Skip to content

MiderHive 1.2.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 18 Sep 08:25
v1.2.0
4bad258

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

1.2.0 is the "connect anything, and make the memory real" release. Eight AI coding tools now
connect in one click through a single integration registry, the MCP server grew from 18 to 27
tools so an MCP-only agent can do everything an HTTP client can, the workbench is fully bilingual,
and — the big one — the knowledge base can finally use real semantic embeddings from a model you
already run, instead of the built-in n-gram matcher. Alongside that: a batch of hardening fixes found
by fuzzing, measurement and adversarial review, with the numbers published.

⚠️ Read this before running

  • This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
    warning on first launch. Choose More info → Run anyway. If you would rather not, try the
    portable ZIP first — it needs no installation.
  • Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need the
    VC++ Redistributable installed.

What's new in 1.2.0

Real semantic search — bring your own embedding model

The built-in embedder is an offline n-gram matcher: good for recall, not true semantics. It
still works out of the box (and still needs no model bundled), but you can now point the CLI at an
OpenAI-compatible embedding endpoint you already run — Ollama, LM Studio, vLLM, a gateway:

agent-cli knowledge add --title "…" --content "…" \
    --embed-url http://127.0.0.1:11434/v1/embeddings --model nomic-embed-text
agent-cli knowledge search --q "…" --mode semantic \
    --embed-url http://127.0.0.1:11434/v1/embeddings --model nomic-embed-text
  • agent-cli embed is there to test the endpoint first; MIDERHIVE_EMBED_KEY carries a bearer
    token if your endpoint needs one (it stays out of your shell history).
  • The request leaves the CLI you ran, never the service. MiderHive itself still makes exactly
    one outbound call (the daily update check); no background code path can reach the network.
  • MCP-only agents get the same power without installing the CLI: knowledge_add and
    knowledge_add_version accept an embedding (64..4096 numbers) plus the embedder name.
    In 1.1.x only knowledge_search accepted a vector — an MCP agent could search a custom
    dimension it had no way to write into. That asymmetry is gone.
  • Failures are loud, not silent: https (this build has no TLS), an unreachable host, a non-2xx or
    non-JSON response, or an out-of-range dimension all fail with a readable reason. Nothing silently
    falls back to the built-in embedder — a silent fallback would make "why is semantic search wrong?"
    impossible to answer.
  • One dimension is bound to one model (vectors of the same width share a table), so a second model
    needs another dimension. Trying to mix them is rejected with a message naming the bound provider.

One-click connect for eight tools

  • Connect wizard (also Settings → Agents): pick a tool, click once. It provisions the
    identity, emits that tool's own config format — JSON, Codex TOML, DSH profile patch, Hermes
    YAML, environment variables, or an instructions block — and, where the file can be merged safely,
    writes it for you, backing the original up as .miderhive.bak. A live "Connected" appears
    once the agent actually comes online.
  • Eight clients from one registry: Claude Code, ChatGPT / Codex, Factory Droid, DSH, Hermes,
    ZCode, Cursor and GitHub Copilot — with per-client install detection and real config-path
    resolution. Adding a ninth tool is a data change, not a code change.
  • Codex TOML was broken for real Windows paths: it emitted double-quoted strings, where \Q is
    an invalid escape, so the whole config.toml failed to parse. It now emits TOML literal
    (single-quoted) strings.
  • apply-config refuses what it cannot merge safely: configs using tab indentation or YAML
    block scalars (key: |) are declined with a paste-it-yourself hint instead of risking a corrupted
    file — and the generated file is replaced by rename, so an interrupted write can no longer leave
    your config missing.

27 MCP tools, so MCP-only agents are first-class

Nine more tools close loops that previously required raw HTTP: knowledge_add_version, knowledge_get,
knowledge_versions, skill_register, message_reply, usage_report (idempotent), usage_daily,
usage_breakdown, and hive_status — one call that answers "am I connected, who else is here, what
did we spend".

