v0.4.0 · LIVE ON NPM

AI agents that live
in your browser.
Local, private, yours.

Pick any LLM. Bring your own keys. Build agents once, share them as markdown files, run everything on your machine. No cloud, no lock-in, no surprises on your bill.

~/sidekit · install
# 1. Install the local companion
$ npm install -g @try-sidekit/companion
# 2. Start it (keep this terminal open)
$ sidekit-companion
# 3. Chrome extension — in review, sign up below for launch
Free & open source
MIT licensed
Works offline with Ollama
No accounts, no telemetry

Today's AI makes you pick: powerful or private.
We pick both.

ChatGPT is powerful but locked to one model and one vendor. Local tools are private but primitive. Try Sidekit is the third option — real agents, your models, your machine.

01 / THE VENDOR TAX

Three AI subscriptions shouldn't be normal.

ChatGPT Plus, Claude Pro, and a dozen "AI-powered" apps. You already pay for the APIs through your other tools. Try Sidekit lets you use them directly — one interface, one tab of credentials.

02 / PRIVACY THEATER

Your data shouldn't travel to summarize a page.

Every AI sidebar ships your browsing to someone's server. Try Sidekit runs a sandboxed companion on 127.0.0.1. The only outbound request is the one you explicitly chose.

03 / LOCK-IN

Your workflows should outlive any one vendor.

Custom GPTs die when OpenAI deprecates them. Claude Projects don't port. Try Sidekit's Skills are markdown files — version them, share them on GitHub, compose them. Portable by design.

Built for the way you already work.

Try Sidekit lives in your browser's side panel, reads your tabs on permission, remembers per-agent context, and switches models on demand.

ANY LLM

Pick a different brain for every agent.

Research agent on Claude. Quick tasks on GPT-4o-mini. Private work on Ollama. Bring your own keys — they live in your OS keychain, not ours.

OpenAI · Anthropic · Google · Ollama
SKILLS

Agents as markdown files.

Each agent is a folder with a SKILL.md that defines its role, tools, and permissions. Drop in, share via GitHub, remix. Try Sidekit can also generate them from descriptions.

~/.sidekit/skills/my-skill/SKILL.md
SANDBOXES

Each agent gets its own room.

Per-agent directory. Per-agent permissions. Per-agent network allowlist. Agent A can't read agent B's files, and neither can touch anything outside their sandbox.

Filesystem isolation + network allowlists
GRACEFUL

Agents tell you what they can't do.

When an agent hits a wall, it says so. No hallucinated search results, no fake citations. It can also ask to load a skill mid-conversation to unlock new capabilities.

request_capability · list_skills · load_skill
COSTS

No bill-shock.

Token counts and dollar cost per turn, per agent. You see what you're spending as you go. Set Ollama as the default and the number stays at $0.

Per-turn meter · per-agent totals · USD breakdown
CONVERSATIONS

Separate chats per task.

Research notebook agent — one thread per project. Summarize agent — one thread per article. Switch, archive, delete. Each conversation is its own context window.

Unlimited threads · per-agent history

Five minutes to first agent.

Two pieces, one workflow. Install the companion, install the extension, you're chatting.

01 · INSTALL

Run the companion

npm install -g @try-sidekit/companion — one command, all platforms.

02 · CONNECT

Add the extension

Chrome extension launching soon. Join the waitlist to be notified the moment it's live.

03 · CONFIGURE

Paste one API key

Or skip and use Ollama locally. Stored in your OS keychain.

04 · USE

Pick a skill, chat

Built-in skills cover summarizing, research, and search. Or write your own.

Radically local.

We'll save you the privacy-policy bureaucracy: we don't have servers. We couldn't collect your data if we wanted to.

  • No accounts, no login Try Sidekit runs on your machine. There is nothing to sign up for.
  • No analytics, no telemetry We don't track which pages you use it on, how often, or what you ask.
  • Keys in OS keychain macOS Keychain, Windows Credential Vault, or GNOME Keyring. Never plaintext.
  • Per-agent network allowlists An agent can only make HTTP requests to hosts you explicitly approve.
  • MIT licensed, fully open source Audit exactly what it does. Fork it. Change it. Self-host forever.
🔒

Only 127.0.0.1.

The companion binds exclusively to localhost. No external access possible. CORS locks to Chrome extension origins only.

The only outbound requests are the LLM API calls you configured — OpenAI, Anthropic, Google, or nothing (Ollama stays offline).

No proxy. No logging. No servers. We don't operate infrastructure.

Read full privacy policy →

Get the extension first.

Chrome extension is in review. Drop your email and we'll ping you the moment it's live on the Web Store.

No spam. One email when the extension launches.