Ódinn Forge

Your AI assistant, on your computer, under your control.

Ódinn Forge is an AI assistant that runs on a computer you control. It can remember useful context, research the web, use approved tools, and help you carry work across multiple conversations.

It is built for everyday users, business owners, and independent professionals who want more than a chatbot—but do not want an AI quietly clicking buttons, changing accounts, or hiding what it did.

Odinn Forge v1 supports the local, single-user workflow described in the v1 compatibility policy. Core advanced services, optional plugin modules, third-party packages, and remote multi-user hosting have separate support boundaries.

What can I use it for?

Everyday work

• Research a topic and turn the results into a useful summary.
• Keep separate conversations, saved details, and goals for different projects.
• Remember preferences and project context you choose to save.
• Draft plans, documents, checklists, and communications.
• Use the web without mixing its browser activity into your personal browser profile.
• Review a clear history of tool use and important decisions.
• Inspect workspace directories, metadata, text, literal matches, and diffs through bounded read-only tools with sensitive-file filtering.

Small business owners and independent professionals

• Research competitors, suppliers, products, or market changes.
• Keep client and internal projects separated inside one workspace.
• Preserve useful business context between work sessions.
• Schedule repeatable tool-based tasks.
• Choose the AI service that fits your budget and privacy needs.
• Keep an activity trail so you can see what ran, when it ran, and whether it succeeded.

Discord and Telegram are supported local single-user integration slices with environment-only credentials, explicit inbound policies, safe diagnostics, durable session binding, and no-replay recovery. Their live services remain provider-dependent. Slack, Microsoft Teams, WhatsApp Business, and the shared channel plugin interface remain experimental; see docs/channels.md.

An optional read-only Microsoft Graph slice can inspect one explicitly configured Microsoft 365 account's email and calendar data. It is disabled by default, local single-user only, fixed-origin, capability-gated, bounded, and content-free in durable evidence. It cannot send mail or mutate calendars; see docs/microsoft-graph-read.md.

Ódinn is not a finished industry-specific business suite. It is a general assistant you can connect to the models and tools that make sense for your work.

Why use Ódinn instead of a regular chatbot?

Your work stays organized

Ódinn groups conversations, goals, memories, and activity by project. You can return later without rebuilding all the context from scratch.

You choose the AI

Use OpenAI / ChatGPT, OpenRouter, a local model through Ollama, or another compatible AI service. You are not locked into one model company.

OpenAI, OpenRouter, and Ollama are tested as primary connections. Other listed services may depend on shared compatibility support or on features controlled by that provider. See AI provider support for the plain-language labels shown during setup.

Important actions are visible

Reading a public webpage is different from clicking Buy, sending a form, or changing an account. Ódinn treats those actions differently and asks for approval before browser actions that can change something outside your computer.

Memory is inspectable

Ódinn can suggest useful details to remember, but you decide what to keep. You can review, edit, scope, or forget saved memories instead of trusting a hidden profile you cannot inspect.

There is a record of what happened

Tool runs, approvals, and important system events are recorded locally. This makes it easier to understand failures, review past work, and verify that the assistant followed the expected path.

Quick start

What you need

• Linux, macOS, or Windows
• A supported x64 machine for the normal standalone download. Node.js is embedded and verified; do not install it separately.
• For arm64 or another unsupported standalone target, use the advanced generic package with Node.js 24 or newer.
• An account with a supported AI provider, or a local AI model through Ollama

Install a release

Download the latest standalone package for Linux x64, macOS x64, or Windows x64 from GitHub Releases, then follow the platform instructions in the user guide. Standalone release packages contain the built application, runtime dependencies, and a verified Node.js runtime; normal installation does not require Node.js, pnpm, or a source checkout.

After installation, run:

odinn onboard

Setup will help you:

1. Choose an AI provider and model.
2. Sign in or connect a local model.
3. Review what the assistant is allowed to do.
4. Test a real AI response before saving the setup.
5. Open the local console in your browser.

To open Ódinn again later:

odinn start

To check releases and protect local state:

odinn update check
odinn state status
odinn backup

Odinn checks an update before installing it, keeps the previous application available when possible, and will not roll back if doing so could damage newer saved data. Uninstall keeps your saved state unless you explicitly ask to remove it. See the user guide.

