Meandr Apps · RAG

Your repos and docs, searchable by every agent

Meandr RAG indexes your git repositories, documents and the notes your agents share. Claude, Cursor, Codex and any MCP client search it through your workspace, with no vector database or embedding API to set up.

Try for free See how it runs

Run it

From the Apps page to a searchable index in three steps

Meandr builds and versions the app. You pick where it runs and what it reads, and it starts indexing at once.

  1. 1

    Pick a machine

    Run it on a machine your team already has, or order one. It needs 512 MB, 1 GB is recommended, and 2 GB of disk for its index.

  2. 2

    Choose the repositories

    Pick the ones your team shares from GitHub, private or not, or paste the URL of any public repository.

  3. 3

    Attach it to workspaces

    It shows up as a server. Add it to any workspace, and every agent there gets its tools at the workspace URL.

Tools

Seven tools your agents can use

Results are written for a model: each passage names its source, file, lines and commit, so the agent can cite where an answer came from.

  • search

    Passages for a question, matched by meaning and by exact words, then reranked. Narrow it to some sources, to code, docs or notes, or to a path.

  • read

    A file as indexed, in numbered lines, so the agent can open more around a result and cite the lines it used.

  • sources

    What is indexed: each repository with its commit, files and passages, indexing progress, errors and the models in use.

  • rememberwrites

    Saves a note for every agent that searches: a decision, a fix, a convention, a pitfall.

  • add_documentwrites

    Adds a web page, a PDF or text, turned into Markdown and indexed beside your repositories.

  • forgetwrites

    Deletes a note or a document. Its history keeps it, so it can come back.

  • history

    Every version of a note or document: who wrote it and when.

  • The three that write are marked as such, so a policy can hold them for a person, or keep them from an agent altogether.

What it reads

Code, docs and what your agents learn

Git repositories

Code and docs, picked by file type. Dependencies, build output, lock files, generated files, binaries and files over 1 MiB are skipped.

  • Narrow each repository with paths to include or skip
  • Follow the default branch, or name one

Documents

Web pages, PDFs, Markdown and plain text, added by URL or as text. Pages become Markdown without their scripts and menus.

  • Adding a URL again replaces its document
  • Up to 4 MiB of text each

Notes agents share

An agent that finds a fix or makes a decision saves it, and the next agent finds it in its search.

  • Every version kept, with who wrote it
  • A deleted note can be restored

It stays current on its own

  • On every pushRepositories shared from GitHub are fetched when you push, and hourly besides.
  • Every 5 minutesPublic repositories added by URL are checked on a schedule you can change.
  • Only what changedA changed or moved file is indexed again; the rest keeps its vectors.

Models

Embedding and reranking built in, your own model optional

Included

Embedding

Qwen3 Embedding 0.6B, run by Meandr. Every passage is embedded with its path and heading, so a search for a concept finds the right file.

Included

Reranking

The best candidates from meaning and from exact words are reranked on Meandr's GPUs, so the passage that answers comes first.

Optional

Context, from your model

Connect Anthropic, OpenAI, Mistral, Ollama or any OpenAI-compatible API, and it writes a line of context for each passage, searched and embedded with it. Each passage is sent to that model with its document.

Price

Pay for the machine and what it reads

Run as many RAGs as you like, and share each with as many workspaces as you like. A RAG shared with five workspaces is indexed once.

  • MachineYour plan's machinesFrom 512 MB; several apps and servers can share one
  • Embedding$0.10 per million tokensCounted while it indexes and searches
  • Searches$0.25 per 1,000Searches reranked on Meandr's GPUs

Questions about Meandr RAG

Where does my code go?
The clone and the index stay on your team's machine. Passages go to Meandr's embedding and reranking models to be read, and to your own context model only if you connect one.
How does it read private repositories?
Through the Meandr GitHub App, which only reads the repositories you choose. For each sync the RAG gets a token for one repository, valid for an hour, and it never sees your GitHub credentials.
Which git hosts work?
Repositories shared with your team from GitHub, public or private, and any public repository over HTTPS by its URL.
Can agents change what is indexed?
They can add notes and documents, never repository files. Those three tools are marked as writing, so your policies can hold them for approval or deny them, and every version stays in history.
Do I need a vector database or an OpenAI key?
No. The index lives with the app, and embedding and reranking are built in. A key is only needed for the optional context model.

Give your agents your team's knowledge

Run Meandr RAG on your machine, pick the repositories, and every workspace you choose can search them.

Try for free

Questions? Email us