RushDB agent plugins
RushDB for coding agents and AI assistants. Give your agent a database it can search, query and analyze: RAG over your records, structured and semantic search, relationship traversal, and agentic analytics such as counts, groupings and time series. Every package ships the RushDB MCP tools and skills, built from one source in this repository.
Where a platform has hooks, the package also offers automatic context: it recalls relevant earlier work at session start and stores finished turns in your RushDB project. It is off by default; you turn it on yourself.
Work in progress. The install commands below work once this repository and its listings are published.
Install
| Platform | Install | Automatic context | MCP tools and sign-in |
|---|---|---|---|
| Claude Code | /plugin install rushdb --marketplace rush-db/agent-plugins |
yes, off by default | hosted https://mcp.rushdb.com/mcp, OAuth |
| Qwen Code | qwen extensions install rush-db/agent-plugins:rushdb |
yes, off by default | hosted, OAuth (not yet verified in Qwen) |
| Gemini CLI | gemini extensions install https://github.com/rush-db/agent-plugins |
yes, off by default | local @rushdb/mcp-server, API key |
Antigravity (agy) |
agy plugin install https://github.com/rush-db/agent-plugins |
recall only, off by default | local @rushdb/mcp-server, API key |
| Codex (marketplace) | codex plugin marketplace add rush-db/agent-plugins, then codex plugin add rushdb@rushdb |
yes, off by default | hosted, OAuth |
| ChatGPT and Codex (app) | the OpenAI app directory, once listed | no | hosted, OAuth |
| Pi | pi install npm:@rushdb/pi |
yes, off by default | local @rushdb/mcp-server with an API key, or hosted |
| DeepSeek Harness | dsh plugin --profile <p> add @rushdb/dsh |
yes, off by default | built-in RushDB tools, API key |
- Antigravity reads its own manifest from this repository, and the RushDB
guidance loads as an always-on rule. If you already use the Gemini CLI
extension,
agy plugin import geminiconverts it instead. Antigravity'sStophook doesn't receive the answer text, so the plugin only recalls there; it stores turns only when you ask through the MCP tools. - Qwen Code installs the Claude Code package and runs its hooks, but it
doesn't pass plugin settings to them. Export
RUSHDB_API_KEY(andRUSHDB_AUTO_CONTEXTto turn context on) in your shell, or use the config file below. - Codex: the first command adds the marketplace and the second installs the
plugin (or pick it in
/plugins). Codex asks you to trust plugin hooks: trust them in/hooksif you want automatic context. - ChatGPT and Codex app: the directory build has no hooks, so no automatic context. Install either the app or the Codex marketplace plugin, not both.
Get an API key at app.rushdb.com. Automatic context always needs one (or a self-hosted engine), even where the MCP tools sign in with OAuth, because hooks can't use the agent's OAuth token.
What you get
- MCP tools to read the schema, find, create and update records, aggregate and group, run semantic, vector and text search, follow relationships, and work with imports and connections.
- Skills, synced from rush-db/skills:
rushdb-query-builder,rushdb-data-modeling,rushdb-faceted-searchandrushdb-agent-memory. - Commands in Gemini CLI:
/rushdb:recall,/rushdb:rememberand/rushdb:ask. - Automatic context, optional, described below.
Automatic context
One switch, auto_context:
| Value | What it does |
|---|---|
off (default) |
Hooks exit at once: no network calls, no local files. |
recall |
At session start, recall relevant earlier context from your RushDB project and add it to the session. |
recall+persist |
Also store each finished turn, and the summary written at compaction, in your RushDB project. |
The first of these that is set wins:
- the plugin setting, where the platform has one (
claude plugin configure rushdbin Claude Code,gemini extensions config rushdbin Gemini CLI); - the
RUSHDB_AUTO_CONTEXTenvironment variable; ~/.config/rushdb/config.json($XDG_CONFIG_HOME/rushdb/config.jsonwhenXDG_CONFIG_HOMEis set), a file you write yourself (mode0600; the agent never writes it):
{ "apiKey": "<your RushDB API key>", "autoContext": "recall+persist" }
Recalled context arrives under the heading "Context recalled from RushDB (historical, may be outdated)" and is marked as data from earlier sessions, not instructions. RushDB never blocks your work: if the network or the API is down, hooks log the error and exit cleanly.