The console normally opens at http://127.0.0.1:18790/. That address points to your own computer, not a public website.

Already use OpenClaw or Hermes? Setup can detect compatible installations and offer to copy supported sign-in details or state. It does not change the original installation.

Run from source

If you are a developer or want to try the current repository checkout:

corepack enable
pnpm install
pnpm check
pnpm odinn onboard

See Getting started for local-model setup, scripted installation, headless systems, and troubleshooting.

Privacy in plain language

Ódinn enables no product telemetry by default. Its settings, memories, conversations, browser profile, and activity records stay in its local state folder by default. Operators may explicitly configure the Gateway's optional, fixed content-free operational telemetry with ODINN_OTLP_ENDPOINT; status and diagnostics expose only bounded local counters and categorical state.

Local-first does not mean nothing ever leaves your computer:

• If you use a cloud AI provider, that provider receives the prompts and context needed to answer you.
• If Ódinn visits a website, that website receives normal web traffic.
• If you enable third-party extensions or tools, they may have their own data practices.
• A local model can keep AI requests on your machine, but web activity still reaches the websites you ask Ódinn to visit.

Choose providers and tools that are appropriate for the information you handle. Do not give any AI system secrets or sensitive client data unless you understand where that data will go.

Safety and control

Ódinn is designed to make consequential choices explicit:

• It uses a separate browser profile instead of silently taking cookies from your everyday browser.
• Browser clicks, typing, and keypresses require approval by default.
• New agent, skill, and extension packages start disabled.
• Setup changes are reviewed, tested, and backed up before replacing a working configuration.
• Built-in automatic improvements are limited to reversible reliability settings. They cannot rewrite the application, disable safeguards, change credentials, or grant themselves new permissions.

You can weaken some of these controls, but Ódinn makes that an explicit choice. Read SECURITY.md before enabling remote access, relaxing approvals, allowing private-network access, or installing third-party code.

Honest limits

• It is not a safety-critical system and should not make medical, legal, financial, or other high-stakes decisions for you.
• Its worker processes help contain crashes; they are not a security sandbox for hostile code.
• It cannot guarantee that an action on an outside website can be undone.
• AI responses can be wrong, and cloud services can be unavailable, rate limited, or changed by their providers.
• The normal local console is for one person on one computer. Do not expose it directly to the public internet.
• Multi-user hosting is available for experienced operators, but people who do not trust one another should use separate operating-system accounts, containers, or machines.

The surface matrix distinguishes stable v1 interfaces, internal implementation details, experimental interfaces, provider-dependent behavior, platform-dependent behavior, and unsupported behavior.

Advanced features

Ódinn includes core advanced services for deeper control and inspection:

• Runemark — run verification checks whether a run met specific acceptance rules.
• Gatewatch — policy safety previews capability intersection, commands, file locations, and approval requirements without executing the operation.
• Norn Restore — restore points saves selected local files and previews a restore before applying it.
• Raven Route — model routing chooses among configured models using recorded reliability, speed, cost, policy, rollback, and verification results.

The following user-facing features are optional plugin modules and remain disabled until enabled individually:

• Rune Key — scoped temporary access gives one narrowly defined permission to one run.
• Saga Archive — portable run bundles exports a redacted, verifiable copy of a run.
• Worldtree Paths — scenario comparison compares work in separate copies of a workspace.

These tools improve visibility and recovery for local work. They do not make outside actions perfectly reversible, and copied workspaces are not security sandboxes.

Ódinn also includes opt-in runtime foundations that remain experimental unless the surface matrix says otherwise:

• Bounded agent-run graphs can dispatch up to eight read-only child nodes, with at most four concurrent, through the durable jobs boundary. Child authority is intersected with the parent grants, and durable projections retain digests rather than prompts or model output. See agent manifests and run graphs.
• Skills, MCP, durable workflows, event ingress, and project context are separately gated runtime foundations. They remain disabled by default or require explicit operator activation and do not enlarge the stable local single-user promise.
• Host-capability plugin seams govern paired computer.screen and provider-injected read-only email/calendar integrations. The current foundation does not bundle ambient desktop access or desktop mutation. The optional Microsoft Graph adapter supplies the first concrete read-only email/calendar slice. See the plugin system notes.
• Authenticated remote-node reads expose only node.status and node.diagnostics for exact operator-allowlisted HTTPS nodes. Model input cannot choose routing or credentials, and durable evidence retains only digests, counts, and fixed statuses. See remote-node reads.

