Self-hosted & private
Inference runs on local Ollama. By default no ticket text is sent to any third-party API, so customer data physically stays on your box — the cleanest path to GDPR, HIPAA, and air-gapped support.
Self-hosted · open source · runs on Ollama
Resolvr — Self-Hosted, Open-Source AI Customer Support Agent (Runs on Local Ollama)
Resolvr is an open-source, self-hosted AI customer support agent that classifies a support ticket, retrieves knowledge-base articles via RAG, decides resolve-vs-escalate behind a safety gate, and drafts a reply — running on local Ollama (qwen2.5:14b) so customer data never leaves your server, at zero per-token cost.
What is Resolvr?
An agent-first support pipeline you own end-to-end — not an inbox with a chatbot bolted on.
Open source & hackable
The whole pipeline is on GitHub — FastAPI + React, MIT-style. Read it, fork it, swap the model, point it at your own knowledge base.
Zero per-token cost
No per-resolution or per-token fee. The only cost is the server you already run. An opt-in cloud-API fallback covers the hardest tickets when you want frontier quality.
Safety-gated, not reckless
Honest scope: Resolvr does autonomous triage, RAG answering, and safety-gated drafting + escalation — not end-to-end account changes or refunds. Tool/action execution is on the roadmap.
Stack: FastAPI (Python) + React/Vite · Ollama embeddings + cosine RAG · deterministic classify/decide · pytest eval-gated.
How does Resolvr work? The classify → RAG → safety-gate → draft pipeline
Four tools run in sequence. Classification, retrieval, and the resolve/escalate decision are deterministic; the LLM only writes the final reply.
1 · Classify
classify_ticket
A deterministic classifier tags the ticket with category (billing, technical, account, how-to, refund, security, legal, abuse), intent, sentiment, and priority. No LLM call — so the routing is reproducible and testable.
2 · Retrieve (RAG)
search_kb
Semantic retrieval over your help-center articles: the query is embedded with a local Ollama model and ranked by cosine similarity. Grounding the reply in real KB passages is what stops the agent from inventing policy.
3 · Decide
decide_action
A configurable confidence threshold decides resolve-vs-escalate. A hard safety gate forces escalation on security, legal, abuse, and refund tickets; low retrieval confidence escalates with full context attached.
4 · Draft
draft_resolution
For resolvable tickets, local Ollama drafts a reply strictly from the retrieved KB. Escalations get a courteous holding reply plus an internal routing note for the human who picks it up.
Can it run fully self-hosted so customer data never leaves the box?
Yes. Because inference runs on local Ollama, no ticket text is sent to any external model by default — which is a materially stronger privacy story than “privacy-first” SaaS that PII-masks and then sends the rest to a cloud LLM.
- No third-party processor — there's no external API call to cover with a DPA, which simplifies the GDPR and data-residency story for EU SaaS and regulated teams.
- True local inference — data physically stays on your infrastructure, so it fits air-gapped and sovereignty requirements.
- Opt-in fallback — the cloud-API tier is off by default, so the privacy guarantee holds unless you explicitly enable it.
What does it cost vs Intercom Fin or Zendesk AI?
Resolvr has zero per-token and zero per-resolution cost — the only cost is the server you already run. That's the wedge for teams hit by per-resolution SaaS pricing, where bills scale directly with ticket volume.
One honest caveat so this stays credible: below roughly a couple of million tokens a day, a cloud API can be cheaper than running dedicated GPU infrastructure. Resolvr's value is privacy and control plus no marginal cost at volume — and the opt-in cloud fallback covers low-volume or hardest-ticket cases without locking you in.
Resolvr vs Intercom Fin, Zendesk AI, Chatwoot & Zammad
As of 2026, based on public documentation. The dimensions that matter for a self-hosted, privacy-first deployment.
| Product | Self-host | Local LLM | Data leaves box | Per-resolution fee | Built-in RAG | Resolve/escalate gate | License |
|---|---|---|---|---|---|---|---|
| Resolvr | Yes | Yes (Ollama) | No (local by default) | None | Built-in | Yes | Open source |
| Intercom Fin | No (SaaS) | No | Yes (cloud) | ~$0.99/resolution | Yes | Handoff only | Proprietary |
| Zendesk AI | No (SaaS) | No | Yes (cloud) | Per-resolution add-on | Yes | Handoff only | Proprietary |
| Chatwoot + Captain | Yes | Varies | Depends on model | None (self-host) | Add-on | Helpdesk handoff | Open source |
| Zammad | Yes | Varies | Depends on model | None (self-host) | Add-on | Helpdesk handoff | Open source |
Chatwoot and Zammad are excellent self-hostable helpdesks with AI add-ons; Resolvr is the agent itself — the classify → RAG → resolve/escalate pipeline — rather than the inbox around it.
When should an AI support agent resolve vs escalate?
A safety gate is a rule layer that overrides the model: certain ticket categories must always reach a human, and anything the agent isn't confident about is escalated rather than guessed.
