Get started

Install rh and capture your first session.

A few minutes. At the end, one of your tools is connected and a new line in the capture log shows your first session is on the record.

Linux or macOS (or WSL on Windows) and one supported tool. No model, database or account.

Hand it off

Let your agent do the setup.

Paste this into any coding agent. It follows the setup brief, which is the three steps below written for an agent.

prompt
Read https://rivethub.io/llms.txt and set up RivetHub on this machine.

Or do it yourself

  1. Install rh

    Linux or macOS
    $ curl -fsSL https://get.rivethub.io/local.sh | bash

    This installs rh and starts a node on this machine that keeps everything in one file, ~/.rivethub/memory.sqlite. On Linux the node runs as a service, so it starts again after a restart.

    What this does, and how to keep it offline

    The installer picks the build for your machine, verifies it against the release's SHA256SUMS, puts rh and rivethub-migrate in ~/.local/bin, and starts the node. The node listens only on 127.0.0.1:5174. Linux builds need glibc 2.35 or newer.

    Read the installer before you run it, or install by hand.

    To keep everything on your network, add -s -- --no-cloud (or set RIVETHUB_CLOUD=off). The node's config then carries cloud: {disabled: true}: no cloud sign-in, crash reports, usage counts, cloud update checks or sharing through rivethub.cloud, and users can't turn them back on. Updates come from GitHub; sharing goes only to the team servers you list in memory.share.allowed_servers.

    On Linux with systemd, the installer then hands the node to a user service, rivethub.service (rh service init), so it starts again after a restart: when you log in, or at boot if lingering is on for your user. The installer says which, and gives the command that turns lingering on; it does not run it. Add --no-service to skip the service.

    On macOS, under WSL without systemd, or with --no-service, the node does not start by itself after a restart; start it with rh local --ensure. With --no-start nothing is started: run rh local --ensure (rh local --no-cloud for an offline node).

  2. Connect a tool

    One command adds the capture hook and the recall (MCP) entry to a tool's own settings. It prints the diff and backs the file up before it writes; add --dry-run to look first.

    $ rh harness connect claude-code
    ToolCommand
    Claude Coderh harness connect claude-code
    Codexrh harness connect codex
    Grok Buildrh harness connect grok-build
    OpenCoderh harness connect opencode
    DeepSeek Harnessrh harness connect dsh

    Codex asks you to trust the two new hooks the first time you open it after connecting: choose Trust all and continue. Until you do, Codex does not run them and nothing is captured, including from codex exec; rh harness status says whether Codex trusts them.

    rh harness status lists the tools it found and what is wired. OpenCode captures through a small plug-in rather than a hook; How it works lists what each tool captures, and the runtime's README covers the details. rh harness disconnect undoes any of it.

    Upgrading from an older rh or OpenCode 1.x

    OpenCode 2 is the supported version. If you connected OpenCode with rh 0.8.5 or older, run rh harness connect opencode again after updating. OpenCode 1.x is captured from rh 0.8.6 and gets no further work.

    DeepSeek Harness connects from rh 0.8.4; on an older rh the command answers with an error, so run rh update first.

  3. See your first captured session

    Open the tool you connected and do a short task. Then check the capture log:

    Claude Code; the other tools log to codex-capture.log, grok-capture.log and so on
    $ tail -n 3 ~/.rivethub/claude-capture.log
    2026-11-11T16:44:12Z Stop claude-code:1111…5555: updated conv … +4 msg (had 0)

    A line like that means the session is on the record. Now ask the agent, in a new session, to search RivetHub memory for something you said in the first one. It calls memory_search and gets your own words back.

    Did it work? One click tells us whether this guide does its job. If you accepted site analytics, it records which button you pressed; otherwise nothing is sent.

    Stuck? rh doctor checks the node and rh harness status shows what is wired. Support lists the usual fixes and what to include when you ask for help.

After your first session

What to turn on next.

Bring in the sessions you already have

Your tools have been keeping their own history all along. One command reads every session that Claude Code, Codex, Grok Build, OpenCode and DeepSeek Harness have stored on this machine and imports them into your local record.

every tool on this machine
$ rh capture backfill

Add --dry-run to count the sessions without adding any, or --harness codex to read one tool. The command-line reference covers where each tool keeps its history.

Session data in another format can come in too, such as a data export from SpaceXAI, Anthropic or OpenAI. rh has no importer for those, so hand the file to your agent: the setup brief tells it the format your node accepts.

prompt
Read https://rivethub.io/llms.txt and import the sessions in this export into RivetHub. Show me a few first.

Summaries, topics, the wiki and semantic search

Out of the box you have capture and keyword search. Point rh at a chat model and it writes summaries, files sessions into topics and builds plain Markdown wiki pages under wiki_dir. Point it at an embedding endpoint and search understands meaning as well as words. Any OpenAI-compatible endpoint works, local or hosted. Whatever endpoint you choose will see the session text you send it.

~/.rivethub/config.yaml, added to what rh local wrote
runtime:
  default_agent: memory
agents:
  memory: { provider: local, model: my-model }
providers:
  local: { type: openai-compatible, base_url: http://127.0.0.1:8000/v1, model: my-model }
memory:
  sqlite:
    workers: true
    tagging: true
    wiki_dir: ~/.rivethub/wiki
    embed_endpoint: http://127.0.0.1:8001/v1
    embed_model: my-embedding-model

Check it with rh config validate, then restart the node: rh service restart, or rh stop and rh local --ensure if it does not run as a service. How it works explains what each part does.

More machines, or a team

A standalone node serves one person on one machine. When you want every machine you use, or several people, on one record, move to a datahub: Postgres, the memory workers beside it, and a users file. rh memory bundle packages this node's sessions for the datahub to ingest, so nothing is left behind.

How sharing works   Datahub setup in the operations guide

Read your memory in an app

The RivetHub app, for reading, searching and tidying your memory, is in preview for Linux and has no release yet. Until it ships, your agents search your memory over MCP and you can read your wiki Markdown files directly under ~/.rivethub/wiki. Where the app stands

Keep it current

rh update downloads the latest release, verifies it, migrates your memory and swaps the binaries, keeping the previous ones beside them. rh update --channel beta takes prereleases.