AGIW Suite
A monitor for local models
Mac monitor for local model observations, resource readings and deliberate controls.
Apple silicon · macOS 13+1.0.0 (7)
Understand local model state
on your Mac.
Inspect model observations, resource pressure and recorded routes. Check when a reading was collected, keep missing information visible, and deliberately request supported model or recovery actions.
Show what the evidence supports.
Keep unknown and stale state visible.
Inspect the model. Trace the signal. Choose the action.
AGIW Suite brings model details, evidence sources, and deliberate controls into one Mac view.
Inspect
Open the model inspector
See a model’s exact ID, context, quantization, capabilities, loaded instances, source, and observation age when the runtime reports them. Missing fields stay unknown.
Inspect the released source ↗ (opens in a new tab)Browse models with resource context
Core 6 filters to pinned model IDs; All discovered shows the full Mac inventory. Memory pressure, best-effort GPU readings, and a load headroom estimate help you decide what to keep resident.
Read the release details ↗ (opens in a new tab)Diagnose
Check each evidence source
Feeds & evidence labels each source as live, recorded, or stale and shows its detail. Old readings lose their live status, so a quiet graph is not mistaken for a current signal.
See the feed rendering ↗ (opens in a new tab)Follow recorded clients and routes
The map shows client model identities, saved choices, and available route receipts. Their age is visible in the inspector; a recorded choice is not presented as live generation.
Read the map’s explanation ↗ (opens in a new tab)Act
Check readiness, then request recovery
Check readiness inspects route evidence without calling a model. A separate Fix inference action checks configured Mac, Nisi, and Windows states in order; route recovery may send one short Windows probe.
Read the action boundaries ↗ (opens in a new tab)Control a model deliberately
Request load or unload through the local lms CLI and inspect the resulting job state. Automatic unloading is off by default and, when enabled, protects an established resident Nisi pair.
Extend
Explore bundled Nisi
The release includes Nisi 0.2.0’s engine, journal, and CLI. Its native fixed self-check runs offline with zero model calls; the local-model JSON example uses separately configured models.
Read the Nisi scope ↗ (opens in a new tab)Connect optional Jev
A signed helper stores your TypeSafe key in the login Keychain. An explicit connection check sends one fixed synthetic request over HTTPS to test connectivity; coding-task routing is configured separately.
Read the Jev scope ↗ (opens in a new tab)Follow the signals.
Choose the action.
AGIW makes local evidence visible and exposes deliberate controls. These diagrams show how observations reach the app and how a model action travels back to the local tools.
01 / Observation & control
Conceptual flow- Source
Local evidence
LM Studio inventory and reported activity.
Optional input: recorded route receipts - Collect
Local observer
Collects bounded observations and serves local snapshots.
127.0.0.1 - Inspect
AGIW dashboard
Shows model state, evidence sources, and observation age.
Swift + WebKit
- Choose
Your request
Select a model and request load or unload.
User action - Check
Observer boundary
Checks Host, Origin, and action input size and shape.
Request validation - Execute
Local model tools
The local CLI handles the requested model operation.
lms · job state
Inspect the result. The app exposes the action’s job state including whether it completed or failed.
02 / Read the state correctly
Evidence guideA model is loaded
It is resident in memory. Request activity needs its own evidence.
Fresh, validated
busy signal?
Observed activity
Read the reported work and its evidence source.
Activity unknown
A loaded model alone cannot establish active generation.
From workflow to working tools.
The coding workflow provides the context; Nisi and AGIW have their own public releases. Dates below mark those releases.
- Workflow
Initiate online code mode
A local coding setup for checking available routes and coordinating bounded work. It supplies the workflow context for the two software projects.
Workflow context · separately configured Nisi 0.2.0
The public Node.js CLI and library give an application a bounded draft, check, test, review, and repair workflow. The CLI includes fixed demonstrations.
Explore the Nisi case studyAGIW Suite 1.0.0 (7)
The Mac monitor makes local model and route evidence visible, provides explicit model controls, and includes Nisi’s fixed offline self-check.
View the AGIW release ↗ (opens in a new tab)
Historical inference records
Read the work behind the numbers.
These examples use different workloads and timing definitions. Read each date and method with its result. They do not measure an AGIW speed improvement or a Mac-versus-PC ranking.
Median effective completion rate for seven completed requests in each task. Completion tokens are divided by the entire streaming request duration, including time spent waiting or reasoning.
