Potpie
Potpie maps large codebases into a knowledge graph, then runs specialised AI agents on it to plan implementations, generate tested code, debug production incidents and open pull requests. Open source, self-hostable, built for enterprise engineering teams.
What is Potpie?
Potpie is an AI-native software delivery platform published by Momenta Softwares Inc., aimed at engineering organisations whose codebases have outgrown generic coding assistants. Its central mechanism is stated plainly on the site: your code is mapped into a knowledge graph that agents use to reason and execute tasks. A proprietary Context Engine sits on that graph and supports multi-hop reasoning across components, which the vendor argues is what surface-level assistants cannot do beyond a million lines. The platform is organised into four building blocks. Specialists are atomic, task-scoped agents; the published directory includes a Q&A Agent with semantic search and citation tracking, a Debugging Agent covering log analysis, dependency graphs, diff impact and stack traces, an LLD Agent and a Validation Agent handling schema validation, security scanning and type checks. Forge treats code generation as a constraint-satisfaction problem, combining test-driven development with the knowledge graph and running each candidate change in ephemeral zero-trust containers, with CPU, memory and I/O profiling, before any pull request appears. Recipes are durable event-driven workflows started by CI/CD, alerts, cron schedules, webhooks or a git push, with checkpoint-based recovery for jobs that run for hours. Trace investigates production incidents by traversing the service call graph, comparing live behaviour against API and deployment baselines, and exploring branching hypotheses in sandboxes. The model layer is deliberately provider-agnostic: OpenAI, Anthropic, Azure, OpenRouter and MiniMax are named, and the Enterprise tier adds self-hosted language models. Potpie is reachable as a web application, an API, a command-line tool, a Slack app and a VS Code extension, and the core is open source on GitHub with documented self-hosting. Published claims include 63% on SWE-bench Lite, the only figure linked to verifiable evidence through a pull request in the SWE-bench repository, alongside 5,100 or more GitHub stars, testing at fifty million lines of code and a 41% improvement in pull request cycle time. Testimonials are attributed to a CEO at Flobiz, a senior engineer at Astronomer and a VP of engineering at Moniepoint, without named individuals.
What it does
- Index a repository branch and build a queryable knowledge graph of the whole codebase
- Answer questions about unfamiliar code with citations back to the exact source files
- Turn a plain-language request into a structured implementation plan, asking clarifying questions first
- Generate code against the plan, test it in an isolated container and open a pull request
- Diagnose production incidents by traversing the service call graph and validating contracts against deployment baselines
- Run durable multi-agent workflows triggered by CI/CD, alerts, cron schedules or a git push
- Build custom agents for recurring work such as migrations, refactors and architecture reviews
When to use Potpie / When not to
A quick filter to help you decide if Potpie is the right fit.
When to use Potpie
- Engineering teams working in codebases of one million to fifty million lines, where keyword search and generic assistants stop being useful
- Organisations maintaining poorly documented monoliths or legacy systems that need the architecture explained before it can be changed
- Site reliability and platform engineers running root cause analysis on production incidents: one cited customer went from nearly a week to around thirty minutes on a 40-million-line codebase
- Regulated businesses that require audit trails, on-premises deployment and self-hosted language models rather than a shared cloud service
- Teams that want file-accurate implementation plans and clarifying questions before any code is written, instead of a first draft built on assumptions
When not to use Potpie
- Anyone needing a mobile app: Potpie ships as a web application, an API, a CLI, a Slack bot and a VS Code extension, with no iOS or Android release
- Buyers who need a published price before talking to sales, since the paid Enterprise tier is quoted as Custom and the pricing page is only a qualification form
- Free-tier users who cannot accept their inputs feeding internal model fine-tuning, because the privacy policy states no opt-out is offered on those plans
- Teams with a hard EU data residency requirement, as the majority of systems and storage sit in India and no EU region is offered outside on-premises deployment
- Solo developers or small teams after a lightweight in-editor autocomplete, as this is an orchestration layer designed around organisational context
How to use Potpie
A typical end-to-end flow, from setup to results.
