What is Osmium Coder?

A native desktop agent IDE for Mac, Windows, and Linux. You open a folder, and an agent reads and edits it while you keep or discard each change. It includes a Monaco editor, a real terminal, git, a debugger, ripgrep search, and MCP support in one window.

Yes, for local models. The full IDE and the Agent, Plan, Ask, Analyze, and Debug modes work with models running on your own machine, with no API key, no card, and no usage metering. You create a free account with an email code to download. You only pay if you want frontier cloud models or the heavier Architect, Arena, and Fleet modes.

Which local models can I run?

Anything Ollama or LM Studio can serve, plus local GGUF files through llama.cpp. Osmium ships a curated catalog of 16 models with real VRAM figures, from Qwen2.5 Coder 1.5B at about 2 GB up to Llama 3.3 70B at about 40 GB, and it reads your hardware to tell you which ones fit before you download anything.

What hardware do I need?

A 4 GB GPU is enough to run Qwen2.5 Coder 1.5B or 3B. Around 8 GB gets you Qwen2.5 Coder 7B, which is the model most people should start with. Apple Silicon works well because unified memory counts as usable VRAM. Osmium rates every model as fitting, tight, or too big for your specific machine.

Does my code leave my computer?

Not when you use a local model. The prompt, the context, the generation, and the edit all happen on your hardware, and Osmium works with no network connection at all. If you choose a cloud model, then your prompt and attached context go to that provider — which is why the choice is explicit.

Do I have to paste an API key?

No. There is no API key field anywhere in the product. Local models do not need one. Cloud models are proxied for you, or configured by an administrator. If a cloud model has no key behind it, the app tells you to ask an admin rather than showing you a paste box.

How is this different from Cursor, Copilot, or Devin?

Those require their cloud to work at all, and bill per seat or per task. Osmium's core loop runs on a GPU you already own, for free, and the agent writes directly to your working tree instead of a hosted sandbox. We do not claim unattended multi-hour runs or zero errors.

What do the paid plans add?

Starter at $49 and Plus at $89 add frontier cloud models — Claude, ChatGPT, Grok, Gemini, DeepSeek, and Composer — with an included monthly usage grant, proxied so you never handle keys. Plus is more included usage, not more tools. Pro at $140 permanently unlocks Architect mode, Model Arena, and Autopilot Fleet.

Can I use it offline?

Yes. Once a local model is pulled, the editor, terminal, git, debugger, and agent all work with no connection. Your sessions and history live in a local SQLite database on your machine.

Who sells this and can I cancel?

Zenguard Private Limited, billed through Razorpay on osmiumhq.com. Subscriptions can be cancelled at any time, and yearly plans receive a prorated refund for unused months where applicable. The free tier needs no payment method at all.

Osmium Coder — the agent IDE that runs on free local models

PLATE 02The free part is the whole IDE

Not a trial. Not a teaser tier.

Most AI editors give you a free plan that runs out. Osmium gives you the entire IDE and every agent mode, running on models on your own hardware, with nothing metered. There is no key to paste and no card to add, because there is nothing to bill — the compute is yours.

No key, ever
There is no API key field anywhere in the product. Local models talk to localhost.
Nothing metered
Local generation costs us nothing, so we count nothing. Run it all day.
Works offline
Pull a model once and the agent keeps working on a plane or an air-gapped box.

Your GPU

7 of 16 models run on this. We would start you on qwen2.5-coder:7b.

Checked against GPU memory only. The app also checks system RAM and free disk on your actual machine.

PLATE 03Your machine

Your code never leaves the box.

With a local model, the whole loop — prompt, context, generation, edit — happens inside your own hardware. Nothing is uploaded, nothing is logged by us, and there is no vendor in the path. Switch to a cloud model and the picture changes; that is why the picture is on the page.

Prompt, context, and generated code stay inside your machine.

Composer · Models

Ln 10TypeScriptCloud locked

PLATE 04The IDE

One window. Editor, terminal, git, agent.

This is a real desktop IDE, not a chat box with a file picker. Monaco for editing, a proper terminal, git status and commits, a debugger, ripgrep search, and the agent dock — all in the same window, all working on the folder you opened.

demo

$ █

Ln 4TypeScriptOllama · qwen2.5-coder:3b

Explorer · Your actual folder tree.
Editor · Monaco, tabs, split view, problems.
Terminal · Real shell via node-pty.
Git · Status, diff, stage, commit.
Agent · Six modes, one context.

Everything in the IDE →

PLATE 05Six modes

Read-only when you want it. Writing when you ask.

The mode decides what the agent is allowed to do. Ask and Analyze never touch a file. Plan writes a plan and stops. Agent and Architect edit, and every edit lands as a diff you approve.

Agent
01
Edits files and runs commands
Give it a task. It reads the repo, edits files, runs commands, and shows you a diff to keep or discard.
Architect
02
Big multi-file changes
For work that spans many files at once — migrations, renames, restructuring a module end to end.
Plan
03
Plan before touching anything
Produces a plan first and waits. Useful when the task is ambiguous or the blast radius is large.
Ask
04
Read-only answers
Questions about the codebase with no edits. Nothing is written while you are in Ask.
Analyze
05
Full architecture report
A structured read of how the project actually fits together. Read-only.
Debug
06
Debugger in the chat column
Breakpoints, stepping, and variable inspection driven from chat. F5 to start, F9 to break, F10 and F11 to step.

