Model-backed conversation
Connect your compatible inference endpoint for incremental replies. Without a configured model, the ordinary path is labelled demo logic.
Model and speech configuration →creekX Agents · v0.10 early access
Support conversations.
Evidence you can inspect.
A private workspace for guided support, source-backed excerpts and operator follow-through. Built to work beside your creekX media infrastructure.
Separate application · Node.js 22.13+ · Single workspace · Assisted pilot
Choose the right path
Three modes. A clear job for each.
Flow Studio
Guide intake through deterministic prompts, captured fields and branches. Test scenarios and coverage before publishing a version.
Build a guided flow →Evidence Studio
Search a published knowledge snapshot locally and quote excerpts with numbered citations. Inspect matching terms, source hashes and scores.
Publish an evidence profile →Flow and evidence modes are separate session choices in v0.10.
From draft to release
Test the boundary before customers find it.
- PreparePublish verified knowledge and define supported questions.
- EvaluateCheck expected sources, forbidden results and explicit abstention.
- ReleaseReview a signed publication preview tied to the current draft and tests.
- InvestigateTrace a session to its fixed profile version and source evidence.
Keep operators in the loop.
Conversation review
Inspect transcripts, assign work and preserve review context. Runtime evidence reports show citation metadata without question and answer text.
Confirmed local tickets
Review and confirm a ticket proposal before it is saved. Track local ownership and follow-up; external help desks require separate integration.
Operator text handoff
Queue a waiting conversation and reply from the operator console. This does not establish a telephone transfer or joined human audio room.
Evaluate what ships today.
| Available for evaluation | Still outside verified scope |
|---|---|
| Roles, SQLite snapshots, private exports and encrypted backup tools | Customer tenancy, document-level permissions, SSO/MFA and full deletion/revocation workflows |
| Recorded-turn speech endpoints and optional browser speech | Full-duplex conversation, physical-device acceptance and guaranteed low audio latency |
| Explicit model snapshots when vision is configured | Continuous video agents, avatar rendering and carrier transfer |
| Plan/quote previews and completed-operation counts | Payment processing, authoritative minute billing and a published Agents subscription price |
Evidence retrieval is lexical and extractive. Scores are not confidence probabilities, and citations do not prove factual correctness. Existing snapshots retain copies after original sources are archived.
A release with a review record.
v0.10 local verification: 200 automated checks passed, zero failed, one physical-device check skipped. The reference evidence benchmark passed 4/4 cases with MRR 1.000. Those results apply to the tested local release, not a production SLA.
Verification record · Archive SHA-256 · Deployment boundaries
Documentation for the whole pilot.
- Documentation overview
- Download and quick start
- Evidence and benchmarks
- Guided workflows
- Models, speech and visual input
- Security and deployment
- Operator API
- Operations and recovery
Read the release introduction and the pilot acceptance checklist.
Start with one useful job
Build a supervised pilot.
Tell us the support task, data requirements, expected users and intended hardware. Agents is offered through a separate early-access evaluation. Media Server pricing and trial terms do not establish an Agents entitlement.