Your AI-native consulting agency running 24/7 for you.
26
stars
121
commits
Python
primary language
Sep 10, 2026
updated
README
Your AI-native consulting agency, running 24/7 for you.
Works with GitHub, GitLab, any OpenAI-compatible model and any MCP client.
See it live: berwick.ai is a real consulting practice running on Openvisor, white-labelled and in production.
How it works · What you get · Quick start · Production · Bring your own model · Make it yours · Docs
Watch the full demo (62 seconds, MP4)
Quick start
With your AI assistant. Copy/paste:
Read instructions at https://github.com/scalevisor-io/openvisor/blob/main/docs/DEPLOY_WITH_AI.md to deploy my own Openvisor instance.
Or by hand.
git clone https://github.com/scalevisor-io/openvisor.git && cd openvisor
cp .env.example .env # update envs!
echo "127.0.0.1 openvisor.local app.openvisor.local mail.openvisor.local mcp.openvisor.local" | sudo tee -a /etc/hosts
make dev
| Landing | http://openvisor.local |
| App | http://app.openvisor.local (sign in with ADMIN_EMAIL / ADMIN_PASSWORD) |
| http://mail.openvisor.local (Mailpit) |
Two switches worth knowing: the sign-in captcha only runs in a secure context, so on plain http set ALTCHA_ENABLED=0; and OPENHANDS_ENABLED=0 (the default) pushes a deterministic scaffold through the real PR → merge → demo path at zero token cost, 1 builds with the agent.
How it works
Coding agents write code. Openvisor is the consultancy around them: the customer portal, the review gate, the sandbox, the demo and the billing. You stay the consultant: the agent can never price, approve or advance a project.
| Step | What happens | |
|---|---|---|
| 1 | Deposit | A customer describes the project on your white-label site, answers your onboarding questions and connects a GitHub or GitLab repo. |
| 2 | Evaluate | Moderation, feasibility and a credit estimate run automatically. Doubts are flagged to you, never auto-accepted. |
| 3 | Price | You approve or answer with a direct quote. The customer tops up credits through Stripe, invoiced with tax. |
| 4 | Plan | The agent writes a plan the customer approves in the thread before anything is built. |
| 5 | Build | One disposable sandbox per run, secrets as env vars, a live console with token counters and a Stop button. |
| 6 | Publish | Leak scan, then a pull request under your review. Security review and auto-merge if you allow it. |
| 7 | Demo | Boot-tested, then live on <project>.<your-domain> behind basic auth. Merge the PR and the demo redeploys within a minute. |
| 8 | Deliver | Follow-up requests get their own run, PR and thread. The customer approves delivery. |
What you get
| 🧑💼 Customer portal | Deposit, live status, one thread per project and per request, PR links, usage, delivery approval. |
| 📺 Live build console | Phases, commands, edits, browsing, git, security review, token and credit counters, Stop. |
| 🔌 MCP for their agent | A project token lets Claude Code, Cursor or any MCP client consult the codebase, delegate work and search your knowledge. |
| ⏱️ Programs and routines | Runnable repos on a cron schedule or an inbound webhook; scheduled prompts that become ordinary requests. |
| 🎨 Custom branding | Your brand, consultant name and focus from .env; your landing copy; 20 prompt templates you edit as files. |
| 📚 Knowledge bases | Add knowledge to enrich your harness. Local folder, git repos, MCP servers, Context7, web search. Switched on/off per project. |
| 🛠️ Tools | MCP servers the agent acts through (GitHub, GitLab, web research, yours), with a tool-poisoning scan on every enable. |
| 🧠 Any model | OpenAI-compatible endpoints, chosen per project, with a one-token probe for chat, reasoning effort and vision. |
| 💳 Billing | Stripe integration with prepaid credits at your markup, per-model prices, invoices with automatic tax. |
| 🔒 Sandbox and leak scan | One container per build (Sysbox in production), an egress allowlist, and a scan of staged files, commit messages and added lines before any push. |
| 🔐 Encrypted secrets | Envelope encryption for every secret and Memory value. Secrets reach the sandbox through a sourced-then-deleted file, never docker -e. |
| 🛡️ Guardrails | A forbidden-actions floor the model cannot lower, a retrieval score floor against corpus extraction, proof-of-work captcha, rate limits, an audit log. |
| ☸️ Compose or Kubernetes | One VM with docker compose behind Traefik, or the Helm chart built from the same images. |
Production
Recommended deployment: with Kubernetes so you have truly-isolated pods to run each agentic task.