PLATE 06Model catalog

Pick a model your GPU can actually run.

Osmium ships a curated catalog with real VRAM numbers, and reads your hardware to tell you what fits before you download gigabytes. Any other Ollama tag works too — the catalog is guidance, not a whitelist.

Your GPU

Start with qwen2.5-coder:7b on this hardware.

Browse all models →

Composer · Models

Ln 10TypeScriptCloud locked

Scores are product guidance from our own use, not benchmark results.

PLATE 07Keep or discard

The agent proposes. You decide.

Every change arrives as a diff in your working tree. Keep it, discard it, or roll the whole conversation back. Chat restore and code restore are separate, so you can rewind the code without losing the thread.

Inline diffs
See exactly which lines moved before anything is permanent.
Code restore
Roll the working tree back to any checkpoint in the session.
Chat restore
Rewind the conversation without touching your files.

PLATE 08Terminal, git, debugger

The boring parts, done properly.

An agent that cannot run your tests is a suggestion engine. Osmium has a real shell, real git, and a real debugger, so the agent can run the command, read the failure, and fix it.

Terminal
Multi-tab shell under the editor. The agent uses the same one you do.
Git
Status, diffs, staging, and commits without leaving the window.
Debugger
Breakpoints and stepping, driven from chat in Debug mode.
Search
Ripgrep across the workspace, plus an index for @-context.

PLATE 09 · Contents

01Monaco editor
02Agent mode
03Architect mode
04Plan mode
05Ask mode
06Analyze mode
07Debug mode
08Integrated terminal
09Git status and commit
10Inline diff review
11Keep or discard
12Chat restore
13Code restore
14Ripgrep search
15Quick open
16Workspace index
17MCP servers
18Live preview
19Model Arena
20Autopilot Fleet
21Problems panel
22Ollama
23LM Studio
24Local GGUF
25Mac, Windows, Linux
26Offline capable

PLATE 10MCP

Give the agent your own tools.

Osmium speaks the Model Context Protocol, so any MCP server you point it at becomes something the agent can call — your database, your ticket tracker, your internal API. Per-workspace, and off by default.

MCP setup →

.osmium/mcp.jsonper workspace

{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@mcp/postgres", "$DATABASE_URL"]
    },
    "linear": {
      "command": "npx",
      "args": ["-y", "@mcp/linear"]
    }
  }
}

postgres.queryread the sessions table schema
linear.issuepull acceptance criteria for OSM-214
fs.writeapply the migration to your working tree

PLATE 11Arena and Fleet

Race models. Queue work.

Arena runs the same prompt through several models side by side so you can see which one actually handles your codebase. Fleet queues agent jobs to run in the background while you keep working.

Architect, Arena, and Fleet are on a 14-day trial, then part of the Pro plan.

Arena · same prompt, four models

qwen2.5-coder:32blocalpassed 12/12
claude-sonnetcloudpassed 12/12
qwen2.5-coder:7blocalpassed 10/12
deepseek-coder-v2:16blocalpassed 9/12

Fleet · background jobs

Upgrade prisma to 6.xrunning
Add tests for auth guardqueued
Backfill JSDoc in services/queued
Fix eslint in src/utilsdone

PLATE 12Cloud, when you want it

Frontier models without the key management.

Some jobs want a frontier model. Turn one on and Osmium proxies it for you — Claude, ChatGPT, Grok, Gemini, DeepSeek, Composer. You still never paste a key, and usage is metered against your plan so there is no surprise provider bill.

No keys
There is no paste field. Keys live with the admin or with us.
Metered, with a hard stop
Spend stops at your balance instead of overrunning.
Local stays free
Turning cloud on does not start metering your own hardware.

PLATE 13Privacy and offline

Local-first is an architecture, not a setting.

Your work lives in SQLite on your machine. Osmium runs with no network at all if the model is local, and there is a kill switch for everything that reaches out.

Local database
Sessions and history stay in a file you own.
Offline capable
No connection needed once a local model is pulled.
Kill switch
One control freezes outbound calls and spend.

PLATE 14How we are different

Three honest differences.

01
It runs on hardware you already own
Cursor, Copilot, Windsurf, and Devin all require their cloud to function. Osmium's core loop runs on your GPU, for nothing.
02
It edits your repo, not a sandbox
The agent writes to the folder you opened, on your disk. There is no VM to watch and no PR to import.
03
You never handle keys
No provider signup, no key rotation, no per-seat API spend to reconcile.

vs Cursor →vs Devin →vs Copilot →vs Windsurf →

PLATE 15Pricing

Free is the product. Paid is the cloud.

Everything that runs on your own hardware is free and always will be. You pay when you want frontier cloud models, or the heavier agent modes.

