Skip to content

Meru v0.7.0

Choose a tag to compare

@aarora79 aarora79 released this 03 Nov 01:07
· 141 commits to main since this release
1b1437e

Release v0.7.0 - The router picks which model answers, and answer models can run on your own network

November 2026

An answer model can now run on another model server: one on this computer, or a GPU box on your own network that speaks the OpenAI chat completions API. Meru asks before a question goes to that box. With routing on, the fast model answers easy questions, and you can ask a larger model when an answer falls short. A new edit_file tool changes part of a saved file.


Upgrading from v0.6.4

Breaking changes

There are no breaking changes in this release. A config without the new keys loads and answers as before: the main model writes every answer, and nothing leaves the computer. The transcripts and meru.db keep their format. Older clients keep working with the new merud. A consent question reaches one as an ordinary approval, and a no keeps the turn on this computer.

New config keys

Key Default What it does
[[endpoints]] none An answer model on another model server: name, reach ("local" or "near-edge"), url, model, context_length, and optional vision, tools and api_key ("secret:<name>"). [models] main, large or a set's main names it as "@name".
[models] large "" A bigger answer model, local or "@name", that runs only when you ask for it.
[router] answers "main" "routed" lets the fast model answer easy questions; "main" keeps every answer on the main model.
[router] easy_above 0.8 How sure the router must be before the fast model answers, with answers = "routed".
[builtin] tools and confirm adds edit_file edit_file is on by default and asks before each edit.

If your config lists [builtin] tools by hand, Settings and meru tools show edit_file as New until you add it.

Upgrade instructions

The one-line installer on a Mac. Run it again; it keeps your config.toml:

curl -fsSL https://github.com/aarora79/meru/releases/latest/download/install.sh | bash

The Mac installer. Download the disk image,
open it, and run "Install Meru.app". It leaves your config and memories alone.

Linux. Run install-linux.sh again; it keeps your config.toml:

curl -fsSL https://github.com/aarora79/meru/releases/latest/download/install-linux.sh | bash

From source.

cd meru
git pull origin main
git checkout v0.7.0
go install ./cmd/merud ./cmd/meru
launchctl kickstart -k gui/$(id -u)/com.meru.merud   # Linux: systemctl --user restart merud

Then, on a Mac that reads Apple Notes: macOS drops the Full Disk Access grant
when the merud binary changes. In System Settings, open Privacy & Security,
Full Disk Access, and turn merud off and on again.


Major features

Answer models on another model server

An [[endpoints]] entry names an answer model on vLLM, llama.cpp's server, LM Studio or Ollama's own /v1, and main = "@name" points at it:

[models]
main = "@gpu"

[[endpoints]]
name           = "gpu"
reach          = "near-edge"
url            = "http://192.168.1.50:8000/v1"
model          = "Qwen/Qwen3-32B"
context_length = 32768
api_key        = "secret:gpu_key"   # only if the server needs one
  • Two reaches. local takes a loopback URL. near-edge takes a computer on your own network or private cloud network. Config refuses a public IP address when it loads, and merud checks every address a host name resolves to before it connects. Meru has no reach for the internet.
  • The fast and embedding models stay home. Routing, skill picks, summaries, indexing and search always run on the local Ollama.
  • One API. Every endpoint speaks OpenAI chat completions, with streaming, tool calls and pictures.
  • Keys stay in secrets.toml. A missing key leaves merud running, and a question for that model says what to add.

PR #214, PR #218

Meru asks before a question leaves the computer

Before a turn goes to a near-edge endpoint, Meru.app, meru chat and meru ask "Send this question to @gpu?" and list what the turn carries. The answers are This time, This chat or No.

  • "This chat" lasts until merud restarts. Meru never remembers a yes beyond that.
  • A no keeps the turn on this computer: the fast model answers, and a notice says why. An incognito chat, a scheduled job and meru run --json always stay here.
  • An answer from the endpoint says where it came from, such as "answered by @gpu (near-edge)".
  • The transcript records each answer to the question, never the prompt.

PR #215, PR #218

The router picks which model answers

With [router] answers = "routed", the fast model answers a plain question the router is sure about, such as "what is the capital of Japan?". Every other question goes to the main model. Picking a local model never asks you anything.

  • Escalation. If the fast answer runs out of room, uses every tool round, gives up or breaks a tool call, Meru asks the main model once. Meru.app folds the first try into a closed "First try" block, and meru chat folds it to one line.
  • The large model, on request. Set [models] large, then use "Ask the large model" in Meru.app, /large in either client, or meru -large. Meru offers it after a main answer ends short. It never runs on its own.
  • Benchmarks. meru check records which model answered, and make bench ANSWERS="main routed" runs each model set both ways.

"main" stays the default until more benchmark runs settle it.

PR #217


What's new

Tools

  • edit_file replaces exact text in a file in ~/meru-output/, keeps the old copy as a hidden .bak, and asks before each edit. (#214)
  • write_file and edit_file return a link to the file, and an answer that leaves it out gets a "Saved:" line with the link. (#214)

Documentation

  • AGENTS.md's first rule, ARCHITECTURE.md and the 300, 200 and 100 pages now say no prompt goes to the internet, and none leaves the computer without your yes. (#215, #218)
  • A FAQ page on answer endpoints, the model-router design, and the slide decks and poster for this release. (#214, #217, #218)

Bug fixes

  • write_file no longer writes a second copy under ~/meru-output/meru-output/ when the model gives a path that starts with the folder's name. (#214)
  • An answer that claims it changed a file, after a turn that changed nothing, now gets caught. (#214)
  • A tool-call parse error that Ollama returns before any text now retries the round, instead of ending the turn. (#214)
  • analyze_table accepts its lists sent as a string. (#214)
  • A follow-up no longer searches the web for a name from the chat, and a question about an attached file doesn't search the web first. (#214)
  • Clicking a link to a missing file names the file, instead of "open: exit status 1". (#214)
  • The meru binary is no longer tracked in git. (#216)

Closed issues

Issue Title Closed by
#78 A model router: answer easy questions with the fast model, the rest with the main model PR #217

Pull requests included

PR Title
#218 Keep answer endpoints on this machine or your own network
#217 Let the router pick which model answers
#216 Stop tracking the meru binary, and ignore it
#215 Ask before a turn goes to an answer endpoint off this machine
#214 Answer with a model from another model server, and fixes from test sessions

Contributors


Full changelog: v0.6.4...v0.7.0