Your codebase
Source code, README, configs, website, database
How it works See how Zipchat reads your code and answers technical questions.
Use Cases
See all use casesIndustries
View all industriesCapabilities
See all capabilitiesUse Cases
See all use casesIndustries
View all industriesCapabilities
See all capabilitiesConnect your GitHub, GitLab, or Bitbucket repo. Zipchat reads your codebase, your marketing website, and any connected sources, then answers technical support questions on website chat and email. No documentation sprint, no stale docs, no engineering interruptions.
7-day free trial · 30-day money-back guarantee
Your codebase
Source code, README, configs, website, database
Zipchat Code AI
Codebase-aware retrieval, always current
Customer touchpoints
Website chat, email support, human handoff
There is no manual training. No article writing. No cold-start period where the AI guesses at your product. Zipchat reads your actual codebase and starts answering technical questions from day one. Here is the exact sequence.
Connected
Step 1
Sign up at zipchat.ai and select Codebase as your agent type. Enter your marketing website URL so the AI understands what your company does. Then provide your Git repository URL and an access token.
Zipchat verifies access, clones your repository, and begins indexing. Initial indexing completes within minutes for most codebases. You can extend the knowledge base later with additional sources via Custom Tools.
github.com/acme 
Source code 
README & comments 
Configuration files 
Marketing website 
Docs, wikis & tickets Step 2
Once access is confirmed, Zipchat clones your repository into an isolated, secure environment. It reads and indexes your source code, README files, configuration files, and code comments. It crawls your marketing website for product context. PDFs, custom text, an optional PostgreSQL connection, and any tool you connect via Custom Tools cover anything else.
You do not write articles. You do not train the AI on example Q&A pairs. The knowledge base builds itself.
Re-indexing runs weekly on Starter and Growth, daily on Pro, Scale, and Enterprise. Ship a new release and the AI learns the change on the next sync. Your docs can go stale. Your code does not.
Correction added Step 3
Spend five minutes in the AI Training panel. Write a core prompt: product name, tone, what to do when context is missing, and topics that escalate immediately (outage reports, billing disputes, security incidents). Add chat starters and proactive campaigns triggered on your pricing page, docs portal, or trial sign-up flow.
For actions, add Custom Tools. Connect Jira or Linear to create tickets automatically. Connect Zendesk, Intercom, Gorgias, Freshdesk, or Salesforce to route escalations into your existing support queue. Connect any internal API in plain English. No coding. No webhooks to maintain.
Test Chat mode previews answers before any customer sees them. If a response is wrong, correct it in one click and the AI never repeats the mistake. Debug mode shows you the exact part of the codebase or knowledge base used to build each answer.
Website chat
Email supportZendesk handoffJira ticket creation
Internal team accessStep 4
One knowledge base, one AI, deployed on website chat and email. The same agent answers a prospect on your pricing page, a customer in your support inbox, and an internal teammate on Slack with the same accuracy and the same source of truth.
The analytics dashboard tracks every question the AI answered, every question it could not, and every escalation. You see which product areas generate the most support load, where to invest in product fixes, and where the docs need to be filled in.
Zipchat Code does not use generic AI training data to answer your customers. Every answer comes from your actual codebase, your marketing website, and any additional sources you connect. Ship a new feature on Monday and the AI knows about it by Tuesday. No documentation sprint required.
Connect GitHub, GitLab, or Bitbucket with an access token. Zipchat clones your repo and indexes source code, READMEs, configuration files, and code comments. Private repos fully supported. Self-hosted GitLab supported.
Your company website is crawled automatically at setup. The AI uses it for product positioning, feature descriptions, pricing context, and anything a prospect would ask before evaluating your API or product.
Connect Notion, Confluence, Jira, Linear, Zendesk, Intercom, Google Drive, ReadMe, or any other tool via Custom Tools. The AI calls the API during conversations and incorporates the response into its answer.
Connect a read-only PostgreSQL database for live data access. The AI runs scoped queries during conversations to answer about subscription status, usage data, or customer records, never exposing raw SQL or schema.
Upload PDFs, paste text, or add individual URLs. Architecture decision records, internal onboarding docs, runbooks, security FAQs, anything not accessible via API goes in here.
Your core prompt tells the AI how to speak, what to escalate, and what to do when context is missing. Every answer comes back in your voice, not as a generic chatbot reply.
