cd /news/ai-tools/local-privacy-focus-on-cache-ai-cli-… · home topics ai-tools article
[ARTICLE · art-98371] src=github.com ↗ pub= topic=ai-tools verified=true sentiment=· neutral

Local+privacy+focus on cache AI CLI and llama.cpp branch + Qwen3.x

Developer alainnothere released privibe, a fork of Mistral Vibe's CLI coding agent reworked for private, local-first development with no callbacks, running against local llama.cpp servers via an OpenAI-compatible API. The fork removes cloud/account features, telemetry, and remote auth, adds KV cache persistence for session resume, and requires Python 3.12+ and uv. It is developed and tested against a local llama.cpp server, with Anthropic and Mistral backends inherited but unverified.

read5 min views1 publishedAug 16, 2026
Local+privacy+focus on cache AI CLI and llama.cpp branch + Qwen3.x
Image: Michielbdejong (auto-discovered)

CLI coding agent for private, local-first development.

privibe is a fork of Mistral Vibe reworked to not do any call back home of any kind and then run against local models first, I want to be able to use it and know it will not be sending data of any kind anywhere.

The cloud/account machinery is gone, the Mistral SDK is now an optional extra, and a lot of work has gone into making it fast and pleasant to use against a local llama.cpp server.

If you want the original, hosted, Mistral-centric experience, use upstream Mistral Vibe. If you want a small coding agent you can point at your own local llama.cpp server, this is that.

Step-by-step version of this demo: docs/getting-started.md.

Resuming and switching between sessions, plus context files in action: the lets-document skill compiles what one session learned into a context file, and a second, unrelated session picks up from that knowledge without redoing the research (played at 4x):

Resuming a session with its KV cache restored from disk: with the companion llama-server build, evicted conversations are saved to disk and reloaded on resume, so picking an old session back up does not reprocess the whole conversation:

Note on backends:this fork is developed and tested against a localllama.cppserver (via its OpenAI-compatible API). The Anthropic and Mistral backend adapters inherited from upstream are still in the code but arenot actively tested here— treat them as unverified.

Requires Python ≥ 3.12 and uv.

git clone https://github.com/alainnothere/privibe.git
cd privibe
uv sync

I set up an alias on .bashrc

alias privibe='uv run --project /yourPathToPrivibeHere/privibe privibe'

so you can later just use it with privibe from whenever you are, or inside privibe folder with

uv run privibe

On first run it writes a config to ~/.privibe/config.toml

. Point the model entries at your server (e.g. a local llama.cpp instance) and you're going.

There's also an ACP entrypoint for editor integrations:

uv run privibe-acp

For the most part... exactly the same, clone the source and uv run privibe... I use it daily on git bash and works correctly, I have also tried powershell and I know it works.

  • Built around a local llama.cpp server (OpenAI-compatible API). The Mistral SDK is an optional extra (uv sync --extra mistral

), not a requirement, and there's no API-key onboarding gate. - The upstream Anthropic and Mistral backend adapters are still present but are not tested in this fork (see the note at the top).

  • Removed the cloud and account features (nuage/teleport), telemetry, tracing, the update notifier, plan offers, data-retention, and remote auth — along with their dead code and tests. De-branded from Mistral Vibe throughout. If you find anything calling home is not intentional, I miss it, please create an issue and we'll nuke that code into outer space....

  • Conversation state is held in a structure ( ConversationList

) that keeps the prefix immutable, so the server's prompt/KV cache stays valid across turns. - KV cache is preserved on --resume

/--continue

by restoring the original system prompt instead of regenerating it (which would invalidate the cache). - Context-size auto-detection with a /detect-context-size

toggle (re-runs on model swap), and an opt-in model warmup. - Per-message reasoning effort: /effort

cycles off/low/medium/xhigh for new messages without invalidating the KV cache. Needs the companionllama-server build; stock servers silently ignore the stamps. - Long sessions stay responsive: the transcript is windowed behind a "Load more" button, and streaming output is render-throttled so drawing keeps up even when the model generates faster than the terminal can paint.

  • Tolerates SSE keep-alive pings from newer llama.cpp servers.

  • Hashed-line file tools split into explicit single-line, block, and delete operations, plus find_symbol

. - Per-agent file undo stack with arestore_file

tool; writes are serialized. - Tuned for smaller local models: best-effort indent correction on edits, hashed line addresses re-pointed when earlier edits shift them, and naive reads of huge files return a head preview plus guidance instead of flooding the context.

  • Cross-dialect path translation (Windows / WSL / Git Bash / Cygwin) with a [paths]

config section. @

-mention file completion backed by a stateless git enumeration.

Console mode:--console

runs the same agent as a plain-text REPL (no colors, no TUI), including tool output and session resume.Rewind: Alt+Up browses your previous messages; fork the conversation from any of them, and optionally restore files to that point (the file restore is gated behind a typed confirmation code).Agent steering: queue messages mid-turn instead of cancelling the agent.- Double-press Ctrl+C

to exit (no more accidental one-key quits). - Clipboard fixes for Linux/X11 terminals, a /autocopy

toggle, and a warning when clipboard tools are missing. /resume

session picker showing the folder and a short preview of each session, with search: type to filter,-word

to exclude, results ranked by match. Conversation history is re-rendered on resume.- File edits show a diff in the TUI.

  • Model name + tokens/sec in the context footer.
  • Configurable tool-result preview length ( /preview-lines

), scrollback (/scrollback

), and an/llm-debug

dump toggle. - Subagent work is preserved when you cancel mid-execution.

  • Bundled sample skill (lets-document) with auto-discovery and load-error surfacing.
  • Model selection skips entries with a missing API-key env var and falls back to a valid model; config upgrades append commented stubs for new keys.
  • Datetime-based version stamping; .deb

and Windows-zip build scripts.

Config lives in ~/.privibe/config.toml

(set PRIVIBE_HOME

to relocate it). The [paths]

section (documented in privibe/core/config/default_config.toml

) controls cross-dialect path translation. Models, providers, and feature toggles are set there too; several have in-app /commands

(/model

, /config

, /autocopy

, … — see /help

). The active model can also be picked for a single run with --model

.

Apache-2.0. privibe is a fork of Mistral Vibe (© Mistral AI); see LICENSE

.

── more in #ai-tools 4 stories · sorted by recency
── more on @alainnothere 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/local-privacy-focus-…] indexed:0 read:5min 2026-08-16 ·