prom.codes gives your AI agents two things they are missing: typed, scoped, governable access to your codebase — and a durable memory that survives context-window resets. Three MCP servers — context, memory and a token saver — EU-hosted, bring your own model. No data leaves the EU without your per-tenant opt-in.
prom.codes ships three independent MCP servers — a code context engine, an agent memory and a token saver — plus an EU-sovereign foundation under all of them. Each is replaceable and installs in one line; together they let an agent reason about a real codebase, remember what it learned, and answer leaner — without dragging the whole repo through every prompt. In our own benchmarks, a small, cheap model paired with prom.codes matched the accuracy of a frontier model on the same task — at a fraction of the cost.
The flagship MCP. Tree-sitter parsing across 14 languages, semantic chunking along symbol boundaries, and hybrid retrieval fusing lexical, semantic and symbol-graph signals. Eleven typed tools — search_code, get_symbol, find_references, find_callers, find_callees, expand_context, index_status and more. Recall@10 100% on 12 OSS repos: the file you need to edit lands in the top 10 every time.
See language coverage →A standalone MCP server that gives your agent a persistent memory: facts, decisions and procedures survive context-window resets and session restarts. With cross-encoder re-ranking, focused recall matches the accuracy of replaying the entire conversation — at ~28× fewer tokens (measured on LoCoMo). Local-first SQLite, project memories as git-versioned markdown, secrets rejected on every write.
How memory works →A third MCP that installs an efficient-output rule into your agent — leaner answers, the substance kept verbatim. ~20–25% fewer output tokens on the balanced default (up to ~30% aggressive), with code, caveats, destructive-command warnings and ordered steps never compressed. Pure JS, runs with npx, no key required to try.
How the Saver works →Indexer, code-context store and mail all operated by us inside the EU. A per-tenant region mode pins which providers may ever see your code; no data leaves the EU without your explicit opt-in.
Region modes explained →Most "RAG over code" splits files into fixed line windows and hopes the right chunk ranks. prom.codes parses every file with Tree-sitter and chunks along symbol boundaries — whole functions, classes and methods — then links them in a reference graph. A query lands on a symbol, and the graph pulls in its callers and callees as context.
The second MCP gives your agent a persistent memory — facts, decisions and procedures that outlive context resets and session restarts. Recall is hybrid (keyword ⊕ vector), then a cross-encoder re-ranks: focused recall matches replaying the whole conversation at a fraction of the tokens. Ask in time — "our latest decision on X", "the first approach we tried" — and a temporal pass surfaces the right version. Local-first SQLite, project memories as git-versioned markdown, secrets rejected on every write.
Your IDE talks to a local MCP client. The client speaks MCP to prom.codes, which fans out to the indexer and the code-context store we operate inside the EU.
prom.codes is built and operated by eNetworkers GmbH from Germany. The default deployment is EU-first: your code, embeddings and metadata stay inside the EU. If you choose to opt in to an American model provider for a workspace, that decision is explicit, per-tenant and reversible — keys can be rotated or revoked at any time.
We're onboarding agentic-coding teams in small waves. Drop your email — we'll send setup instructions and an API key when your tenant is ready. No marketing, no tracking, no third-party scripts.