it, and collect the usage receipt afterward. Separate workers do not share a local ledger automatically.

This SDK support alone does not qualify a hosted integration. Credential transfer, host identity, consent, cleanup and end-to-end inference need their own authorized validation. Existing production provider connections keep their current authentication and billing unless their owner explicitly selects a different connection.

Canonical evaluations

The canonical LumiBot actor and semantic judge both use this subscription connection. Ordinary provider API keys and broker credentials are stripped from the isolated fixture environment. Select a protected developer profile and your authorized ChatGPT usage policy, then run:

python scripts/run_agent_evals.py --case-id research_macro_point_in_time \
    --repeat 5 --max-model-calls 80 --max-cost-usd 0 \
    --output-root /your/private/eval-run \
    --freshness-state /your/private/eval-freshness.json

The append-only call and repetition ledgers permit resume without rerunning passed repetitions. --max-model-calls sets an informational reporting window covering actors, continuations and judges. The runner automatically extends that window without resetting history; it is not a hard spending or request limit. Actual provider quota and account credit settings still apply. Missing authoritative usage retains the reservation. Reports distinguish subscription usage from API spend; zero API spend is not proof of the account’s credit setting.

Only failures with verified no-mutation evidence are treated as retryable infrastructure errors. Deterministic checks still run and retain actual broker receipts. Failures after order actions or with unknown effects, completed decisions and tool failures keep their ordinary grading. Earlier attempts and unknown usage reservations remain in the ledger; live orders are not replayed.

Private CI funding must have its own authorized registration, host identity, protected storage and serialized refresh owner. Public forks receive no refresh credentials. Missing CI authentication fails the release gate; it is not permission to run a paid model or skip the gate.