- Sign up at app.potpie.ai, authenticating through GitHub single sign-on
- Connect your existing engineering systems: source control first, then project management and documentation tools, so the agents inherit full system context
- Index a branch of the repository and let Potpie build the context layer; the documentation advertises a five-minute quickstart
- Describe in plain language what you want built, without writing a specification up front
- Answer the clarifying questions the agent generates on authentication, role-based access control, data sources and security decisions
- Review and refine the generated implementation plan, adding any extra context or requirements
- Let the agent produce code aligned to your stack, tested in an isolated container with your own test command
- Approve the pull request it opens against the base branch you selected
- For recurring work, wire a Recipe to a trigger such as a CI/CD event, an alert, a schedule or a git push
- Alternatively drive Potpie from the CLI, the VS Code extension, the Slack bot or the API, or self-host it using the documented setup
Pros & Cons
Pros
- Repository context is built on an explicit knowledge graph rather than keyword retrieval, which is what makes multi-hop reasoning across services possible
- Output is designed to be checkable: plans cite exact files and carry per-task verification criteria, and generated code is executed in a sandbox before it reaches a pull request
- The agent asks clarifying questions before writing, which removes the silent assumptions that make generated plans expensive to unwind
- Open source with more than 5,100 GitHub stars, documented self-hosting, and on-premises deployment with self-hosted models for sovereignty constraints
- Provider-agnostic orchestration lets you swap OpenAI, Anthropic, Azure, OpenRouter or MiniMax without rewriting workflows
- Execution traces and audit trails give the paper record that regulated environments require
- Unusually clear contractual position on ownership: your code and the generated outputs stay yours, and private code is never used to train public models
Cons
- No public price for the paid tier: Enterprise is quoted as Custom, and the pricing page carries only a qualification form while the actual tier table sits on the homepage
- Free-tier inputs feed internal model fine-tuning and the privacy policy states plainly that no opt-out mechanism is offered on those plans
- The Pro Plan training exclusion is described as being rolled out rather than already in force, so it is not yet a guarantee
- Inputs, logs and feedback may be stored indefinitely to support history, analytics and model training
- SOC 2, HIPAA and GDPR badges appear on the marketing pages with no supporting text, audit date or accessible report, and the Vanta trust centre renders nothing without JavaScript
- The majority of systems and storage sit in India, with no data residency option outside on-premises deployment, and no DPA is published
- Liability is capped at the greater of twelve months of fees or 100 USD, with mandatory individual arbitration in Delaware and a class action waiver
Pricing & Plans
A permanent free plan is available. The Individual - Pro tier is listed at 0 USD per month and includes 50 requests per month, with unlimited requests if you supply your own model provider keys. The only paid tier, Enterprise, is quoted as Custom: no figure is published anywhere on the site. The pricing page states that licences are priced per user, that a platform fee applies based on the number of users supported, and that integrating your existing technology stack is free. No time-limited free trial is advertised, and the self-hosted open source edition carries no vendor licence fee.
- 50 requests per month and unlimited with your own keys
- custom agents
- agentic workflows
- custom tools
- community and email support
- unlimited requests per month
- custom agents
- agentic workflows
- self-hosted language models
- custom tools
- audit trails
- on-premises deployment and a dedicated forward deployment engineer
- Self-hosted open source edition — no vendor licence fee — the core platform from the public GitHub repository
- deployed on your own infrastructure with your own model keys
- governed by its open source licence rather than the hosted terms of service
Data, GDPR & hosting
A consolidated view of how Potpie handles your data.
GDPR overview
Implementation is partial, and the site says so. The privacy policy states that Momenta Softwares does not currently have a formal presence in the EU or UK, and that its data protection policies are being updated to meet the standards of the GDPR and UK GDPR: an intention rather than a claim of compliance. No Article 27 representative is designated and no data protection officer is named. The rights offered are listed concretely: access and portability in CSV or JSON, rectification, erasure, withdrawal of consent, objection, restriction, marketing opt-out and the right to complain to a supervisory authority. Requests go to hello@potpie.ai, erasure under the subject line Data Deletion Request, with an internal appeal route. One gap deserves flagging: a GDPR badge appears on the marketing pages while the policy text stops short of claiming conformity. Personal data is transferred to India and subject to Indian law.
Who owns the data?
The terms are explicit: you retain all rights and ownership in your User Content, meaning the code, data, files and prompts you supply, and also in the outputs generated for you. Potpie states it will not claim ownership of your code or other proprietary material. In exchange you grant Momenta Softwares Inc. a non-exclusive, worldwide, royalty-free licence limited to two purposes: operating the service, and improving it including the underlying models. Private code is not used to train publicly available models nor shared with other customers. The platform itself, its software, architecture, models, branding and documentation, remains the vendor's exclusive property under a revocable licence to you. Feedback you submit passes to the vendor perpetually.
Reuse rights
Because you own the outputs, you may reuse generated code and answers without asking permission, and the terms confirm Potpie disclaims any copyright in AI-generated output. Two caveats are stated. Uniqueness is not guaranteed, so similar code may be produced for other users and you should check licence compatibility before incorporating it; and you remain responsible for reviewing and testing anything generated before it reaches production. The vendor's own reuse of your data is bounded but not symmetrical. De-identified inputs and suggestions feed internal fine-tuning on free-tier plans, where the privacy policy says no opt-out mechanism is offered, while Pro Plan data is excluded by default, a feature the terms describe as still being rolled out. Usage, error and performance data reach PostHog, Sentry and New Relic; payments run through Stripe, authentication through Firebase, transactional email through Resend. Nothing is sold and no cross-context behavioural advertising takes place.
