AMOS Desktop

The open-source native operator for the governed AMOS company brain.

AMOS Desktop lets a person choose where intelligence runs, connect it to the company context and capabilities in AMOS, and safely work across both company systems and an explicitly selected local workspace.

The first run offers three honest starting points:

• My workspace — no AMOS account required; use a provider key or local model for private code, research, documents, and local automation.
• Northwind demo — a short-lived sample company using the real AMOS tool, policy, approval, receipt, and hosted-intelligence boundaries.
• My company — sign in or create an AMOS account for durable organizational memory, connected applications, shared authority, approvals, and proof.

Download for Apple Silicon · Download for Intel Mac · Download for Windows · AMOS Labs · Managed platform

Why it exists

General-purpose AI is powerful, but a company needs more than a model:

• durable context that survives a chat or model change;
• secure connections to the applications where work actually happens;
• identity, tenant boundaries, roles, and policy at the point of action;
• human decisions for consequential work;
• receipts showing what changed, why, and who approved it; and
• a local surface for documents, code, files, and private work.

AMOS Desktop supplies the native execution and interaction layer. The AMOS managed platform supplies the governed company brain. Models remain interchangeable.

AMOS Desktop
  chosen intelligence
  local workspace + private task context
  documents, screenshots, code, and tools
               │
               │ OAuth + tenant-scoped MCP
               ▼
AMOS managed platform
  company memory + connected applications
  engines + automations + goals
  policy + approvals + proof receipts

What works today

• Native macOS applications for Apple Silicon and Intel, plus Windows 10/11 x64
• Browser-based AMOS OAuth 2.1 + PKCE sign-in
• Signed-in company, role, and effective-scope identity
• AMOS engine discovery through a compact MCP tool surface
• Documents, source files, drag/drop, and pasted screenshots
• Local PDF, DOCX, text, and source extraction
• Explicit Use for this task, Keep in private memory, or Add to company memory handling
• Operating-system-protected private memory with reuse, promotion, and permanent forget controls
• Passphrase-encrypted .amos-memory export/import with preview, tamper detection, deduplication, and fork lineage
• Explicit four-hour, server-signed company briefings for read-only offline work
• Encrypted offline outcome drafts with live diff, identity pinning, and explicit reauthorization
• Typed company canvases for metrics, tables, trends, briefs, evidence, approvals, and receipts
• Local workspace grants, search, reads, Git status/diff, and atomic patches
• A bounded project briefing with detected stack, manifests, scripts, Git state, README context, verification commands, and safe suggested next tasks
• Approval-gated shell commands and file changes
• Digest-addressed local task receipts for completed, failed, and canceled work
• Streamed responses with visible planning, action, and evaluation phases
• Productive tasks continue until completion, safe cancellation, or a no-progress safeguard
• Live user steering that joins the same task at the next safe model/tool boundary
• Safe cancellation across model, MCP, web, and local process work
• Encrypted restart checkpoints with identity, company-context, and approval revalidation
• Visible, skill-backed workflows with task-specific verification criteria
• Live work, decisions, approval notifications, activity, and proof
• AMOS-hosted, customer-cloud, provider API, and local-model profiles
• Signed macOS and Windows releases with in-app update notifications

AMOS Desktop is not a second CRM, integration vault, or company database. Shared business state, integrations, governance, and receipts remain in AMOS.

Install

1. Download the correct installer from the links above.
2. On macOS, drag AMOS Desktop into Applications. On Windows, run the signed per-user installer; it adds Start-menu and desktop shortcuts without requiring an administrator account.
3. Open it and choose My workspace, Northwind demo, or My company.
4. For personal work, select a local runtime, provider API, Bedrock, or another compatible endpoint. No AMOS account or company tools are required.
5. For the demo, the browser creates a bounded sample company and securely returns it to Desktop without placing its short-lived key in a URL.
6. For your company, sign in and use AMOS Hosted immediately—or choose another intelligence profile.
7. Grant only the local workspace you want AMOS to inspect and operate.

The app checks for signed updates after launch and every six hours. It notifies you before downloading and never restarts during an active task.

See Desktop installation and releases for packaging, signing, update, and troubleshooting details. See skills and workflows for the engine/skill/workflow contract and tenant-extension safety model.

