RIGHT FILE IN THE TOP 10
76.8%
against 47.0% for grep then read. First right file in a median 224 tokens, not 15,491.
500 SWE-bench Verified issues, retrieval only. The scorecard →
One index · every AI agent · on your machine or in the cloud
Semlith
One index of everything you work on, for every agent you run.
On SWE-bench Verified, the file the fix edits is in Semlith's top 10 for 76.8% of 500 issues, against 47.0% for grep then read. Retrieval only. See the benchmarks →
Every agent asks the same index — and gets facts, not guesses.
On your laptop, in a cloud session, in a CI job. Each answer states the commit it was read at. Semlith runs no chat model, your agent keeps its own.
It reads GitHub, GitLab, Confluence, Jira, Your folders, Uploads.
01 — The problem
Every agent starts from zero, inside one repository.
Your agents run everywhere now — a terminal, an editor, a cloud session, a CI job. Each one sees a sliver of what you know.
02 — Reach
One index. Every agent. Every machine.
Semlith runs on your machine and is free. Semlith Cloud is the paid, hosted version. It keeps one index that every agent can reach, on any laptop, in the cloud or in CI, for you alone or your whole team.
SEMLITH · FREE
Semlith
Free and open source. It indexes your code and docs on your machine for the agents you run there.
Install SEMLITH CLOUD · PRO · TEAM
Semlith Cloud
Paid and hosted. One index of all your repos, docs and tickets that every agent can reach. On Pro it's yours alone. On Team the whole team shares it.
See Semlith Cloud Same tools, same answers — pick where it lives:
- reaches itneeds CloudIn your terminalClaude Code · Codex · OpenCode · Gemini CLI · Copilot CLI
- reaches itneeds CloudIn your editorCursor · Copilot in VS Code · Zed · Cline
- reaches itneeds CloudOn your desktopClaude Desktop · ChatGPT desktop
- reaches itneeds CloudIn a cloud sessionAgents running on the web, sandboxed to one repository
- reaches itneeds CloudIn CI and routinesOrg tokens are machine identities — free, never a seat
- reaches itneeds CloudOn your other machinesIndex once; every laptop asks the same store
MCP URL · one for every client
https://cloud.semlith.com/acme/mcp
claude mcp add --transport http semlith \ https://cloud.semlith.com/acme/mcp \ --header "Authorization: Bearer sml_live_k9f2…"
One hosted copy of your index. Tokens are shown once and scoped to the stores you allow. Queries are never metered.
03 — Breadth
The answer is wherever it is. Semlith finds it there.
Every repository, page, ticket and file you pick goes into one index…
…and each piece lands on a map by what it's about — not where it's stored.
The question lands beside its answers: the code, the Confluence page, the Jira ticket, the PDF on a drive. (Exact names still work: it checks the words and the meaning, then combines both.)
04 — Proof
Facts from the graph. Never a model's guess.
Semlith maps which piece calls which — across repositories. So an agent can ask what uses this? and what would this change reach? There is no generative model in the retrieval path, so the index cannot make anything up.
Extracted — the file says so
Resolved — one definition matches
Ambiguous — it says how many
Inferred — matched by name only
Hover a piece →
Watch it light up everything it touches, across repositories.
05 — Impact check
Know what a pull request reaches.
On every pull request, Semlith posts one status check and one comment: what changed, which callers it reaches — even in another repository — and what it leaves dangling. No model reads the diff. Every finding is a graph fact with a file:line.
When there are findings, conclude
One caller sits in another repository. Neutral asks for a look; it never blocks a merge.
semlith/impact—neutral4 findings, 1 across repositories
semlith commented on acme/api #412 · Add order splitting
Changed · 3 symbols
create_order(cart) → create_order(cart, split)src/orders.rs:142
total(order) (body only)src/orders.rs:201
struct SplitPolicy (new)src/orders.rs:118
Reaches · callers not in this PR
handle_checkoutacme/web › src/checkout.ts:88resolvedanother repository
retry_ordersrc/jobs.rs:57resolved
test_create_ordertests/orders.rs:19resolved
Left dangling
none
No caller yet
SplitPolicy::default :130
Related documents · name a touched symbol
acme/docs › specs/orders.md §3.2resolved
acme/docs › adr/0007-split-shipments.pdf p.2inferred
Semlith posts facts from the index, never an opinion about the change.
On every Semlith Cloud plan, for GitHub pull requests, through the GitHub App. The comment can be turned off; the check stays. See plans →
06 — The ledger
A record of what every agent read.
Every retrieval — from a laptop, a cloud session, or an impact check — is a row, and each row is chained to the one before it. Change one, and the chain visibly breaks. Click a row to try.
✓ All 6 rows check out. Each one carries a fingerprint of the row before it.
