aigent-hive

v0.9.5

Published

4 days ago

Provider-neutral local agent harness for subscription-authenticated hosts

0High
0Medium
0Low

gvm1229

Aigent Hive

A provider-neutral local harness for Codex, Claude Code, and Gemini Antigravity.

English · 한국어

Hive gives subscription-authenticated agent hosts one consistent setup, Skill routing, project knowledge, durable role/run state, usage safeguards, and safe update contracts. It never asks for model-provider API keys, calls model-provider APIs, or replaces the host's own model runtime.

Stable 0.9.4 is the current release for this source.

Install the current stable release

0.9.4 is published on npm as latest, with a normal GitHub Release and annotated Git tag.

npm install -g aigent-hive

Or pin the exact version:

npm install -g [email protected]

The npm installer requires Node.js and npm. The installed hive runtime is a native Rust binary and does not require Node.js.

Expected stable version label:

AIgent Hive v0.9.4 (released 2026-08-14)

macOS and Linux with curl

curl --proto '=https' --tlsv1.2 -LsSf \
  https://unpkg.com/[email protected]/install.sh | sh

Windows PowerShell 5.1+

irm https://unpkg.com/[email protected]/install.ps1 | iex

Windows Command Prompt

curl.exe -fLo install-aigent-hive.cmd https://unpkg.com/[email protected]/install.cmd && install-aigent-hive.cmd

The direct installers fetch the same native package bytes from npm, verify the embedded exact-version SHA-256, and record direct-install ownership. They do not require npm, Node.js, or PowerShell 7.

Optional one-prompt setup

If you want Codex, Claude Code, or Gemini Antigravity to guide the entire user-level installation, paste the following prompt instead of manually following the first three setup steps below. It is optional: the four-step setup remains the predictable manual path.

I want the optional one-prompt Aigent Hive setup. Work only at user scope; do not inspect,
initialize, or change any project, repository, folder, or current working directory.

Install the current stable release 0.9.4. The stable install guidance is
https://github.com/gvm1229/aigent-hive#install-the-current-stable-release.
Detect my operating system and active host (Codex, Claude Code, or Gemini Antigravity), asking
me if either is unclear. Check whether Node.js and npm are available. If they are missing,
give me the official OS-specific Node.js installation command and request any approval the host
requires before installing it. Then install the exact Hive release I selected using the official
method in the linked guidance, verify `hive --version`, and activate only my host with
`hive install --scope user --host <detected-host> --apply --output json`.

Then begin interactive global setup in this conversation. For a first setup, ask only whether I
want English or Korean first; continue one question at a time. For existing settings, first ask
whether I want to change one setting or review everything. Do not start project setup afterward:
offer the separate project-setup prompt instead. Never ask for provider API credentials or install
an optional third-party Skill.

This option installs only the current stable release.

Project maintainers can consult the release verification notes.

Supported targets

| Platform | Native target | 0.9.4 gate | | --- | --- | --- | | macOS Apple Silicon | aarch64-apple-darwin | Candidate runtime qualified | | macOS Intel | x86_64-apple-darwin | Candidate runtime qualified | | Linux x86_64 | x86_64-unknown-linux-musl | Release qualification in progress | | Linux arm64 | aarch64-unknown-linux-musl | Release qualification in progress | | Windows x86_64 | x86_64-pc-windows-msvc | Candidate runtime qualified |

Codex and Antigravity have real-host qualification evidence. Claude Code packaging and projection are covered by fixtures, but a real subscription-backed session remains unverified. Stable 0.9.4 uses explicit macOS ad-hoc signing and publishes Windows as unsigned unless free SignPath Foundation signing is approved. See the code signing policy for the exact trust boundary.

First setup

Follow these four steps in order. Repeat step 2 for each host, step 4 for each project, and step 3 whenever global preferences change.

1. Install the Hive CLI

Use one command from Install the current stable release above. The npm installation provides the hive command; it does not yet activate Hive inside a host.

2. Activate Hive for this host

In a terminal, activate the host projection:

hive install --scope user --host codex --apply --output json

Replace codex with claude or antigravity for that host. This operation restores an authenticated known prior user installation before updating it to the current projection; it still refuses unknown or modified ownership manifests.

3. Configure global preferences

Open Codex, Claude Code, or Gemini Antigravity and paste this shared prompt:

Configure or reconfigure my global Aigent Hive preferences for this host. Do not inspect or configure a project, repository, folder, or current working directory. Start the interactive user-scope setup.

Use this prompt for the first setup and later preference changes. It configures only your user-scope language, Wiki, user contexts, persona, Skills, and update preferences; it never inspects the current folder or creates a project harness.

All built-in Skills are active by default. If you prefer a smaller set, choose Skills one by one during setup; setup-hive always remains active. You may select multiple user contexts and add a short description. They help Hive understand you globally, but never choose a project workflow, implementation approach, delivery priority, or active Skill set. Your persona and selected host also never change the active Skill set. Users with an earlier recommended-suite setting keep its exact existing Skill set until they review and approve a new preview.

4. Configure one project

Open the exact project in the host and paste this separate prompt:

Configure the local Aigent Hive harness for this project. Use my existing global Hive preferences, inspect only this project, show the exact write preview, and ask me only about choices that require my approval.