Choose where intelligence runs

AMOS Desktop uses a provider-neutral, OpenAI-compatible model boundary. Supported profiles include:

• AMOS Intelligence (default) — zero-config AMOS-managed, capability-routed inference; included plan credits apply first and additional usage is metered
• Amazon Bedrock — customer- or AMOS-controlled AWS inference
• Compatible endpoint — a customer-controlled HTTPS endpoint
• Provider API — including the Moonshot/Kimi API
• AMOS Local — a bundled, managed local runtime with hardware-aware models
• Compatible local runtime — an explicitly customer-managed Ollama or llama.cpp endpoint

Changing the intelligence does not reconnect the company or change the user's AMOS authority. Models without reliable tool use should be limited to observe-and-draft workflows.

AMOS Intelligence uses the same short-lived AMOS identity as the company connection. Desktop requests the stable auto model alias, while the managed platform owns provider/model routing. Desktop exposes Efficient, Balanced, Deep, and Frontier capability profiles rather than leaking the current implementation model. That lets AMOS move from Bedrock to an AMOS-hosted model or route by workload without shipping a new desktop build. Explicit provider keys, private compatible endpoints, customer Bedrock, and local models remain available and are never overwritten after the user selects them.

AMOS Local assesses the computer, recommends a curated profile, shows resumable model-download progress, supports explicit removal, and can activate a visibly separate local-only operating mode. Signed Desktop releases include and supervise the pinned runtime on loopback; users do not install or operate a second application, and model weights remain installed across AMOS updates. In local-only mode AMOS and public-web tools are not exposed to the model. Desktop follows the operating system's light or dark appearance by default, with an immediate header switch and persistent overrides.

AMOS Local models

AMOS Local is the governed local-runtime experience, not a claim that AMOS trained the underlying open-weight models. The signed catalog currently contains:

AMOS profile	Underlying model	Approximate download	Recommended memory	Best fit
Compact	qwen3:4b	2.6 GB	12 GB	Fast summaries, drafting, extraction, and lightweight workspace tasks
Balanced	qwen3:8b	5.2 GB	16 GB	Stronger reasoning, coding, documents, and everyday local tool use
Capable	gpt-oss:20b	14 GB	24 GB	Primary interactive text, coding, retrieval, and tool work
Vision	qwen3.6:27b-q4_K_M	17 GB	32 GB	Secondary multimodal profile for tasks that contain images
Vision Max	qwen3.6:27b-q8_0	30 GB	64 GB	Experimental higher-precision multimodal profile

On first launch, Desktop starts the included, checksum-verified runtime on an AMOS-owned loopback port and disables its cloud features. The user chooses a model and selects Install; Desktop handles the resumable download, activation, lifecycle, and persistent model directory. The model never receives AMOS company or public-web tools in local-only mode. A catalog change requires a new signed AMOS Desktop release—it cannot be introduced by model output. AMOS selects a default local context from installed memory: 16K below 16 GB, 32K on 16 GB systems, 64K on 32 GB systems, and 128K on 64 GB systems. Advanced users can explicitly raise the bounded limit to 262K on a model and machine that have been measured at that window. The context compiler still targets a smaller information-dense working set for routine turns.

The catalog's recommendation is evidence-driven rather than “largest model that fits.” On the 64 GB qualification machine, GPT-OSS 20B matched both Qwen 27B quantizations on the hard AMOS suite while running more than five times faster. Qwen Q4 is recommended only when an image requires a multimodal model; Qwen Q8 remains available for experiments but showed no quality gain on the current suite. See the qualification report.

For work beyond the laptop's dependable capabilities, the user can switch to AMOS Intelligence, Bedrock, a provider API, or a compatible private endpoint without reconnecting the company.

AMOS Desktop 0.9 adds an explicit offline company-context grant. While online, a signed-in user can ask AMOS for a short-lived, server-signed copy of the already bounded resume_company briefing. Desktop verifies the signature and current identity, encrypts the grant with operating-system protection, and exposes one read-only tool in local-only mode. It never carries credentials, write authority, approvals, or permission to replay work.

AMOS Desktop 0.10 adds outcome-level offline drafts. A local model can stage human-readable proposed work, but it cannot queue AMOS tool arguments or actions. On reconnect, Desktop checks the exact user and tenant, compares the captured section fingerprints with a fresh read-only company briefing, and shows any drift in Decisions. Continuing only fills the online Operator composer; the user must still press Run, and current AMOS policy, approvals, and receipts govern everything that follows.

