Vibe coding with memory that persists.
Hydrate connects to Mistral Vibe via its MCP server and gives it the same persistent memory layer already shipping for Claude Code. Prior decisions, session summaries, and compressed project docs inject automatically into the context before each prompt. On-prem, auditable, EU AI Act aligned.
MCP tools Mistral Vibe invokes automatically.
Hydrate exposes an MCP server at localhost:24312/mcp.
Add it to Mistral Vibe once and it gains three tools. Mistral decides
when to call them based on the conversation context - the same
mechanism it uses for every other MCP tool.
hydrate_recall Returns prior decisions, session summaries, and pinned canon relevant to the current task. Mistral invokes it automatically when context from past sessions would help.
hydrate_compress Compresses long prose through the local summariser before it reaches the model. Measured 15x-100x ratios on real project docs.
hydrate_read_file Reads a workspace file, auto-compressing when it is bigger than 4 KB. Prevents large files from dominating the context window during vibe coding sessions.
Setup: two lines
Run hydrate setup to start the local server, then add
http://localhost:24312/mcp as an MCP endpoint in Mistral
Vibe's settings. That is the entire integration. No daemon, no
cloud account, no new UI to learn.
The server handles tool registration, session capture, and recall
automatically. Every Mistral-originated request carries
X-Hydrate-Source: mistral so savings attribute cleanly
per client in the audit log.
Want hook parity with Claude Code? Run the fork.
Stock Mistral Vibe exposes one experimental hook (post_agent_turn),
which is enough for session capture but not for automatic context injection
before every prompt. Hydrate maintains a fork of Mistral Vibe that adds the
five missing hook events (user_prompt_submit,
session_start, pre_compact,
pre_tool_use, post_tool_use). With the fork
installed, Vibe gains the same hook surface as Claude Code: facts inject
before the model sees your prompt, compact-survival snapshots fire before
context wipes, and tool calls become visible to the dashboard.
Stock Mistral Vibe
- MCP tools — model invokes
hydrate_recallon its own initiative - Session capture via
post_agent_turn - No automatic pre-prompt injection
- No compact-survival snapshot
- No tool-guard visibility
Hydrate fork of Mistral Vibe
- Everything stock has, plus —
- Automatic context injection before every prompt
- Compact-survival snapshots at
pre_compact - Session-start recovery re-injection
- Tool-use visibility (pre + post)
The fork tracks mistralai/mistral-vibe upstream and adds nothing
Hydrate-specific to the hook layer. The five new events are designed to be
contributed upstream; if Mistral merges them, the fork becomes unnecessary.
Switch tools. Keep context.
Hydrate does not partition memory by client. A decision captured from a Claude Code session injects into the next Mistral Vibe prompt. A file compressed by the Copilot extension is available to Mistral without re-uploading. One store, all your tools.
The same flow works in reverse: Mistral captures, Claude Code recalls. All three surfaces share one store.
Five turns. Two tools. One memory.
A two-developer experiment ran the round-trip end to end: Carol opened a
Mistral Vibe session and proposed three architectural facts via the MCP
surface. Alice opened Claude Code on the same project, saw Carol's
facts injected automatically by the UserPromptSubmit hook,
extended them with two more, and committed. Carol's next Vibe session
received Alice's additions back through the same store.
No fact was retyped. No transcript was copy-pasted. No cloud relay was involved. The store is one local SQLite database; the surfaces are two assistants from two different vendors. This is the capability MemPalace, Pieces, Supermemory, and Mem0 do not currently demonstrate.
Claude Code
shippingNative hooks - UserPromptSubmit, Stop, PreToolUse, PostToolUse. Zero configuration. Works on day one with your existing Claude subscription.
Claude Code page →VS Code + Copilot
v1 launchVS Code extension registering an @hydrate chat participant plus three Language Model Tools. Same memory store - decisions from Copilot flow to Mistral and vice versa.
Copilot page →Mistral Vibe
v1 launchMCP server integration. Mistral Vibe connects to the Hydrate MCP endpoint and gains the same recall, compress, and file tools the other clients use.
Details ↓Your sessions stay on your machine.
On-prem binary
hydrate-server runs locally. The MCP endpoint is
localhost: Mistral Vibe never routes your session
data through a Hydrate cloud service.
Encrypted at rest
The local SQLite store uses AES-GCM encryption. Keys are derived from your machine identity. The file is unreadable if copied to another machine.
Secrets scrubbed
The internal/scrubber package strips common secret
patterns before any storage or LLM processing. Credentials visible
in a Mistral session are never persisted or transmitted.
EU AI Act aligned
hydrate compliance report maps to each applicable Article.
Every figure is backed by a row in hydrate_retrievals
and reproducible with a single sqlite3 query.
In open beta — install live.
Hydrate for Mistral Vibe launches alongside Hydrate's public v1 release. Press, researchers and anyone writing about the launch can request early access.