Investor brief Synaptic · 2026
Synaptic is the persistent, auditable code-intelligence layer for coding agents and engineering teams. It turns a repository into a source-linked graph, so a change can be scoped with real evidence — instead of re-reading the codebase every time.
crates/synaptic-server/src/lib.rs:L491A real excerpt from Synaptic's own repository graph — 6,741 nodes, 21,327 edges, 154 communities.
01 The problem
To make a change today, a coding agent searches, greps, and re-reads files, reconstructs the architecture inside its context window, then edits — often without a complete, system-wide view of what the change touches. The reconstruction is thrown away after every task.
A large, multi-language codebase.
Guess at relevant files by name and text.
Load raw source into the context window.
Rebuild the architecture from scratch, every task.
Change code with a local, partial view.
Find the downstream impact after it ships.
02 The Synaptic approach
One extraction turns a repository into a queryable graph of symbols, dependencies, APIs, SQL objects, events, and runtime relationships. Every edge keeps its source location and is labeled extracted, inferred, or ambiguous — so an answer is auditable, not a guess.
One repo, a monorepo, or a federated fleet.
A persistent graph.json of nodes and typed edges.
Ranked subgraphs, reverse impact, structural search — from CLI, REST, or MCP.
A relevant slice plus blast radius and evidence — not the whole codebase.
03 What it does
Grouped from the capabilities shipping in the product today.
Ranked subgraph search, source reading, shortest paths, community structure, and per-file outlines.
query_graphSynQLget_sourceReverse traversal connects a change to callers, references, dynamic risks, and the tests that cover it.
affectedpredictaffected_testsProvenance on every edge, real test execution, and honest handling of what static analysis cannot see.
speculatesql audithazardsDependencies that are not imports — protocols, processes, languages, and repositories.
30 toolsdiff04 Benchmarks & evidence
These are the current reproducible benchmarks — on Synaptic's own code and a hand-labeled corpus. They validate correctness on representative shapes, not universal production accuracy.
Fewer tokens to answer a structural question than reading the files it references (27–38× across six questions). Exact cl100k counts on Synaptic's 199-file, 510,966-token source.
On the hand-labeled corpus, with 0 distractor nodes leaked — the impact walk is not over-broad.
Across five boundary kinds and six couplings, with six look-alike distractors correctly left unconnected.
| Correctness metric | Result | Scope |
|---|---|---|
| Call edges (P / R / F1) | 100 / 94 / 96 | 17 labeled edges |
| Blast-radius recall / exclusion | 100% / 100% | distractor nodes held out |
| Affected-test recall | 100% | multi-hop; unrelated test excluded |
| Cross-language (P / R / F1) | 100 / 100 / 100 | 6 couplings, 6 look-alike distractors |
| Corpus size | 11 fixtures · 6 language families · 41 labeled symbols (all resolved) | |
05 Why now
Models keep getting better and cheaper at generating code. As agents take on larger changes across bigger systems, reliable understanding of those systems — not generation — becomes the constraint. That layer looks like infrastructure.
A person holds the mental model and makes the change.
Agents draft code; a person still supplies and checks the system context.
Larger, multi-file, cross-service changes made with less human mediation.
06 Technical moat
The value is in the extraction and the graph, not a prompt wrapper. This is infrastructure work.
Structure comes from tree-sitter parsing across 30+ languages, each grammar built and tested in isolation in CI — not from a model that might hallucinate an edge.
Calls, imports, inheritance, HTTP/gRPC/queue/IPC/FFI handoffs, and SQL lineage as first-class typed relationships, resolved across service and language boundaries.
Every relationship is labeled EXTRACTED, INFERRED, or AMBIGUOUS, and unresolved dynamic dispatch is surfaced as a hazard rather than invented.
Separate repositories federate into one graph with real cross-repo edge resolution, so architecture stays visible past a single repo boundary.
File watching, git hooks, and per-commit caching keep the graph current without a full rebuild; an AST cache makes warm rebuilds fast.
A read-only MCP 2026-07-28 server with stateless discovery and 30 tools gives agents ranked, cited context. The graph is built for machines and humans alike.
07 Who we're building for
The product capabilities ship today. The commercial packaging below is where they turn into what an organization buys.
Current product capability and future commercial packaging are called out deliberately. Nothing here implies traction that does not yet exist.
08 Business model
This is the current strategy, not a finished go-to-market. The engine is source-available; the hosted and enterprise control plane is where revenue is designed to sit.
The source-available engine (FSL-1.1-ALv2) drives adoption and trust with individual developers.
A managed control plane adds shared graphs, hosted MCP access, history, and roles.
Self-hosted deployment, signed licensing, identity, governance, and support.
09 Traction & velocity
The story here is development velocity and surface area, not vanity metrics. Everything below is verifiable in the repository.
10 Competitive landscape
Synaptic does not replace the editor, code search, or the security scanner. It preserves a structural answer those tools can consume.
Sourcegraph · SCIP
Compiler-backed definitions and references.
Dependency and impact paths, including inferred runtime handoffs, with source-linked caveats.
Cursor · Copilot
Semantic retrieval into a coding session.
A persistent graph queryable again from editor, CLI, CI, REST, or MCP.
CodeQL · Semgrep
Purpose-built data-flow and taint analysis.
Everyday architecture, blast-radius, test-selection, and SQL-lineage questions.
Vector retrieval
Fuzzy recall of similar-looking code.
Exact, typed relationships with provenance — not similarity.
Synaptic's thesis: persistent, queryable, source-linked structural intelligence designed for both humans and agents — the layer underneath coding agents, not another wrapper around them.
11 Team
Designed and built Synaptic end to end: the Rust extraction engine and graph, 30+ language extractors, the reproducible benchmark suite, the MCP integration, and the hosted Next.js control plane deployed on GCP.
See the work on GitHubSynaptic is a solo-built project, and the surface area above is the case for the founder: a source-available engine, a hosted platform, and a reproducible benchmark suite shipped to public release v0.6.4 — engineering-led, with evidence over claims.
The near-term hiring plan follows the work — extraction coverage, platform reliability, and go-to-market — as usage warrants it.
12 Current investors
None yet.
Synaptic has taken no outside investment. It has been built independently by its founder. The zero here is deliberate — a starting point, not a placeholder.
We're beginning conversations with investors who believe reliable, auditable understanding of software will be core infrastructure for autonomous development. No round terms are set here — this is the opening of that conversation.