One VM: a real .env, apex and wildcard DNS records, a wildcard TLS certificate in traefik/certs/, Sysbox for hardened sandboxes, make prod. Or Kubernetes with the Helm chart in k8s/ (make helm-bpdr). The same docs/DEPLOY_WITH_AI.md prompt handles production too.
Bring your own model
Three OpenAI-compatible endpoints (chat, embeddings, reranking): OpenAI, Anthropic, Mistral, OpenRouter, a European gateway or your own vLLM server, per instance or per project. A model needs a row in backend/app/static_data/per_model_price_table.json (or prices on its saved endpoint), otherwise the platform refuses to run it rather than bill zero.
Make it yours
| What | Where |
|---|---|
| Brand, consultant name, consulting focus, colours | the # Brand block in .env |
| Landing copy | landing/src/data/site.example.yml → site.yml |
| Offer catalogue, onboarding questions, Memory suggestions | backend/app/static_data/*.example.json |
| Knowledge | the ./knowledge folder, git repos and MCP servers from the admin Knowledge-bases page |
| Agent behaviour | the versioned prompts in backend/app/agents/prompts/ (rules) |
Docs
| Goal | Start here |
|---|---|
| Develop on it | CLAUDE.md - conventions, workflows, gotchas |
| Understand a subsystem | docs/CODE_MAP.md - the tour, and the dev pipeline end to end |
| Call the API | docs/API_CONTRACT.md |
| Bill customers | docs/stripe-billing.md |
| Run builds in parallel | docs/PARALLEL_BUILDS.md |
| Test | make test (1,000+ backend tests) and the end-to-end scenario in .claude/commands/e2e-check.md |
Stack: FastAPI, SQLAlchemy, Celery, Postgres + pgvector, Redis, Meilisearch, Traefik, OpenHands SDK, React + Vite, Astro, Helm.
Status
Alpha. Openvisor already runs a real consulting practice in production (berwick.ai); expect breaking changes between releases, always with migrations.
Contributing
Issues and pull requests are welcome. Read CLAUDE.md first: one branch and one pull request per change, make test green, docs updated in the same PR.
License
Apache-2.0. "Openvisor" and "Scalevisor" are trademarks of the project maintainers; the license does not grant rights to use them. If Openvisor earns you money, consider sponsoring its development.
Contributors
121 commits
See what people are saying
Similar repos
Languages
Python
74.7%
TypeScript
18.0%
Astro
3.0%
CSS
2.5%
Shell
1.3%
Your AI-native consulting agency running 24/7 for you.
26
stars
121
commits
Python
primary language
Sep 10, 2026
updated
README
Your AI-native consulting agency, running 24/7 for you.
Works with GitHub, GitLab, any OpenAI-compatible model and any MCP client.
See it live: berwick.ai is a real consulting practice running on Openvisor, white-labelled and in production.
How it works · What you get · Quick start · Production · Bring your own model · Make it yours · Docs
Watch the full demo (62 seconds, MP4)
Quick start
With your AI assistant. Copy/paste:
Read instructions at https://github.com/scalevisor-io/openvisor/blob/main/docs/DEPLOY_WITH_AI.md to deploy my own Openvisor instance.
Or by hand.
git clone https://github.com/scalevisor-io/openvisor.git && cd openvisor
cp .env.example .env # update envs!
echo "127.0.0.1 openvisor.local app.openvisor.local mail.openvisor.local mcp.openvisor.local" | sudo tee -a /etc/hosts
make dev
| Landing | http://openvisor.local |
| App | http://app.openvisor.local (sign in with ADMIN_EMAIL / ADMIN_PASSWORD) |
| http://mail.openvisor.local (Mailpit) |
Two switches worth knowing: the sign-in captcha only runs in a secure context, so on plain http set ALTCHA_ENABLED=0; and OPENHANDS_ENABLED=0 (the default) pushes a deterministic scaffold through the real PR → merge → demo path at zero token cost, 1 builds with the agent.
How it works
Coding agents write code. Openvisor is the consultancy around them: the customer portal, the review gate, the sandbox, the demo and the billing. You stay the consultant: the agent can never price, approve or advance a project.
