Product reference
Find the capabilities that connect your tools, project knowledge and requests.
View pricing and full limitsOne base URL and one key bring your everyday coding terminals onto the same path without changing how you use them.
Chat Completions and Responses; existing SDKs and scripts switch by changing one base_url.
Terminals such as Claude Code connect directly through ANTHROPIC_BASE_URL.
Pick Codex CLI, Claude Code, Cursor, Continue and more; copy a config with the key already filled in.
The setup guide waits for the key's first real request and shows success, latency, model and protocol.
SSE streaming, tool calls and multi-turn messages behave like the native API.
Explain something once per project; afterwards any terminal can retrieve and reuse it.
Stable facts in five scopes: stack, structure, conventions, pitfalls, stage summary. Each entry can be disabled.
Five kinds of entries (facts, decisions, preferences, pitfalls, structure) with state, hit count and source; confirm, disable or delete.
With learning consent, conversations can produce suggestions. Review their sources and approve them for selection in relevant requests.
The “tell the assistant” form writes a rule that takes effect immediately.
⌘K searches projects, tasks, memories and requests, narrowable by project.
Control profile injection and memory enhancement per project and compare the effect any time.
Split one big change into parallel, traceable executions with usage and results collected in one place.
In-flight concurrency, burst ceiling and a live pulse per key; overruns return a readable reason.
Name a repository, branch, project and allowed actions; an agent engine executes in the cloud.
Turn-by-turn requests and replies next to a work-state panel of progress, decisions, pitfalls and to-dos.
Ordered steps, each with instructions, allowed actions and a verify command; save as a template and rerun.
Set an equivalent-token ceiling per run; it stops at the limit instead of silently burning allowance.
Route by stage between the fast and the deep model, wasting nothing on small things and settling for nothing on hard ones.
Retrieval, rewriting and summaries go fast; architecture and complex fixes go deep.
High-difficulty turns are escalated automatically, with the reason noted in the request details.
Each key is limited to the models it may use, preventing accidental use of expensive ones.
Models convert at their ratio into one allowance number.
Upstream failures retry by credential weight and health, with every reason recorded.
Repeated system prompts, project profiles and long context can reuse upstream prefix caching to reduce upstream costs. Plan quota is still metered at the model multiplier; cache hits do not reduce quota deductions.
System prompts, profiles and tool definitions are recognised as a stable prefix and cached with no request changes.
A task's history is cached incrementally by turn; long sessions hit more the further they go.
Requests distinguish cached and uncached tokens. Usage shows estimated upstream cache costs; these estimates do not represent reductions in plan quota or user payments.
Set TTL and minimum cache length per project or key; switch it off for sensitive projects.
When a profile or memory changes the matching cache is invalidated so stale context is never reused.
Planned: cache hit rates, estimated upstream cost savings and latency trends by day and Key. Savings estimates do not represent reduced plan quota deductions.
Split keys by purpose; each has its own limits, usage and project.
Create, rotate, enable, disable and revoke; the secret is shown once, at creation.
Model allowlist, IP allowlist, concurrency, RPM and a per-period usage cap.
Select a project explicitly or through a supported client workspace to use its profile and memories. Associating a key alone does not select a project.
Concurrency and burst curves, daily usage bars, recent requests and last use.
Every call lands on one line of the ledger; allowance and spend can be checked any time.
Input, output, cache-hit and cache-miss tokens plus cost and duration.
Today / 7 days / 30 days by day, model or key, with CSV export.
Five plans, 30-day periods, an allowance bar and days remaining at a glance.
Top-ups, subscription charges and failed-payment refunds, line by line.
A reminder at 80% so a task does not stop halfway.
No guessing when something breaks: request body, parameters and failure reason are kept together.
Filter by key, model, status and time; infinite scroll; open the detail drawer.
Parameters, usage and cost, peak marker, failure reason code and message body; copy as curl.
Pick a key and a model and chat; tokens, cost, TTFT and duration update live.
In-flight concurrency, requests this second and this minute against their limits, refreshed every 5 s.
TTFT P50 / P95 and total duration, split by model and key.
Clear account, data and permission boundaries so teams can trust it.
Phone binding and rebinding, WeChat binding, passwords, session list and remote sign-out.
Choose not to store message bodies and keep only metering and status.
Cloud agents only run allowed actions and never exceed the terminal's own permission model.
Concurrency and allowance are checked live; anomalies leave replay and reconciliation records.
Key, plan, top-up and permission changes are logged line by line.
Check tool access and plan requirements, then verify your first request in setup.