One index for everything you have.
Nothing leaves where it lives.
inseam is an open retrieval layer that sits between your AI and your data — your inbox, your files, your notes, the calls coming in, the tools your team uses. It doesn't copy anything. It runs on the hardware you already own. It's the layer every other AI-over-your-stuff product skips past.
There are a lot of these.
Four reasons this one is different.
AI-over-your-data is a crowded category. Most products in it converge on the same shape: copy your corpus into a vendor cloud, charge per seat, search it with whatever search API the vendor decided to ship. That's not the right shape. Four decisions make inseam a different thing.
Every channel, every medium, reachable from anywhere.
inseam indexes addresses, summaries, and embeddings — not the underlying content. A phone call comes in: the Twilio plugin picks it up, inseam indexes it, the caller's number ties it to the customer record already in the index. The audio stays at Twilio; the conversation is now retrievable from anywhere else, scoped to that customer. The same shape covers email, drive, notes, chat, tickets, voice memos, calls, photos, the niche tool your team depends on — one index over all of it, fetched live only when something actually needs the bytes. Nothing copied. Nothing moved.
Gemma, and the local-LLM cohort actually work.
Most RAG hands a small model fifteen marginal chunks and hopes. A 3B on-device model collapses under that noise. inseam hands the model a navigable tree — workspace down to folder down to document down to section — and lets it walk until it has the slice it can actually reason over. The shape that makes the index fit on the device is the shape that makes the model on the device useful. No data egress. No API bill. Quality that used to require a frontier model with a cloud RAG stack behind it.
an ecosystem you're not locked into.
MIT-licensed, community-developed. No enterprise edition, no proprietary index, no migration path because there's nothing to migrate off of. The plugin contract is intentionally small — small enough that an AI can fill one in for the niche tool your team uses, in an afternoon, against framework-enforced fixtures and tests. The connector catalog isn't bounded by a vendor's headcount — it's bounded by how many people have an agent and ten minutes.
so agents can speak between parties, not just within them.
Every other system treats access as "is this user inside our workspace?" inseam splits it: Connection (the credential), Source (the address), Access (who's allowed to retrieve what) — three things, kept separate on purpose. The identity signals already on your data — From/To, account ids, party fields — become scoped read paths. A customer authenticates to your agent and asks about their thread; the agent answers without leaking into anyone else's. Same for clients, patients, counterparties, contacts. AI that operates between people, with the context for each, over time.
The architecture underneath is small and worth the read. How it works walks through the index shape and the three-way split.
Different people show up here
for different reasons.
The thing underneath is the same. The reason it matters depends on who you are and what you're trying to do. Pick the path that's closest.
Ship product agents that scope to a customer's slice. Write a plugin for the weird internal tool in an afternoon. Embed the retrieval layer instead of building one from scratch per source.
A personal AI that actually knows what you know — over email, files, notes, photos, code. Running on your laptop, your NAS, your phone. No subscription. No egress. Works on the plane.
Fix the data layer first; the AI follows. A lean addition on top of what you already run — not a re-platforming project, not an upload-everything migration, not a vendor you can't leave.
Comparing against Glean, Onyx, Bedrock, an MCP build-it-yourself? The axes that matter — index shape, ranking, permissions, lock-in, deployment posture — written plainly, with the comparison table.
What you're getting.
What you're not.
The marketing page is usually where software lies to you. Instead of adjectives, here's the actual deal.
- Open source. The kind you can read, run, and fork.
- Yours to run — your machine, your server, your call.
- Built so the community extends it. No gatekeeper.
- Real about who-sees-what. Per person, not per workspace.
- The layer between AI and your stuff. Not another place to put it.
- Something that gets out of the way once it's working.
- A vendor cloud. A SaaS with a per-seat invoice.
- A second copy of your data on someone else's hardware.
- A chatbot. inseam isn't the AI — it's what the AI talks to.
- A list of supported tools you have to pray yours is on.
- An enterprise upsell gating the features that matter.
- A pivot waiting to happen — open source, no investor to please.