NexusDesk connects knowledge retrieval, business tool calls, human handoff and ticketing into a single support workflow — so an agent can answer questions, execute tasks, and hand the judgement calls back to a person.
No Redis. No Kubernetes. One Postgres handles both storage and the task queue; one Compose file brings it up.
These are architecture trade-offs, not a feature list.
Postgres carries data storage, the task queue and the cross-process message channel. No Redis. If you deploy into a customer's intranet, sit through security review, or run with a small ops team, every middleware you drop is one less review, one less runbook, one less thing that can fail at 3am.
Agents do not perform irreversible actions on their own. Take ticket creation: the agent can only draft a ticket, and a person has to approve it before anything is created. A conversation can be handed to a human operator at any point.
Agent configuration is draft / published / rollback, with version diffs. Model calls are metered and quota-limited, runs produce a full event log, and operator actions land in an audit trail.
Organised around what you are trying to do, not around our module list.
Configure and publish agents that run an asynchronous ReAct loop for reasoning and tool calls, with caps on rounds and tool calls.
Bring in documents, configure chunking, preview content, build vector indexes and test retrieval quality.
Manage connections and versioned profiles for chat, embedding and rerank models. Any OpenAI-compatible service works.
Register business APIs and manage credentials, parameter validation and execution permissions, with read and write side-effect classes.
Take over a conversation, draft a ticket, confirm it — so problems that need follow-up enter a real process.
Embed the support widget, use channel endpoints, or integrate through the open platform API.
Inspect run records, tool calls and evaluation results to debug problems and keep improving.
Multi-tenant isolation, role-based access, model call quotas and metering, sensitive-word review.
Applications and rotating keys, signed webhooks with retries, idempotency keys, SSE streaming and call logs.
You need Docker and Docker Compose v2. The trial mode uses a simulated model, so no model API key is required.
# from the repository root
sh nexusdesk quickstart
# from the repository root
.\nexusdesk.ps1 quickstart
Prefer not to clone? A single Compose file is enough — see
Quick start. Once it is up, open
https://localhost:8080 and go to sample-support in the conversations workbench.
We would rather you find out early that this is not your tool than after you deploy it.
| Dimension | NexusDesk | General LLM app platforms | Knowledge-base Q&A platforms | Open-source helpdesk |
|---|---|---|---|---|
| Focus | Support and service collaboration | General orchestration | Knowledge-base Q&A | Human agent ticketing |
| Ticket and human handoff loop | Built in | Build it yourself | Build it yourself | Built in |
| Write actions need approval | Built in | Build it yourself | Build it yourself | Not applicable |
| Redis required | No | Yes | Yes | Yes |
| Agent versioning and rollback | Built in | Partial | Partial | Not applicable |
| Model quotas and metering | Built in | Partial | Partial | Not applicable |
| License | Apache-2.0 | See their repo | See their repo | See their repo |
Details about other projects should be checked against their own documentation. This table describes design differences only and is not an assessment of any product.
We do not have customers we can name yet, so we are not going to write “industry leading”. What we can offer is something else: the entire codebase is public, auditable, and you can deploy it yourself to check. Whether the architecture holds up, whether there is a hidden paid dependency, whether your data ever leaves your network — you can verify all of it without asking us.
About the current state, honestly:
The project is under active development. The core support workflow runs, but it is not yet “production ready out of the box”. Retrieval quality, large-document handling and task recovery are still being validated; observability, backup and recovery, and capacity testing are still being filled in.
We keep shipped capability and roadmap separate, and we do not present the roadmap as if it were delivered. See the validation record in the repository for what has been verified.
Stuck on something? Just ask. Whether it is a failed deployment, a capability question, or “could you do this for our scenario”, reach us through Contact.
One command, two containers, no model API key. If it looks like a fit afterwards, we can talk about private deployment or an industry package.