Bilingual workbench

Qt's own translation catalog is now installed, so Qt-supplied strings (dialog buttons, input context
menus) follow the UI language instead of staying English in Chinese mode. A regression test switches
the live window to English and asserts no Chinese text remains in any widget — labels, buttons,
combos, headers, placeholders and tooltips.

Hardening and reliability

Found by fuzzing, measurement and adversarial review rather than by users:

  • A dimension mismatch in semantic search used to return an empty result, silently. It now fails
    with an error listing which dimensions (and providers) the database actually holds.
  • knowledge remove left orphan vectors behind — entries whose versions had different dimensions
    kept rows in the other per-dimension tables. Deletion now covers every dimension of that uuid.
  • Opening a newer database — or restoring a newer backup — is refused instead of being silently
    half-understood, with a message telling you to upgrade MiderHive or restore an older backup.
  • MCP same-name registration race: when two agents of the same name started together, the loser
    exited. It now re-reads agents.json and reuses the key.
  • knowledge/search clamps limit to 1,000; previously an unbounded value let one request assemble
    an arbitrarily large result set.
  • Semantic recall's k cap moved from 10,000 to 50,000 — on sparse-tag large databases the old cap
    systematically under-returned results.
  • The full-text index now covers latest versions only, so index size no longer grows with version
    history.

Verification

  • 1,661 unit-test checks across 532 assertion sites (about 1,000 of those checks are one
    assertion re-run to seed 1,001 messages for the pagination test — the site count is the honest
    measure).
  • 44 integration assertions over real HTTP: multi-agent lifecycle, Chinese retrieval, the async
    task state machine, idempotent reporting, usage alerts.
  • 59 MCP assertions driving miderhive-mcp.exe as a real stdio client, including the new
    write-with-your-own-vector round trip at 256 dimensions.
  • 72 offscreen GUI assertions driving the real workbench headlessly.
  • 39 config-merge fuzz cases, plus the per-client connect matrix and the external-embedding
    round-trip — all three now run in CI on every push (they previously ran only when someone
    remembered, which is how the tab/block-scalar bug survived).
  • Measured, not assumed: published concurrency numbers (docs/hardening-report.md) show the
    long-suspected global lock is not a bottleneck — mixed load scales 3.10× at four threads, pure
    reads 9.47×; pure writes scale 1.14× because every commit fsyncs, which a lock-free SQLite
    measurement proves is the disk, not the lock.

Which file should I download?

File Best for Notes
MiderHive-1.2.0-x64.msi A normal install Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features.
MiderHive-1.2.0-win64-portable.zip No installation Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt Verifying integrity See the verification section below.

Install / run

  • Installer: double-click and follow the wizard, or install silently with
    msiexec /i MiderHive-1.2.0-x64.msi /qn
  • Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with the
MIDERHIVE_PORT environment variable). Settings let you switch the theme, language, font size,
refresh interval, and take backups.

Upgrading

  • From any earlier version: install this MSI over the old one — the same upgrade identity is
    used, so the previous version is removed first and nothing is left behind. Your data directory is
    kept as-is.
  • First launch after upgrading does a one-time index rebuild (the database schema moves to
    version 2): knowledge vectors start recording their dimension, and the full-text index is rebuilt
    to cover latest versions only. On a large knowledge base this takes a moment; it happens once.
    If the rebuild is ever interrupted, close and reopen the app.
  • Databases written by a newer MiderHive are refused with a clear message rather than opened
    half-understood. Backups made by a newer version are refused on restore for the same reason.
  • If a shortcut still shows an old icon after upgrading, that is the Windows icon cache: run
    ie4uinit.exe -show, or restart Explorer.
  • The built-in updater will also offer 1.2.0 by itself (checked at most once a day; the download is
    SHA256-verified before it runs).

Verify the download (optional)

Get-FileHash .\MiderHive-1.2.0-x64.msi -Algorithm SHA256
# compare with the matching line in SHA256SUMS.txt

Feedback

Issues and ideas: https://github.com/SiliconCoderJames/MiderHive/issues

Assets 6
Loading