AI assistant
On this page
AI assistant
Project conversations#
The assistant works beside the project editor. Each conversation belongs to one project, so its context and Edit operations cannot cross into another source tree. Conversation history can be reopened, renamed or deleted; widen the assistant panel when reading longer answers, diffs or Markdown tables.
At the start of a conversation, OPR supplies the project identity, source snapshot, pinned factor contracts, SDK rules, data semantics and relevant documentation. The user prompt still defines the research question or requested change.
Ask & Edit#
| Ask | Discuss research design, SDK usage, data semantics or an implementation approach. It does not alter the project. |
| Edit | Generate a revision-bound proposal for one or more files. Review the exact source diff before choosing Apply or Reject. |
| Apply | Before writing, the platform confirms that the proposal still matches the current draft, then creates a new revision. Broad or destructive changes require an extra confirmation. |
| Undo | Reverse the applied proposal while preserving the project history. |
Models & BYOK#
Available built-in models appear in the composer; unavailable deployments are not selectable. The catalog includes Kimi K3, DeepSeek V4 Flash, DeepSeek V4 Pro and GLM 5.2. Model and reasoning controls sit beside Send, together with the effective context usage and gateway limits.
BYOK accepts an OpenAI-compatible model at a public HTTPS endpoint. Add the endpoint, model name, API key and declared limits in Personal settings, then verify the connection. Private, loopback and link-local destinations are rejected, TLS is verified and provider redirects are not followed.
Built-in usage follows the quota configured by the administrator. BYOK calls use the user's provider account, while OPR still enforces request size, concurrency and security limits.
BYOK configuration#
| Parameter | Default | Description |
|---|---|---|
| model | — | Exact model identifier sent to the provider's OpenAI-compatible chat-completions request. It must match the provider deployment, not a display nickname. |
| base_url | — | Public HTTPS OpenAI-compatible base URL. OPR targets its chat-completions route, verifies TLS, rejects redirects, and blocks private, loopback, and link-local destinations. |
| authentication | Bearer | Bearer sends Authorization: Bearer <key>. X-API-Key sends the key in the X-API-Key header. Choose the scheme documented by your provider. |
| context_window | 131,072 | Provider-documented total context window, 4,096–2,000,000 tokens. Effective input is bounded by this declaration and the gateway, and estimated input + requested output must fit inside the declared context. |
| maximum_output | 16,384 | Provider-documented maximum generated tokens, 256–262,144, and no greater than the declared context. A single OPR request currently asks for at most 32,768 output tokens; lower account, model or gateway limits still apply. |
| provider_concurrency | 4 | Maximum simultaneous requests allowed by your provider, 1–64. Declaring more does not bypass the OPR gateway and may cause provider-side 429 errors. |
| provider_rpm | 20 | Provider requests-per-minute allowance, 1–10,000. OPR applies the lower effective rate together with its own gateway ceiling. |
Connect an external Agent#
Create a personal credential in Agents, copy the one-time setup instructions, and send them only to the private Agent conversation that needs access. The credential works for the period you choose and can be revoked at any time.
| Parameter | Default | Description |
|---|---|---|
| name | — | A recognizable name for the Agent or task. |
| scopes | least privilege | Choose only what the Agent needs for this task. Read and write permissions are separate. Paper access never grants Live access; Live permissions must be selected explicitly and still cannot activate trading, control a running Live deployment, manage trading credentials, or place real orders. |
| lifetime_days | 30 | Credential validity from 1 to 180 days, default 30. It expires automatically and can be revoked immediately; the secret is displayed only once at creation. |
| Task | Minimum scopes |
|---|---|
| Read Hunter profile | profile:read |
| Read projects and data | artifact:read · data:read |
| Develop and run | + artifact:write · backtest:run · backtest:read |
| Batch research | + study:run · study:read |
| Read paper trading | + paper:read |
| Operate paper deployment | + paper:write |
| Read sanitized Live data (opt-in) | + live:read |
| Prepare an inactive Live configuration (opt-in) | + live:create |
| Cancel runs | + backtest:cancel |
Use the setup instructions#
The copied instructions include the MCP endpoint, Bearer credential, granted permissions and expiry. They direct the Agent to initialize the connection, then discover the current references, tools and task prompts before working.
What an Agent cannot do#
Project editing, runs, Studies and Paper controls require separate permissions. Read permission never implies write permission. Live access is always opt-in: an Agent may read sanitized Live information or prepare an inactive configuration only when those permissions were selected.
Activation, leverage changes, Live start or stop, credential management and real order actions remain user-controlled. The Agent never receives exchange secrets, and Paper permissions do not grant Live access.