Data retention & training
Hosting summary
Momenta Softwares Inc. is a US company, but the privacy policy states that the majority of its systems and data storage are located in India. Personal data is therefore transferred to and processed in India, subject to Indian law and potentially accessible to Indian authorities, and the policy acknowledges that Indian data protection law may not offer the same level of protection as your own. It adds, in conditional terms, that cloud providers and other vendors might use data centres in the United States, the European Union, Asia or elsewhere; that is not a residency commitment. Protection measures cited are encryption in transit over HTTPS for the website and API, encryption at rest where applicable, and periodic review of collection and storage practices. For on-premises or self-hosted deployments, data does not leave the customer's infrastructure and is not transmitted to the vendor in normal operation, which makes that the only route to keeping data in a chosen jurisdiction. No European hosting region is offered on the cloud service.
Where Potpie works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Potpie.
- Compliance badges without evidence: SOC 2, HIPAA and GDPR logos appear in the security section of the homepage and enterprise page as images only, with no text, audit date, certificate number or report. The Vanta trust centre renders nothing without JavaScript. Do not treat these as attestations
- A direct contradiction on GDPR: the badge claims one thing while the privacy policy states there is no formal EU or UK presence and that policies are being updated to meet GDPR standards. No Article 27 representative is designated and no DPA is published
- Free-tier training with no exit: inputs and suggestions from free plans are used for internal fine-tuning and the policy says no opt-out is offered. The Pro Plan exclusion is described as being rolled out, and a naming inconsistency makes it unclear which regime covers the 0 USD Individual - Pro tier
- Indefinite retention: inputs, logs and feedback may be kept indefinitely for history, analytics and model training, and some data survives deletion requests in backups or for legal and audit reasons
- Jurisdictional exposure: most storage sits in India and is subject to Indian law, while disputes are arbitrated individually in Delaware with a jury and class action waiver and liability capped at 100 USD or twelve months of fees
- Repository access is broad: GitHub OAuth tokens are stored by the vendor to read the code you authorise, so the blast radius of a vendor compromise is your source tree. Scope permissions deliberately and review what each agent may touch
- Skill atrophy and misplaced trust: the platform is explicitly designed to reason about code you have not read, and the terms make you responsible for reviewing generated code. Teams that let plans and pull requests pass unexamined trade short-term velocity for engineers who no longer understand their own system
Setup & Integrations
Technical difficulty
Low to moderate for the hosted service, and squarely aimed at engineers. You sign up with GitHub single sign-on, connect repositories, index a branch, and the documentation advertises a five-minute quickstart. The real effort is integration rather than installation: value depends on also connecting project management and documentation tools, and granting OAuth access to your repositories. Unlimited use of the free plan requires supplying your own model provider keys, which means holding third-party API accounts. Self-hosting is documented but means running your own infrastructure. Enterprise on-premises deployment includes a dedicated forward deployment engineer.
Deployment
Integrations
Behind Potpie
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Potpie.
Frequently asked questions
What does Potpie actually do?
Is there a free plan, and what does it cost to upgrade?
Will my code be used to train Potpie's models?
Who owns the code the agents generate?
Where is my data hosted?
Is Potpie open source, and can I self-host it?
What does it integrate with, and which language models does it support?
Is Potpie certified against SOC 2 or HIPAA?
How do I reach a human, and what is the minimum age?
Should you pick Potpie?
Potpie is one of the more technically legible entries in the agentic coding category. It names its mechanism instead of gesturing at it: the codebase becomes a knowledge graph, a Context Engine reasons across it, and four distinct products sit on top for questions, code generation, durable workflows and incident investigation. The design choices point consistently in one direction, which is verifiability. Plans cite exact files, tasks carry verification criteria, generated changes run in ephemeral sandboxes before a pull request exists, and execution traces are retained for audit. For an engineering organisation that has already tried a generic assistant on a million-line monolith, that is the right set of concerns. Two areas deserve scrutiny before you commit. The first is commercial opacity: the only paid tier is quoted as Custom, the pricing page is a qualification form, and the published rules amount to per-user licences plus a platform fee. You cannot budget from the website. The second is evidence. SOC 2, HIPAA and GDPR badges sit on the marketing pages with no audit date, certificate or accessible report behind them, and the trust centre shows nothing without JavaScript. The privacy policy is more candid than the badges suggest: it admits no formal EU or UK presence, places most storage in India, and states outright that free-tier inputs feed internal fine-tuning with no opt-out. That candour is creditable, but it is also the thing an EU buyer will trip over. The contractual position on ownership is genuinely good, with your code and outputs remaining yours. Balance that against a liability cap of 100 USD or twelve months of fees and mandatory Delaware arbitration. This is a young company, pre-seed funded in February 2026, selling to enterprises. Pilot it on a real repository, and put the compliance claims in writing.
- Choosing a selection results in a full page refresh.
- Opens in a new window.