Phone voice bot
The customer calls, the bot answers — anchored in the code, not in a FAQ. 24/7. Backup capacity when human support is saturated.
AI Voice market: $5B+ in 2026, growing ~40% per year.
Private beta · limited access
3 days to grasp my own code. Never again.
SourceTrek turns your codebase — Python, Java, PHP, WinDev — into queryable memory. Your dev teams, your support, and soon your customers ask questions in plain English. Sourced answers from the source code, in seconds.
60x measured on a production pilot · French hospitality SaaS vendor (1M+ lines, under NDA)
The code is the only truth.
Three audiences are cut off from it.
Code says how. Never what or why. Documentation is outdated the moment it ships — even its author forgets it a few months later. The result: at every level, access to the product's reality depends on a handful of people.
The developer
Established market≈ 60%¹
of dev time spent reading existing code
Understanding a module, assessing impact before coding, refactoring safely, onboarding. Existing tools help — but they stop at the developer.
The support engineer
UnderservedL1 → Dev
systematic escalation on business questions
Support has no access to the code. Documentation lags behind. Every business ticket bubbles up to a developer. A developer handling L1 is a developer who isn't shipping.
The end customer
Untapped segmentStatic FAQ
customer chatbots feed on static knowledge bases
The leading customer chatbots pull answers from FAQs and articles. No one anchors the bot directly to the product's source code. The window is open.
¹ Reading vs. writing code: the order of magnitude "reading >> writing" has been documented since Knuth (1984) and confirmed in several recent studies (Minelli & Mocci 2015, GitHub Octoverse 2022). The exact ratio varies 50–70% depending on the codebase and the study.
The solution
A queryable memory
A queryable memory
of your code.
SourceTrek reads your source code, builds a semantic map you can query, and answers questions in natural language — every answer points to the file and line involved. No FAQ to maintain. The memory follows the code.
-
Index your repo
Multi-language by design. Self-hostable or private cloud. Automatic re-indexing on push.
-
Ask in natural language
From your IDE, a web chat, or embedded in your support tool. The same answer adapts to the profile of the person asking.
-
Sourced answer
Synthesis, then direct citations to files, classes, lines. Verifiable, traceable. No answer without a source.
-
Learns your jargon
Your company's domain terms are picked up through usage. The bot speaks your language, not generic jargon.
chat.source-trek.com
How are employer social-security contributions calculated in the payroll module?
Contributions are computed in 3 steps: identify the applicable salary brackets (social-security ceiling, tranches A/B/C), apply each contribution rate per the collective bargaining agreement, then adjust for current exemptions (Fillon relief, urban free zones, apprentices). The total is consolidated by category before payslip generation.
reservations/supplements.py:42
pricing/seasonal_rules.java:88
loyalty/adjustments.php:115
Demo · polyglot monorepo typical of a SaaS vendor (PHP front, Java calc, Python business rules)
Next question
Three audiences, one indexing
The dev use case is available today in beta. The support and end-customer use cases run on the same technical foundation and arrive in phases.
Available
Dev productivity
For engineering teams
Instant understanding, impact analysis before coding, safe refactoring, accelerated onboarding. Access via your IDE or web chat.
Roadmap T+3 to 6 months
Support excellence
For L1/L2 support engineers
Dedicated persona, non-technical language, self-learning business glossary. Multimodal screenshots in the conversation. Support finds instead of escalates.
Vision T+12 months
Customer self-service
For your end users
Chatbot or voice bot embedded in your product, anchored in your source code — not in a drifting FAQ. A premium differentiator for SaaS vendors.
60x on a SaaS vendor
in production.
The MVP is running in internal testing at a French SaaS vendor in the payroll & HR-compliance sector, on a production product of over one million lines of multi-language code. The case below is real; the partner's identity is under NDA, verifiable in due diligence.
"How are employer social-security contributions calculated in the payroll module?"
Real business question asked to SourceTrek on a repo of more than one million lines. Synthetic answer with direct citations to the source code — understandable by a developer and by a functional analyst.
Without SourceTrek
2 h
code study, senior dev
With SourceTrek
2 min
full grasp of the behavior
×60
speedup on this case
"Business question solved in 2 minutes, versus the usual 2 hours. The first tool that speaks to my support team just as well as to my developers."
— CTO, French payroll SaaS vendor · 1M+ lines · testimony verifiable under NDA
Case measured on a representative business question. Expect variability depending on module complexity and codebase.
Technical coverage
Polyglot by design.
Polyglot by design.
WinDev as a bonus.
The typical company is polyglot. So is SourceTrek. Native WLanguage support also unlocks a French market barely addressed by American tools.