Zipchat Code combines RAG (Retrieval Augmented Generation) with direct codebase exploration in an isolated E2B sandbox. The AI answers as a product expert and never exposes source code, file paths, or internal implementation details to end users. When the answer is not in your knowledge base, it says so and escalates rather than inventing one.
Each time a customer sends a message, Zipchat runs a six-stage retrieval loop. The loop pulls the most relevant code, README, or doc, verifies it against a confidence threshold, then either delivers a grounded answer or escalates to a human, all in real time, with no manual routing required.
The raw message is classified by topic, language, and urgency. Ambiguous phrasing (for example "the webhook is failing") is resolved against conversation history so follow-up questions stay in context.
Semantic vector search and keyword BM25 search run in parallel across your indexed codebase, README files, marketing website, and connected docs. The top candidates from both are merged and re-ranked by relevance score.
When retrieval alone is not enough, the AI uses shell-level access in an isolated E2B sandbox to navigate directories, read related files, and cross-reference behavior across multiple parts of the codebase before forming an answer.
The language model drafts an answer constrained strictly to the assembled context. It formats the reply in the customer's language and applies your brand tone guidelines. Source code, file paths, and credentials never appear in the draft.
A confidence score is evaluated against your configured threshold. If the answer meets the threshold, it proceeds to delivery. If not, it forks to the escalation path.
Confident answers are sent to the customer with a citation reference, in their language and your brand tone. Doc links and product references are included where relevant. No source code, file paths, or internal identifiers are ever exposed.
Uncertain answers trigger a handoff to a human agent via Zendesk, Intercom, Gorgias, Freshdesk, or Salesforce, with the full conversation context pre-loaded so the agent never has to ask the customer to repeat themselves.
You configure the AI once. It then runs natively on website chat and email, escalates into your existing support stack, and exposes a REST API for in-app or in-product surfaces. Same knowledge base, same brand voice, same accuracy threshold.
Answers technical pre-sales and support questions in real time. Deploys to your marketing site, docs portal, or in-app help widget. One line of JavaScript embeds it.
The AI reads incoming support emails and replies using your codebase. Supports multi-turn threads. Drafts can be reviewed in Zendesk, Intercom, or Freshdesk before sending.
Escalates into Zendesk, Intercom, Gorgias, Freshdesk, or Salesforce with the full conversation transcript and confidence score, so the agent has everything needed to resolve the issue.
Custom integrations: in-app chat, voice transcripts, internal copilots, or your own native app. JSON in, JSON out. Same retrieval pipeline behind every channel.
Most AI support tools search your documentation. Documentation goes stale. Your codebase does not. Zipchat Code reads the actual source of truth, the code your product runs on, and answers from that. This is the difference between 60-70% accuracy on docs and 96% accuracy on code.
RAG grounding on live code
Every answer is composed only from retrieved chunks that exist in your codebase, README files, or connected sources. The model is instructed not to generate facts that are absent from the retrieved context. No internet browsing. No generic training data.
Confidence threshold
Before delivery, each draft is scored. If retrieval confidence falls below your configured minimum, the AI does not guess. It acknowledges uncertainty and escalates to a human agent. You set the threshold per use case.
No source code exposure
The AI answers as a product expert. It never shares source code, file paths, credentials, or internal implementation details with end users. Customers see product-level explanations. Engineers see nothing they did not already know.
Why this is different from ReadMe, Mendable, Inkeep, and DocsBot
Those tools search your documentation. When a customer asks a question not in your docs, they get a non-answer or a hallucination. When your docs are out of date, and 46% of documentation goes stale within three months, they give the wrong answer confidently. Zipchat Code reads your source code instead. It answers from what your product actually does today, not from what someone documented six months ago. The result: 96% answer accuracy versus the 60-70% accuracy ceiling most doc-based tools hit.
These are outcomes from SaaS companies running Zipchat Code today. Ticket deflection is measured as the share of technical support questions the AI resolves without human escalation. Engineering focus time is measured by the reduction in interruptions via Slack, email, and direct escalation.
Doc-based AI tools cap at 60-70% answer accuracy because half the time the answer is not in the docs. Zipchat Code reads your codebase, so when the docs are missing or wrong, the answer is still correct.