Use this prompt once for each repository. It inherits your global preferences and only changes the named project after showing its exact write preview. If the host is not open in the project, name the project with an absolute path instead:

Configure the local Aigent Hive harness for the project at /absolute/path/to/project. Use my existing global Hive preferences, inspect only that project, show the exact write preview, and ask me only about choices that require my approval.

Do not use the project prompt from your home directory without a project path. A request that includes both scopes completes global setup first and asks before inspecting or changing a project.

Neither prompt authorizes an update, optional third-party Skill, or provider-credential access.

Hive previews its exact owned write set, preserves foreign guidance bytes, and keeps canonical knowledge in Markdown.

Updating

hive update

This performs an immediate version check. If a newer version exists, Hive explains the exact update and asks before invoking the authenticated install owner. Declining, closing stdin, or running non-interactively causes no installation. An existing 0.9.0-test.N, 0.9.0, or 0.9.1 installation keeps its owner evidence and may update to exact stable 0.9.4 through the same confirmation flow.

When daily checks are enabled, a successful check is throttled for 24 hours. An offline or failed check is not recorded as successful, so the next Codex, Claude Code, or Antigravity session retries it.

Hive never installs an update silently.

Move global knowledge to another computer

Use a .hivekb bundle to move global knowledge between computers. --user-root is the user home directory, not its .hive subdirectory. Use the command block for the shell that is running it; do not copy a Windows environment variable into a macOS or Linux shell, or the reverse.

The bundle contains portable canonical knowledge. It does not copy SQLite indexes, runtime state, project-private knowledge, credentials, or absolute local paths.

macOS and Linux shell

Export the global bundle:

bundle="$HOME/Downloads/aigent-hive-knowledge-YYYY-MM-DD.hivekb"
hive knowledge export --user-root "$HOME" --scope global --bundle "$bundle" --output json
shasum -a 256 "$bundle"

On the destination computer, replace the bundle path and expected checksum with the values you received. Run the dry run first. Run --apply only after the file exists, the checksum matches, and the dry run reports no conflict or validation error.

bundle="$HOME/Downloads/aigent-hive-knowledge-YYYY-MM-DD.hivekb"
shasum -a 256 "$bundle"
hive knowledge import --user-root "$HOME" --bundle "$bundle" --dry-run --output json
hive knowledge import --user-root "$HOME" --bundle "$bundle" --apply --output json
hive knowledge lint --target "$HOME" --user-root "$HOME" --output json

Windows PowerShell

Export the global bundle:

$bundle = "$env:USERPROFILE\Downloads\aigent-hive-knowledge-YYYY-MM-DD.hivekb"
hive knowledge export --user-root $env:USERPROFILE --scope global --bundle $bundle --output json
Get-FileHash -Algorithm SHA256 $bundle

On the destination computer, use the returned Hash value to verify the received bundle. Run --apply only after the file exists, the hash matches, and the dry run reports no conflict or validation error.

$bundle = "$env:USERPROFILE\Downloads\aigent-hive-knowledge-YYYY-MM-DD.hivekb"
Get-FileHash -Algorithm SHA256 $bundle
hive knowledge import --user-root $env:USERPROFILE --bundle $bundle --dry-run --output json
hive knowledge import --user-root $env:USERPROFILE --bundle $bundle --apply --output json
hive knowledge lint --target $env:USERPROFILE --user-root $env:USERPROFILE --output json

Automatic dispatch safeguard

When enabled, Hive checks subscription usage immediately before a new automatic dispatch:

hive usage enforce --target <project> --session-id <id> --process-id <pid> --output json
hive run resume --dispatch-intent automatic --target <project> --run <run-id> --capabilities <json> --output json

The first command is only a preflight; it never authorizes dispatch by itself. External runtime cancellation is auxiliary evidence and never replaces durable goal/task state. Ordinary quick-answers and manual work do not run this automatic-dispatch gate.

Load-tested for large knowledge bases

A chunk is a retrieval-sized piece of a Wiki page, not one complete document. To make the scale concrete, the qualification fixture splits 25 deliberately long Wiki pages into 2,000 chunks each: 50,000 search-ready pieces in total. Its portable-bundle scenario also carries a registry of 100 portable collections. This is a deliberately large test fixture, rather than a claim that an ordinary user has 50,000 documents.

| Qualification | Scenario | Latest p95 | Acceptance limit | | --- | --- | ---: | ---: | | Fresh local retrieval | 50,000 chunks | 170 ms | 500 ms | | Warm repeat retrieval | 50,000 chunks | 0.14 ms | 100 ms | | Portable bundle export | 100 collections, 50,000 chunks | 1.04 s | 5 s | | Bundle import and index rebuild | 100 collections, 50,000 chunks | 3.27 s | 15 s |

These release-build, local-SSD measurements are qualification evidence; real timing varies with hardware and the shape of the knowledge. See the qualification record for the method and limits.

What Hive owns

Hive-owned marker blocks and manifest-listed files only
Provider-neutral Skills and thin host projections
Canonical Markdown/YAML/TOML state
Disposable SQLite indexes rebuilt from canonical text
Verified direct-install receipts

Hive does not own provider credentials, model sessions, foreign guidance, OMX/OMC state, Homebrew/WinGet installations, or optional third-party Skills without explicit approval.

Architecture and maintainer docs

Development requires Rust stable, Python 3.13 for conformance tests, and PowerShell 7 for Windows development/release workflows. Consumer installations do not require Python or PowerShell 7.

python scripts/dev-check.py pre-push

License

Apache-2.0. See LICENSE.