Python
native AST parsing
Java
classes, packages, annotations
PHP
composer-aware, namespaces
WinDev / WLanguage
dedicated parser, tables & columns
Architecture built to extend: adding a language means plugging in a connector, not rewriting the system. Open roadmap driven by early customers: .NET, Go, Rust, COBOL.
Under the hood
No heavy framework.
No heavy framework.
Just clean code.
Thirty years of craft in every line. No black box stacked on others, no dependency we can't control. Here's the journey of a question — from your source code to the answer.
Read
Every file is read and understood in depth — structure, classes, dependencies.
Map
All the code lands on a living map. Who calls whom, who depends on what.
Meaning
Beyond syntax, intent. A price computation is a price computation, whatever it's called.
Investigate
For every question, an agent reasons step by step. Searches, cross-checks, verifies its sources.
Answer
Clear synthesis, verifiable citations. The same memory for your three audiences.
Independent core
The engine depends on no external tech. Swap the AI model, swap the storage backend, plug in a custom parser — nothing breaks. A strong enterprise argument: your investment is yours.
Open connector
The tools your teams already use — IDE, chat, support, AI agents — plug into SourceTrek through an open standard. Immediate adoption, no migration. Tomorrow: a marketplace of third-party agents.
Beyond the MVP
Three circles — then
Short term · T+12 to 18 months
Mid term · T+18 to 24 months
Extension · T+24 to 36 months
Vision · T+30 to 48 months
Three circles — then
a universal substrate.
The MVP serves your developers. The next steps open up support, then your customers. Beyond that, SourceTrek becomes the foundation on which the next generation of software talks to the humans who use it.
Multimodal screenshot capture
The user sends the screen that's causing the issue. The bot identifies the component in the code, explains it, suggests a fix. Decisive for support and product onboarding.
Unblocks customer onboarding + direct L1 deflection.
Marketplace of AI agents
Third-party agents — GDPR counsel, CVE security, FinOps, accessibility — plug into the SourceTrek memory to analyze your code on your behalf. A multi-sided platform.
AI agents are the next users: they plug in here.
Legacy code modernization
The semantic map serves as an executable specification. COBOL → Java, WinDev → .NET, legacy PHP → modern. Target: vendors consolidating, IT services firms, M&A audits.
Legacy code modernization market: $25B in 2026. 10x margins vs. classic development.
MVP = developers. Circles 2-3 = support and end customers. Beyond = the universal substrate every AI agent, every legacy modernization, every enterprise software product plugs into to talk to its humans.
Why SourceTrek
No one else covers
No one else covers
all three audiences with a single indexing.
Today's tools have split into two watertight camps: those that speak to the developer about the code, and those that speak to the customer through FAQs. None bridges the two. SourceTrek is the bridge — one source of truth, three audiences, three revenue lines.
🛟 WinDev focus — saving the legacy code no one wants to read anymore
WinDev (PC SOFT's RAD environment, widely used in French enterprise) is fading. WLanguage developers are retiring out of the workforce, the next generation isn't being trained, and large AI models have barely seen this language during training. The result: critical business applications run in production with no one to maintain them confidently.
SourceTrek natively understands WLanguage: call graph, dependencies between windows, classes, HFSQL queries, stored procedures. A developer discovering a 500,000-line WinDev project can ask a question in plain English and get the exact map — without guessing, without re-reading by hand, without hunting down the old-timer who "knew".
For vendors and IT departments still hosting WinDev: SourceTrek turns a risky legacy asset into a documented, transferable, and progressively migratable base.
The questions we hear
most often.
Does my source code leave my infrastructure?
Not by default. SourceTrek can be self-hosted in a private cloud or on-premises. Source code does not leave your perimeter. A SaaS mode is also available (encryption at rest and in transit, tenant isolation).
Which AI model do you use?
Model-agnostic architecture. You can plug in your own provider key (OpenAI, Anthropic, Google Vertex, Mistral) or a local model. No single-vendor lock-in.
Does it replace my team?
No. SourceTrek offloads repetitive code reading and systematic L1 escalation to dev. Your developers keep the high-value work — architecture, refactoring, new features.
How long does the initial indexing take?
From a few minutes to a few hours depending on repo size. Subsequent re-indexing is incremental (modified files only). On the partner pilot (1M+ lines, multi-language), full indexing completes in under an hour on standard hardware.
Which languages do you support today?
Python, Java, PHP, WinDev/WLanguage in the MVP. .NET, Go, Rust, COBOL on the roadmap based on early customer demand. Extensible architecture — adding a new language means plugging in a connector without touching the core.
What's the pricing?
B2B SaaS model per user, or self-hosted per instance. Pricing finalized for the beta exit. Early beta partners get preferential terms for the first 12 months.
Join the private beta
First spots reserved for SaaS vendors and engineering teams who want to save hours of understanding every week. Response within 48 hours.