Tariff	FreeStart The whole IDE on your own models. Forever. $0forever	Starter Add frontier cloud models when local is not enough. $490/yr	Plus The same tools with a bigger cloud allowance. $890/yr	ProPopular Architect, Arena, and Fleet unlocked for good. $1,400/yr
Seats	1	1	1	1
IDE	Full Osmium Coder	—	—	—
Local models	Unlimited — Ollama, LM Studio, GGUF	—	—	—
Modes	Agent · Plan · Ask · Analyze · Debug	—	—	—
Tools	Git · Terminal · Preview · MCP · Search	—	—	—
Cloud models	Not included	Claude · GPT · Grok · Gemini · DeepSeek	—	—
Card required	No	—	—	—
Everything in Free	—	Yes	—	—
Included usage	—	Monthly credit grant	Higher monthly grant	—
API keys	—	Never — proxied for you	—	—
Architect / Arena / Fleet	—	14-day trial	14-day trial	—
Support	—	Email	Email	Priority
Everything in Starter	—	—	Yes	—
Everything in Plus	—	—	—	Yes
Architect mode	—	—	—	Included
Model Arena	—	—	—	Included
Autopilot Fleet	—	—	—	Included
Start	Free account by email code. No card.		More cloud usage, not more tools.

Prices are USD-native. Other currencies convert at checkout via live FX. Billing is handled by Razorpay.

PLATE 16FAQ

Questions people actually ask.

01What is Osmium Coder?: A native desktop agent IDE for Mac, Windows, and Linux. You open a folder, and an agent reads and edits it while you keep or discard each change. It includes a Monaco editor, a real terminal, git, a debugger, ripgrep search, and MCP support in one window.
02Is it really free?: Yes, for local models. The full IDE and the Agent, Plan, Ask, Analyze, and Debug modes work with models running on your own machine, with no API key, no card, and no usage metering. You create a free account with an email code to download. You only pay if you want frontier cloud models or the heavier Architect, Arena, and Fleet modes.
03Which local models can I run?: Anything Ollama or LM Studio can serve, plus local GGUF files through llama.cpp. Osmium ships a curated catalog of 16 models with real VRAM figures, from Qwen2.5 Coder 1.5B at about 2 GB up to Llama 3.3 70B at about 40 GB, and it reads your hardware to tell you which ones fit before you download anything.
04What hardware do I need?: A 4 GB GPU is enough to run Qwen2.5 Coder 1.5B or 3B. Around 8 GB gets you Qwen2.5 Coder 7B, which is the model most people should start with. Apple Silicon works well because unified memory counts as usable VRAM. Osmium rates every model as fitting, tight, or too big for your specific machine.
05Does my code leave my computer?: Not when you use a local model. The prompt, the context, the generation, and the edit all happen on your hardware, and Osmium works with no network connection at all. If you choose a cloud model, then your prompt and attached context go to that provider — which is why the choice is explicit.
06Do I have to paste an API key?: No. There is no API key field anywhere in the product. Local models do not need one. Cloud models are proxied for you, or configured by an administrator. If a cloud model has no key behind it, the app tells you to ask an admin rather than showing you a paste box.
07How is this different from Cursor, Copilot, or Devin?: Those require their cloud to work at all, and bill per seat or per task. Osmium's core loop runs on a GPU you already own, for free, and the agent writes directly to your working tree instead of a hosted sandbox. We do not claim unattended multi-hour runs or zero errors.
08What do the paid plans add?: Starter at $49 and Plus at $89 add frontier cloud models — Claude, ChatGPT, Grok, Gemini, DeepSeek, and Composer — with an included monthly usage grant, proxied so you never handle keys. Plus is more included usage, not more tools. Pro at $140 permanently unlocks Architect mode, Model Arena, and Autopilot Fleet.
09Can I use it offline?: Yes. Once a local model is pulled, the editor, terminal, git, debugger, and agent all work with no connection. Your sessions and history live in a local SQLite database on your machine.
10Who sells this and can I cancel?: Zenguard Private Limited, billed through Razorpay on osmiumhq.com. Subscriptions can be cancelled at any time, and yearly plans receive a prorated refund for unused months where applicable. The free tier needs no payment method at all.

PLATE 17 · Close

Download it. Pull a model. Start working.

Three installers, no card, and a model that fits the GPU you already have.

Osmium Coder

Ln 10TypeScriptCloud locked

Not a trial. Not a teaser tier.

No key, ever

Nothing metered

Works offline

Your code never leaves the box.

One window. Editor, terminal, git, agent.

Read-only when you want it. Writing when you ask.

Agent

Architect

Plan

Ask

Analyze

Debug

Pick a model your GPU can actually run.

The agent proposes. You decide.

Inline diffs

Code restore

Chat restore

The boring parts, done properly.

Terminal

Git

Debugger

Search

Give the agent your own tools.

Race models. Queue work.

Frontier models without the key management.

No keys

Metered, with a hard stop

Local stays free

Local-first is an architecture, not a setting.

Local database

Offline capable

Kill switch

Three honest differences.

It runs on hardware you already own

It edits your repo, not a sandbox

You never handle keys

Free is the product. Paid is the cloud.

Questions people actually ask.