AMOS Desktop 0.11 adds a durable task lifecycle. OpenAI-compatible model output streams into Operator while the Live Work panel shows bounded phases and tool progress. Stop safely aborts the active model request, MCP call, web fetch, or local process tree. Signed-in company tasks receive an operating-system encrypted checkpoint. After a restart or cancellation, Desktop rechecks the exact user and tenant, fetches a fresh resume_company briefing and approval queue, and loads a no-replay continuation into Operator for explicit review. Completed tasks remove their checkpoint.

AMOS Desktop 0.13 makes AMOS Intelligence the zero-config intelligence path. Signing into AMOS supplies short-lived inference authorization automatically; included plan credits apply first and additional usage is metered to the company. The client requests only the stable auto alias so provider/model routing can evolve server-side. Existing BYOK, private endpoint, Bedrock, and local profiles remain explicit alternatives.

New installations default to the Balanced profile rather than maximum reasoning. AMOS Intelligence accepts the stable auto alias and a capability hint, while the managed platform owns the actual routine, balanced, deep, or frontier route. Users can still choose an exact model through a customer Bedrock or provider-key profile. This keeps routine coding, summarization, extraction, and tool work from paying frontier cost by default.

See Intelligence providers.

Security model

The desktop renderer is sandboxed. Provider secrets are encrypted with the operating system's secure storage. Local file tools remain inside the selected workspace, common credential files are blocked, and child commands receive a scrubbed environment.

Company actions always pass through AMOS identity, tenant isolation, RBAC, operation policy, approvals, and proof. A local model cannot weaken those server-side controls.

Read Safety, Authentication, and Security policy before extending local authority.

Build from source

Node.js 22 or newer is required. Node.js 24 LTS is recommended.

git clone https://github.com/amos-labs/amos-agent.git
cd amos-agent
npm install
npm test
npm run desktop

Create an unpacked local application:

npm run desktop:dir

Create unsigned local DMG/ZIP artifacts:

npm run desktop:build

Create an unsigned local Windows x64 installer:

npm run desktop:build:win

Official macOS releases are signed and notarized, and official Windows releases are Authenticode-signed. GitHub Actions publishes both platforms together only after every installer and update manifest passes its platform gate. Local development builds do not contact the production update feed.

CLI for developers and automation

The desktop application is the primary end-user experience. The same runtime is also available as the amos-agent CLI for development, CI, and controlled automation:

npm link
amos-agent login
amos-agent status
amos-agent --cwd /path/to/project

Run one task:

amos-agent --cwd /path/to/project \
  "Resume my company context, inspect this repository, and explain what should happen next."

Scoped AMOS_API_KEY credentials are supported for CI and unattended agent identities. Human users should use OAuth.

The example environment file documents provider and advanced CLI configuration: .env.example.

Built-in local tools

• list_files, read_file, and write_file
• search_files
• desktop_inspect_project
• git_status and git_diff
• apply_patch
• run_bash
• web_fetch and optional web_search

AMOS tools begin with a compact bootstrap:

• amos_get_started
• amos_whoami
• amos_resume_company
• amos_company_overview
• amos_list_engines
• amos_load_engine_tools
• amos_call_engine_tool

Engine-specific tools are loaded only when needed, keeping model context small as the AMOS platform grows.

Project map

• Desktop product and release guide
• Architecture and system boundaries
• Authentication
• Intelligence and infrastructure profiles
• Safety model
• ExpertCache research boundary
• Memory classes and private-memory controls
• Signed offline company context
• Offline draft reconciliation
• Streaming, cancellation, and restart-safe tasks
• Typed company canvas
• Canvas, offline intelligence, and portable memory proposal
• Contributing

Direction

The near-term product path is:

1. signed device identity and policy-controlled environment grants;
2. richer typed canvases and managed result adapters for company data and active work;
3. richer private-memory retrieval and portable continuity;
4. enterprise deployment and fleet-management controls; and
5. Windows on Arm and managed-store distribution when customer demand warrants them.

The governing principle is constant: local intelligence may observe, reason, draft, and execute within explicit grants; AMOS remains authoritative for shared company memory, policy, approvals, and proof.

License

Apache License 2.0. See LICENSE.