- Hard gate — security, legal, abuse, and refund tickets always escalate, regardless of how confident retrieval is.
- Confidence threshold — when the best KB match is weak, the ticket escalates with full context attached instead of getting a shaky answer.
- Eval-gated — a pytest suite enforces 100% must-escalate recall on security/legal/abuse and ≥90% action accuracy before anything ships.
That combination — RAG grounding plus a confidence-thresholded gate — is the direct answer to the “but won't it hallucinate?” objection: low-confidence cases go to a human instead of being sent.
Build it yourself: a self-hosted AI support agent on Ollama (FastAPI + React)
The whole thing is open source. Four steps from zero to a running agent.
- Pull local models —
ollama pull qwen2.5:14bandollama pull nomic-embed-text. - Clone & install — clone the repo and run
make setupto create the Python venv and install the frontend. - Index your KB — drop your help-center articles in; Resolvr embeds them on first boot for cosine-similarity retrieval.
- Run it —
make devstarts the FastAPI backend and React frontend; submit a ticket and watch classify → retrieve → decide → draft.
Full instructions, the eval suite, and a Docker Compose setup are in the GitHub README. Stars welcome — they help the project rank.
When Resolvr is not the right fit
- You don't want to own a GPU/server — a fully managed SaaS like Fin or Zendesk AI will be less operational work.
- You need frontier-model quality on the hardest, most ambiguous tickets — local models trail GPT-class models there (use the cloud fallback, or a managed tool).
- You need turnkey Zendesk/Intercom inbox integration today — Resolvr ships as the agent + a demo UI, not a full helpdesk.
Who it is for: privacy- and data-residency-driven teams, regulated EU SaaS, teams frustrated by per-resolution pricing, and developers who want a self-hostable, hackable agent they fully control.
Frequently Asked Questions
What is Resolvr?
Resolvr is an open-source, self-hosted AI customer support agent that classifies a support ticket, retrieves knowledge-base articles via RAG, decides resolve-vs-escalate behind a safety gate, and drafts a reply. It runs on local Ollama models, so customer data never leaves your server, at zero per-token cost.
Can I run an AI support agent fully self-hosted so customer data never leaves my server?
Yes. Resolvr runs inference on local Ollama by default, so no ticket text is sent to any third-party API. This makes it suitable for GDPR, HIPAA, and air-gapped deployments. A cloud-API fallback exists but is opt-in and off by default.
Is there an open-source alternative to Intercom Fin or Zendesk AI that I can self-host?
Resolvr is an open-source, self-hostable alternative to Intercom Fin and Zendesk AI. Unlike those SaaS tools it runs on your own server with no per-resolution or per-token fee, and unlike helpdesks such as Chatwoot or Zammad it is an agent-first classify → RAG → resolve/escalate pipeline rather than an inbox with bolt-on AI.
How much does Resolvr cost per ticket?
Resolvr has zero per-token and zero per-resolution cost — the only cost is the server you already run. This contrasts with per-resolution SaaS pricing such as Intercom Fin's roughly $0.99 per resolution. Note that below a couple of million tokens a day, a cloud API can be cheaper than dedicated GPU infrastructure, which is exactly what the opt-in cloud fallback is for.
How does Resolvr decide whether to resolve a ticket or escalate to a human?
Resolvr makes a deterministic resolve-vs-escalate decision behind a configurable safety gate. A hard gate always escalates security, legal, abuse, and refund tickets, and low retrieval confidence escalates with full context. Everything else is auto-resolved with a reply grounded in retrieved KB articles.
Do AI support agents hallucinate, and how does Resolvr prevent it?
Resolvr reduces hallucination two ways: the reply is grounded strictly in knowledge-base articles retrieved via RAG, and a confidence-thresholded safety gate hands low-confidence cases to a human instead of sending them. Its pytest eval suite enforces 100% must-escalate recall on security, legal, and abuse tickets.
Which open-source LLM does Resolvr run on Ollama?
Resolvr defaults to qwen2.5:14b on local Ollama and works with other Ollama models such as Llama 3.1 and Mistral. Smaller 7B–8B models run on modest GPUs; the hardest tickets can use the opt-in cloud-API fallback when needed.
Is self-hosted AI customer support GDPR compliant?
Self-hosting is the cleanest GDPR story because no customer data is sent to a third-party processor — there is no external API call to cover with a DPA. Because Resolvr runs true local inference (not PII-masking before a cloud call), ticket data physically stays on your infrastructure, supporting GDPR, data-residency, and air-gapped requirements.
What stack is Resolvr built on, and how do I deploy it?
Resolvr is built on FastAPI (Python) and React/Vite, with Ollama-based embeddings and cosine-similarity RAG, deterministic classification, and a pytest eval-gated safety layer. It is open source on GitHub and can be deployed end-to-end on your own server; the build guide includes copy-pasteable Ollama and FastAPI steps.