Knowledge Management
SourceAnt captures the knowledge behind your code (decisions, constraints, workarounds, conventions, patterns, discussions, reviews, and plans), structures it into a knowledge graph, and serves the relevant slice to your team and your AI tools. This page explains how the knowledge is stored and the two ways you can run it.
How the knowledge is stored
- Captured once, served everywhere. Approved knowledge is stored durably (your workspace on the cloud, your own database when you self-host) and served back over MCP to any tool that speaks the protocol, so it survives tool changes and session resets.
- Humans govern it, not every write. Capture is automatic and never waits on a person. Knowledge carries a status your team controls and your tools can filter on, so approval is how the team marks what it vouches for, on its own schedule, not a gate every write queues behind.
- Typed, not prose. Every context object has a kind: Decision, Constraint, Workaround, Convention, Pattern, Discussion, Review, or Plan. Typed objects link to code and to each other, which is what makes precise retrieval possible.
Three ways to run it
1. SourceAnt Cloud
The fastest path. Connect your repositories at app.sourceant.ai and SourceAnt runs the intelligence for you:
- Automatic context capture, with a proposal queue your team curates on its own schedule.
- The knowledge graph, explorable and served to any MCP client with an API key.
- Analytics on token savings, review quality, context coverage, and triage.
- Applications on top: context-aware review and issue triage, already wired.
See Cloud Setup to get started, and Pricing for the bands and what the models cost.
2. Open core (your infrastructure, your tools)
SourceAnt runs an open core model. The engine is in the open, MIT-licensed core and runs on your own infrastructure with no account and no limits. It gives you:
- Knowledge management over MCP. A bundled MCP server lets any client create, update, search, connect, and retrieve scoped knowledge: decisions, rules, constraints, conventions, notes, and more. Run it as a local stdio server, or serve MCP from the SourceAnt HTTP server with token authentication.
- Durable storage on your database. Knowledge lives in the core's SQL database (PostgreSQL in a normal deployment, SQLite for local use) and survives restarts.
- System topology. Repositories, services, and the relationships between them, stored durably and inferred from the manifests your repositories already publish.
- Review and triage on the same context interfaces, with any LiteLLM-compatible model provider. See Self-Hosting.
- Replaceable adapters and plugins. Swap in structural indexers, graph databases, or custom knowledge repositories, and extend the runtime through the plugin system.
The cloud runs that engine with the managed layer on top: memory that persists and decays, a proposal queue your team curates, contract analysis and structural code indexing, the explorable graph, workspaces and team roles, Lens for reading a review, hosted serving with API keys, and analytics. That layer is what a team works in day to day, and it is what a paid plan is for. If you need it inside your own network, that is the on-prem plan below.
3. On-prem (enterprise)
For teams whose code never leaves the building: the full cloud experience deployed inside your own network, with premium plugins included. Air-gapped setups are supported and it works with your own model endpoints. There is no event metering on-prem, and it is scoped and priced with you rather than off a band, so talk to us.
Which path should you pick?
| You want | Pick |
|---|---|
| The full experience with zero infrastructure: capture, graph, MCP serving, analytics | SourceAnt Cloud, any plan |
| The cloud experience with your AI spend and data on your side | SourceAnt Cloud, with your own model keys |
| The engine over MCP plus review and triage, on your servers, under your full control | Open core, self-hosted |
| The full cloud experience inside your network, air-gapped, premium plugins | On-prem (talk to us) |
Serving knowledge to your tools
SourceAnt speaks MCP (Model Context Protocol), on every path. On the cloud, generate an API key in Settings and point any MCP client at your workspace. Self-hosted, run the bundled MCP server from the repository (or enable MCP on the HTTP server) and point your client at that. Either way, Claude Code, Cursor, Zed, or anything else that speaks the protocol queries the relevant slice before it acts, so it arrives already knowing the decisions behind the code.