- Generation72.302 tokens/s
- Coding72.842 tokens/s
- Reasoning72.240 tokens/s
72.482 tokens/s overall median across all 21 requests. One warmup was excluded; output correctness was not graded.
See all 21 trial values → (opens in a new tab)Server-reported decoding rate for one fixed JavaScript task per layout. Each request recorded 106 prompt tokens and 141 completion tokens; the backend, context capacity, and offload settings varied.
- A Existing layout18.5 tokens/s
- B GPT-OSS on CPU · earlier pass12.3 tokens/s
- C GPT-OSS on CUDA111.4 tokens/s
- D CUDA; Qwen on CPU106.0 tokens/s
A, C, and D came from one pass; B came from an earlier pass. One request per layout gives no repeated-trial uncertainty estimate.
See original layout chart → (opens in a new tab)Clear boundaries for local work.
The app keeps the dashboard, local observer, model services, and optional integrations in separate components. Here is where they connect and which controls apply.
Swift + WebKit
Menu-bar shell and bundled dashboard. The browser view does not hold the optional Jev key.
Python observer
Serves snapshots on 127.0.0.1. Checks the request’s Host and Origin, then validates the size and shape of action inputs.
- Configured locally
LM Studio +
lmsInventory and activity evidence enter the observer; model load or unload needs an explicit request.
- Bundled package
Nisi 0.2.0
An external Node runtime runs the fixed offline self-check. It makes zero model calls and does not run arbitrary repository work.
- Optional · opt-in
Signed Jev helper
Optional service credentials stay in the login Keychain. Its connection check uses a fixed HTTPS request; task routing is configured separately.
- Separately managed
Router + PC workers
The monitor reads bounded observations and offers supported guarded recovery. The external systems own task assignment and execution.
What the controls establish
Loopback binding, request checks, and input limits constrain browser-origin actions. Local processes remain a separate trust boundary. The public release is Developer ID signed and Apple-notarized; those checks verify distribution and do not certify the entire system’s security.
Download AGIW for Mac.
The current public download is on GitHub for Apple silicon Macs running macOS 13 or later. The release page includes a Developer ID signed, Apple-notarized ZIP, a checksum file, and installation instructions.
Connect an agent.
AGIW MCP gives a coding assistant a readiness check, a Nisi workflow on your local models, and optional advisory Jev checks before and after a task.
The setup guide describes unreleased changes and instructs installation from the reviewed source checkout. Its package version remains 0.1.0; that version alone does not identify these changes. Install the pinned checkout below, then register it with your assistant. For Claude Code:
git clone https://github.com/louiscalata/agiw-mcp.git
cd agiw-mcp
git checkout --detach 7fa6d53821aa8bec755351f954eb0a676541bfc0
pipx install .
claude mcp add agiw-mcp -- agiw-mcpConfigure a local OpenAI-compatible model server and two different models for Nisi: one author and one reviewer. Follow the pinned source setup guide, then ask your assistant to “initiate online code mode.” READY covers the local prerequisites. Optional Jev readiness is reported separately in jev.status (READY or NOT_READY) with its consent and key checks. agiw_nisi_run works without Jev; key lookup is skipped until consent is recorded. Record the checkout commit and any local diff when reporting issues.
Jev is optional and needs your own TypeSafe key. Before using it, inspect the disclosure with agiw-mcp consent show; you grant consent yourself in a terminal with agiw-mcp consent grant. Your assistant cannot grant it. Jev sends the supplied task or report text to TypeSafe, so leave secrets out.
An earlier view of AGIW.
These screenshots, taken on 9 October 2026, document TestFlight build 0.1.0 (22). That build has a different, older interface than the downloadable AGIW Suite 1.0.0 (7) build.
There is no App Store release. The GitHub release (opens in a new tab) is the current public download.
View the archived TestFlight screenshots
(opens in a new tab)
(opens in a new tab)
(opens in a new tab)A local coding advisory form.
The earlier build included a form for requesting a bounded proposal from a selected local LM Studio model through Nisi. This screenshot shows the setup form before any task was submitted. It does not demonstrate code execution, completed checks, or a feature of the public GitHub build.
(opens in a new tab)A monitor and a workflow engine.
AGIW shows the state of local model services on a Mac. Nisi coordinates bounded coding work through adapters supplied by its host application. Each project has its own public GitHub release.