| Step | What happens | |
|---|---|---|
| 1 | Deposit | A customer describes the project on your white-label site, answers your onboarding questions and connects a GitHub or GitLab repo. |
| 2 | Evaluate | Moderation, feasibility and a credit estimate run automatically. Doubts are flagged to you, never auto-accepted. |
| 3 | Price | You approve or answer with a direct quote. The customer tops up credits through Stripe, invoiced with tax. |
| 4 | Plan | The agent writes a plan the customer approves in the thread before anything is built. |
| 5 | Build | One disposable sandbox per run, secrets as env vars, a live console with token counters and a Stop button. |
| 6 | Publish | Leak scan, then a pull request under your review. Security review and auto-merge if you allow it. |
| 7 | Demo | Boot-tested, then live on <project>.<your-domain> behind basic auth. Merge the PR and the demo redeploys within a minute. |
| 8 | Deliver | Follow-up requests get their own run, PR and thread. The customer approves delivery. |
What you get
| 🧑💼 Customer portal | Deposit, live status, one thread per project and per request, PR links, usage, delivery approval. |
| 📺 Live build console | Phases, commands, edits, browsing, git, security review, token and credit counters, Stop. |
| 🔌 MCP for their agent | A project token lets Claude Code, Cursor or any MCP client consult the codebase, delegate work and search your knowledge. |
| ⏱️ Programs and routines | Runnable repos on a cron schedule or an inbound webhook; scheduled prompts that become ordinary requests. |
| 🎨 Custom branding | Your brand, consultant name and focus from .env; your landing copy; 20 prompt templates you edit as files. |
| 📚 Knowledge bases | Add knowledge to enrich your harness. Local folder, git repos, MCP servers, Context7, web search. Switched on/off per project. |
| 🛠️ Tools | MCP servers the agent acts through (GitHub, GitLab, web research, yours), with a tool-poisoning scan on every enable. |
| 🧠 Any model | OpenAI-compatible endpoints, chosen per project, with a one-token probe for chat, reasoning effort and vision. |
| 💳 Billing | Stripe integration with prepaid credits at your markup, per-model prices, invoices with automatic tax. |
| 🔒 Sandbox and leak scan | One container per build (Sysbox in production), an egress allowlist, and a scan of staged files, commit messages and added lines before any push. |
| 🔐 Encrypted secrets | Envelope encryption for every secret and Memory value. Secrets reach the sandbox through a sourced-then-deleted file, never docker -e. |
| 🛡️ Guardrails | A forbidden-actions floor the model cannot lower, a retrieval score floor against corpus extraction, proof-of-work captcha, rate limits, an audit log. |
| ☸️ Compose or Kubernetes | One VM with docker compose behind Traefik, or the Helm chart built from the same images. |
Production
Recommended deployment: with Kubernetes so you have truly-isolated pods to run each agentic task.
One VM: a real .env, apex and wildcard DNS records, a wildcard TLS certificate in traefik/certs/, Sysbox for hardened sandboxes, make prod. Or Kubernetes with the Helm chart in k8s/ (make helm-bpdr). The same docs/DEPLOY_WITH_AI.md prompt handles production too.
Bring your own model
Three OpenAI-compatible endpoints (chat, embeddings, reranking): OpenAI, Anthropic, Mistral, OpenRouter, a European gateway or your own vLLM server, per instance or per project. A model needs a row in backend/app/static_data/per_model_price_table.json (or prices on its saved endpoint), otherwise the platform refuses to run it rather than bill zero.
Make it yours
| What | Where |
|---|---|
| Brand, consultant name, consulting focus, colours | the # Brand block in .env |
| Landing copy | landing/src/data/site.example.yml → site.yml |
| Offer catalogue, onboarding questions, Memory suggestions | backend/app/static_data/*.example.json |
| Knowledge | the ./knowledge folder, git repos and MCP servers from the admin Knowledge-bases page |
| Agent behaviour | the versioned prompts in backend/app/agents/prompts/ (rules) |
Docs
| Goal | Start here |
|---|---|
| Develop on it | CLAUDE.md - conventions, workflows, gotchas |
| Understand a subsystem | docs/CODE_MAP.md - the tour, and the dev pipeline end to end |
| Call the API | docs/API_CONTRACT.md |
| Bill customers | docs/stripe-billing.md |
| Run builds in parallel | docs/PARALLEL_BUILDS.md |
| Test | make test (1,000+ backend tests) and the end-to-end scenario in .claude/commands/e2e-check.md |
Stack: FastAPI, SQLAlchemy, Celery, Postgres + pgvector, Redis, Meilisearch, Traefik, OpenHands SDK, React + Vite, Astro, Helm.
Status
Alpha. Openvisor already runs a real consulting practice in production (berwick.ai); expect breaking changes between releases, always with migrations.
Contributing
Issues and pull requests are welcome. Read CLAUDE.md first: one branch and one pull request per change, make test green, docs updated in the same PR.
License
Apache-2.0. "Openvisor" and "Scalevisor" are trademarks of the project maintainers; the license does not grant rights to use them. If Openvisor earns you money, consider sponsoring its development.
See what people are saying
Similar repos
Contributors
121 commits
Languages
Python
74.7%
TypeScript
18.0%
Astro
3.0%
CSS
2.5%
Shell
1.3%