Use Advanced in the console or read the notes under docs/features. Core placement does not make an advanced API a stable public SDK; the compatibility policy remains authoritative. Existing CLI commands, configuration keys, routes, event names, and SDK exports retain their original technical identifiers for compatibility.

Useful commands

odinn status          # Check the current setup
odinn onboard         # Set up, repair, or change the AI connection
odinn start           # Open the local console
odinn sessions        # List saved conversations
odinn runs            # Show recent tool runs
odinn audit verify    # Check the local activity record
odinn doctor          # Create a safe diagnostic summary

Audit migration, cursor streaming, verification, archive, retention, and soak procedures are documented in Audit storage operations.

Documentation

• Documentation hub — the complete navigation index for guides, references, architecture notes, policies, and project records
• Latest release — download the current supported release
• v1 compatibility policy — the stable product promise and its boundaries
• User guide — installation, privacy, and bug reports
• Getting started — setup and troubleshooting
• Operator console — projects, tasks, memory, and scheduled jobs
• Security guide — safe operation and vulnerability reporting
• Capability boundaries — what is tested, experimental, provider-dependent, or unsupported
• Interface reference — CLI and authenticated loopback gateway inputs and outputs
• Comparative evaluation — benchmark methodology, runtime and model comparisons, and performance reports
• Features index — core advanced services and optional plugin modules
• Report a bug or request a feature
• Contributing guide — development setup, pull requests, validation, and releases

The primary user-facing external interfaces are the documented CLI and authenticated loopback gateway. The interface reference describes their inputs and outputs. The surface matrix identifies stable, experimental, provider-dependent, platform-dependent, internal, and unsupported surfaces.

Benchmark snapshot

The external BlueDot agent-benchmarks harness checks artifacts and deterministic assertions rather than treating a plausible final message as proof. The dated snapshot below is informational: it is not a CI gate, a release guarantee, or a model-only score.

2026-08-26 — weekly GitHub Actions comparison. Seven deterministic cases ran five times per runtime with openai-oauth, gpt-5.6-luna, cloud deployment, runtime-default sampling, and the same bounded-filesystem-and-process policy.

Runtime	Verified	Failed	Verified rate	p50	p95
Ódinn Forge 1.1.1	20/35	0	57.1%	3.80s	19.07s
OpenClaw OpenClaw 2026.7.1-2 (0790d9f)	23/35	12	65.7%	26.75s	45.01s
Hermes Agent Hermes Agent v0.19.0 (2026.7.20)	35/35	0	100%	15.98s	23.98s

Ódinn Forge's one-shot agent adapter covers 4/7 cases: 20 supported trials execute and verify, while 15 process-dependent trials are unsupported. Its documented process.exec surface is separately approved, top-level durable /jobs work and is intentionally outside this adapter. Unsupported trials are not failures and remain in the denominator.

Harness: 8ad91f0b9e39. GitHub Actions run. Raw reports are retained as workflow artifacts.

The current benchmark page contains the complete generated matrix and provenance. A scheduled run replaces the marked snapshot only after all three runtime reports pass the publication validator.

For developers and contributors

The repository is a Node.js workspace:

apps/cli/              command-line interface and setup
apps/gateway/          local service and browser-based console
adapters/channels/     messaging-network adapters and agent tool definitions
packages/application/ transport-neutral contracts and migrated use cases
packages/channels/    shared channel contracts and routing abstractions
packages/kernel/       AI providers, memory, sessions, and tool execution
packages/policy/       permissions and safety rules
packages/protocol/     shared request and activity formats
packages/runtime/      host composition for kernel, adapters, and workers
packages/store-file/   local append-only storage
packages/store-sqlite/ durable run and artifact storage
tests/                 unit, integration, CLI, and platform coverage

Common development commands:

pnpm install
pnpm check
pnpm test
pnpm build

Detailed architecture, release validation, storage formats, remote hosting, API routes, and extension contracts live under docs/. Keeping those details there lets this README explain the product before exposing the machinery underneath it.

License

MIT