What leaves your machine
- MCP tools send what the agent passes in a tool call to the RushDB MCP
server or RushDB API you use. Where the plugin declares the hosted server
(Claude Code, Qwen Code, Codex), the agent connects to
https://mcp.rushdb.com/mcpat session start, before any tool is used, and asks you to sign in once. This happens whateverauto_contextis set to. - With automatic context
offnothing else is sent. Hooks make no network calls and write no files. - With
recall, each session start (in Claude Code also on resume, after/clearand after compaction) sends two search requests, one for earlier turns and one for facts: a fixed question ("decisions, preferences, conventions and open work in this project"), a project scope (a hash of the git remote, or of the working directory when there is no remote, unless you setRUSHDB_CONTEXT_SCOPE) and the session id, so the current session is left out. - With
recall+persist, each finished turn sends the latest user message and the assistant's reply, each cut to a bounded length, with the same scope and session id, plus a summary of the two, the host and hook name (such asclaude:Stop), a timestamp and the turn number. At compaction, the compaction summary is sent too. The first delivery checks for and creates two vector indexes in your project,EPISODE.summaryandMEMORY_FACT.text, if they are missing. - Where it goes: only to the RushDB API URL you configured
(
https://api.rushdb.com/api/v1for RushDB Cloud) or to your own engine. It is stored as records labelledEPISODEin your RushDB project. With a self-hosted engine you also configure an embedding provider, and the same text is sent there to be embedded. - On your machine, with
recall+persist: turns wait in a local outbox (outbox/) until they are delivered,sessions/keeps a turn counter per session, and in Codexprompts/holds the latest prompt until the turn ends. Files are named by a hash of the session id. Errors go tohooks.log(rotated at 1 MB, keys masked). All of these live in the plugin data directory, or in~/.local/state/rushdbwhen the platform has none. - Deleting it: stored turns are ordinary
EPISODErecords in your project. Delete them, and the two vector indexes if you no longer want them, in the RushDB dashboard or with the MCP delete tools. Setauto_contextback tooffto stop sending. - DeepSeek Harness uploads session logs to DeepSeek by default.
@rushdb/dshhas an option to keep its context messages out of that upload.
Self-hosted RushDB
Automatic context can reach your own RushDB in one of two ways.
A self-hosted RushDB API: set RUSHDB_API_URL (or the platform's API URL
setting) to your instance, together with its RUSHDB_API_KEY.
A standalone RushDB engine: set RUSHDB_ENGINE_URL and
RUSHDB_ENGINE_TOKEN (and optionally RUSHDB_ENGINE_SCOPE). An engine has no
managed embeddings, so also configure an OpenAI-compatible embedding provider
with all three of RUSHDB_EMBEDDING_API_KEY, RUSHDB_EMBEDDING_MODEL and
RUSHDB_EMBEDDING_DIMENSIONS, plus RUSHDB_EMBEDDING_BASE_URL for a provider
other than OpenAI (the default is https://api.openai.com/v1). Without them,
automatic context stays disabled and the reason goes to hooks.log. The same settings in the config file:
{
"autoContext": "recall+persist",
"engineUrl": "<your engine URL>",
"engineToken": "<your engine token>",
"engineScope": "<optional scope>",
"embedding": {
"baseUrl": "<provider base URL>",
"apiKey": "<provider API key>",
"model": "<embedding model>",
"dimensions": 1536
}
}
Environment variables win over the config file. When RUSHDB_ENGINE_URL is
set, the engine is used instead of the API URL.
For the MCP tools in Claude Code, add the local MCP server next to the plugin:
claude mcp add rushdb-local --env RUSHDB_API_KEY=<key> --env RUSHDB_API_URL=<url> -- npx -y @rushdb/mcp-server@3.4.1
Repository layout
shared/: one copy of the listing text, MCP settings and skills pin used by every manifest. See shared/README.md.skills/: the synced skills, written bypnpm sync:skills.scripts/:build.tswrites every generated file (--checkfails on drift);sync-skills.tscopies the pinned RushDB skills.packages/: the host-neutral runtime and the npm packages.submissions/: drafts of the listing submissions. See submissions/README.md.
See CONTRIBUTING.md and SECURITY.md.
No comments yet. Be the first to write one.