Query text stays yours
The binary keeps what was asked on your machine. Nothing synced to Cloud carries it, and Cloud's own ledger keeps it off unless you turn it on.
Per agent, per repository
See which client read what, and how much it saved over reading whole files.
Free on your machine
Recording is in the binary. Cloud keeps 30 days and exports CSV or JSON; Enterprise streams it to your SIEM.
07 — Always up to date
You save. Or push.
Fix a bug, merge a pull request, edit a spec, drop in a PDF.
Semlith notices.
On your machine it watches quietly — under a second of work per idle minute. In the cloud, a GitHub push re-indexes; GitLab, pages and tickets are checked about once a minute.
It re-reads only what changed.
One file changed? One file re-read — vectors and graph edges together. Everything else stays where it was. Rename a file and nothing is re-embedded, because vectors are cached by what the code says, not where it lives.
Ready to be found.
About a second from save to searchable on your machine. In the cloud, every source states the commit it was read at, and how far behind it is.
~1s
08 — Private by construction
The binary never sends anything. Cloud is your choice.
The free binary
- No account. Nothing to sign in to.
- Loopback only. It listens on 127.0.0.1 — a door that only opens from the inside.
- --airgap refuses every outbound request before a socket opens: model, upgrade, prices, cloud. Without it, the only fetches are model weights and acceleration packs on first use, each pinned by digest, and what you ask for.
- Credential files stay out. Credential files (.env, keys, ~/.ssh) are never indexed unless you pass --include-secrets. A file whose contents look like a secret waits for your decision.
Semlith Cloud, if you choose it
- Read only. It never pushes, edits or deletes. The one write is the impact check, where you allow it.
- Only what you pick. Connector tokens encrypted at rest; org tokens kept only as a hash. Stores never cross organisations.
- No model is trained on it. No third-party AI service ever sees your content.
- Your store is yours. It downloads and opens in the free binary, offline.
check it yourself:sudo tcpdump -i any -n 'not host 127.0.0.1'
09 — Documents, too
Agents make up quotes from files they can't read.
Grep can't open a slide deck or a spreadsheet. Semlith reads 13 document formats — and pictures, by what they show — so the agent quotes the real passage. Measured on 127 questions over 11 formats:
PASSAGE FOUND
98%
with semlith, against 63% with grep, glob and read.
WITHOUT SEMLITH
41 of 127
top quotes were made up — not in the document they named.
AND CHEAPER
$0.043
per session with semlith, against $0.067 without. More accurate and cheaper at once.
- Word
- PowerPoint
- Excel
- OpenDocument
- EPUB
- RTF
- Email (EML, MBOX)
- HTML
- Notebooks
- Markdown & text
- Images, by what they show
- Code, 46 languages
An Opus agent on 53 public documents. It chose semlith on its own in 127 of 127 sessions.
10 — Measured
Every number says what it was measured against.
84%
of two-part questions
against 47% for rg · multi-hop
0.88
recall on "who calls this?"
graph tools, against 0.71 for rg
879k
chunks across 70 repositories
first index about 71 min on an M1 Air, search about 100 ms for a name and 260 to 360 ms for a concept
~1 s
from save to searchable
on your machine, 1.24 s measured with watch running
230/s
chunks indexed per second
Apple M1 Neural Engine, on a laptop
46 · 13
languages · document formats
plus 5 image types
12
AI assistants tested
terminal, editor and desktop clients
3,982
tokens read per question
against 105,132 opening whole files, on semlith's own repository, an upper bound
WHERE IT PAYS OFF
On one small repository, a strong agent with grep does fine on its own. Semlith earns its place on wide and unfamiliar codebases, many repositories, "who calls this?", and documents. Every public benchmark, win or lose →
Index once.Every agent asks.Every read on record.
SEMLITH · RECEIPT
your visit, written down
visitoryou
readthis page
time here0 s
sections seen1 / 12
bytes sent anywhere0
chain-------- ✓
thanks for reading
One line to get it
No account. Nothing to sign up for.
$ curl --proto '=https' --tlsv1.2 -fsSL https://raw.githubusercontent.com/semlith/semlith/main/install.sh | shmacOS (Apple silicon) · Linux (x86_64, ARM) · Windows (x86_64) — then it sets itself up for the AI tools you use.
- Terminal
- Claude Code · OpenAI Codex · OpenCode · IO CLI · GitHub Copilot CLI · Gemini CLI
- Editors
- GitHub Copilot in VS Code · Cursor · Zed · Cline
- Desktop apps
- Claude Desktop · ChatGPT desktop (the Codex app)
Semlith Cloud, the optional hosted index for you or your team, reads GitHub, GitLab, Confluence, Jira and uploads, read only. See Semlith Cloud · What it reads and keeps
github.com/semlith/semlithopen source · Apache-2.0built in Rustno network at question time