See how Zipchat Code handles your product's specific support questions.
Book a demoZipchat Code connects natively to GitHub, GitLab, and Bitbucket. It deploys to your existing website and email channels. For ticketing, escalation, and internal data, it connects to the tools your team already uses, via plain-English Custom Tools.
GitHub
Personal access token or GitHub App. Supports private repos and organizations.
GitLab
Personal access token. Supports GitLab.com and self-hosted GitLab instances.
Bitbucket
App password and username. Supports private repositories.
Zipchat Code connects to Zendesk, Intercom, Gorgias, Freshdesk, and Salesforce for human escalation. It connects to Jira, Linear, Notion, Confluence, Google Drive, ReadMe, and any other HTTP API via Custom Tools. Write plain-English instructions, add your API key, and the AI calls the tool during conversations.
See all integrationsYour code and your customer conversations are sensitive. Zipchat Code is architected with security as a first principle, not an upgrade tier.
Your codebase is cloned into an isolated, ephemeral E2B sandbox. It is not shared with other customers, not used to train any AI model, and credentials are encrypted at rest.
All stored knowledge base data and conversation logs are encrypted with AES-256. All API traffic uses TLS 1.2+. Access tokens for Git providers are encrypted at rest.
Data processing agreements available. Conversation data can be auto-deleted after 30, 60, or 90 days. EU data residency available on Enterprise plans.
Integrate with your identity provider via SAML or OAuth. Assign view-only, editor, or admin roles to team members without sharing credentials.
You provide an access token with read-only scope. Zipchat connects to private repositories on GitHub, GitLab, or Bitbucket using that token. Your code is cloned into an isolated E2B sandbox that is not shared with any other customer. It is never used to train any AI model and is never accessible to Zipchat employees. The AI answers from the code but never exposes file paths, code snippets, or internal implementation details to any end user.
On Starter and Growth plans, the codebase re-indexes weekly. On Pro, Scale, and Enterprise plans, it re-indexes daily. You can also trigger a manual rescan at any time from the Knowledge Base dashboard. Zipchat connects to your main branch by default. If you ship on Friday, the AI knows about it by Saturday morning on daily plans. For teams that ship continuously, daily re-indexing means the knowledge base is never more than 24 hours out of date.
The AI says it does not have that information, collects the customer's email, and forwards the full conversation context to your support team via your configured escalation channel (Zipchat inbox, Zendesk, Intercom, Gorgias, Freshdesk, or Salesforce). The unanswered question is also flagged in Proactive Knowledge Gap Detection for you to address. Your support agent receives the full thread without having to ask the customer to repeat themselves.
By default, it answers questions. With Custom Tools, the AI can take actions during conversations: create tickets in Jira or Linear, look up subscription status in your CRM, check usage data in your database, or trigger internal workflows via any HTTP API. Each tool is defined in plain English. You specify the API endpoint, when the AI should call it, and what to do with the response. The AI executes the call in a secure sandbox and shares the result with the customer. No coding required.
Yes. For GitLab.com, connect with a personal access token. For self-hosted GitLab, provide your instance URL along with a personal access token. Zipchat connects to self-hosted GitLab the same way it connects to GitLab.com. There is no requirement to use a hosted Git provider. This is the primary reason enterprise teams on self-hosted GitLab choose Zipchat Code over competitors that only support github.com.
Your codebase is cloned into an isolated, ephemeral E2B sandbox. It is not shared with other customers, not used to train any AI model, and credentials are encrypted at rest. The AI answers as a product expert and never reveals source code, file paths, variable names, or internal implementation details to any end user. Zipchat does not currently hold SOC 2 certification. If compliance certification is a hard requirement for your procurement process, confirm the current status directly with the Zipchat team at the demo stage.
Customer conversation data is handled in accordance with GDPR requirements. Personal data collected during conversations is stored securely and not shared with third parties. Your codebase is processed only for the purpose of answering product questions and is not used for any other purpose. For specific data residency options or DPA requests, discuss with the Zipchat team during the sales process.
That is the entire point. Support agents, CS managers, sales reps, and product managers can all ask questions in plain English and get accurate, sourced answers, with no engineering background required. The AI answers as a product expert. It never sends technical jargon, code snippets, or file paths to the person asking. New team members ramp up on the product faster because they have an always-available expert to ask.
