SARAH — the all-in-one place for AI agents
What it does
The whole inventory, grouped by what you are trying to get done. Nothing here is an add-on, a higher tier or a beta.
Providers and models
- Any endpointOpenAI, OpenRouter, Groq, xAI, Google, DeepSeek, Mistral, Together, Fireworks and Cerebras — or a base URL and key for anything else that speaks the OpenAI format.
- Opencode Go and Zen, built inA fresh install can talk to something before you add a single key.
- Local modelsOllama on your own machine: no key, no bill, and nothing leaving the network.
- Model per conversationEvery chat, persona and team member carries its own provider and model, so a cheap model can do the grunt work and an expensive one can do the thinking.
- Spend, visibleOverall and per-provider usage, refreshed live, so a runaway run is obvious before the invoice arrives.
- Keys kept localEntered once, stored on your machine, never logged and never handed to a model.
Chat and control
- Chat, Plan, BuildThree modes and one question: how much may the agent touch? Plan cannot write at all.
- Approvals, queueing, steering, stopSend it off, then add a correction mid-run, queue the next job, or stop everything.
- Command CenterEvery live run on one screen: what it is doing now, its full conversation, its tool calls and its files.
- Multi-ChatSeveral agents on one question at the same time. Compare the answers, or let them argue it out and read the transcript.
- Team roomsA shared room with agents and invited guests. Guests join with a code and only get the permissions you hand them.
- Chat searchFull text across every project, filtered by when a chat was last active. A hit jumps straight back to that message.
- Pop-out windowsPull a chat into its own window and dock it back when you are done.
- Spoken repliesRead answers aloud, with a persona voice and a corner avatar that reacts while SARAH works.
Projects
- A folder, plus everything around itFiles, chats, agents, skills, plugins, tasks, notes, memory, environment and activity, all scoped to the project.
- Nothing bleeds between projectsMemory and context are per-project, so a client job never leaks into your side project.
- Activity recordEvery run, tool call and file write in order, so you can see what changed and when.
- NotesFree-form notes that live with the project and stay searchable.
Files, code and running things
- Explorer and editorBrowse and edit in place, with inline ghost completion where Tab accepts the suggestion.
- Run and previewRun code and serve a website from inside the app, with the output where you can see it.
- TerminalThe agent shell commands and your own console share one dock, so nothing needs a second window.
- GitBranch, stage with checkboxes, commit, push, pull, stash, log and per-file diffs. The agent can do the same work headless.
- GitHubPull requests, issues and workflow runs without leaving SARAH.
- Remote projects over FTPPoint a project at an FTP site and edit it as if it were local.
- Regex file searchFinds the code rather than reading files one at a time.
Tools that reach outside
- Plugins.sarah bridges to Discord, mail, licensing, TOTP codes, social posting, game servers and more. Every plugin works in every project and every mode.
- Package any programPython, Node, a C binary, a curl pipeline — point SARAH at a script that already works and it becomes a named tool with arguments.
- MCPModel Context Protocol servers over stdio or HTTP, so tooling you already run elsewhere comes along.
- The webFetch pages and search for documentation instead of guessing at an API.
- Computer controlScreenshots, clicks, typing and key presses, for the jobs where driving another app is the only way through.
- Screen drawCircle the thing you mean on a screenshot and send it with the marks baked in.
- Shot StudioA screenshot gallery with an annotator, for keeping the shot that proves it works.
Skills and memory
- Skills teach judgementMarkdown files that carry your conventions — 202 ready-made in the catalogue, or write your own and keep it with the project.
- Learn by watchingWatch mode turns a job you just did into a skill, so the next run already knows.
- Memory, two scopesPer-project memory is injected automatically; global memory follows you everywhere, and only when you say so.
- Recall on demandAsk about another project and SARAH can look up what it learned there, without that memory polluting this one.
- Memory browserSee and edit every remembered fact instead of trusting the black box.
Safety nets
- Time MachineSnapshots taken automatically and on demand. Revert a file, or the whole project, to any point.
- Undo, not just revertThe last few things you asked for can be unwound one at a time.
- SandboxesDisposable copies of a workspace where an agent can try something risky, several in parallel, without touching the real project.
- Self-edit is journaledWhen SARAH rewrites its own UI, skills or plugins, every change is backed up and undoable, and a syntax error is refused before it lands.
- Plan mode is read-onlyA hard boundary rather than a promise, so looking can never turn into writing.
Sharing and automation
- Share a run or a chatPublishes a static page at /s/<id>, server-rendered with no scripts and home paths rewritten, so a link cannot leak your desktop or a token.
- SchedulesTimers that tick while the server runs, pause when it stops, and fire once if they were overdue on restart.
- Task boardTo do, doing, done over the project task file the agent also reads. Repeating tasks roll over when ticked, so the weekly job never vanishes.
- Auto-updateA quiet pill when a newer build is published, and a changelog before it installs.
Interface
- Split panesTwo resizable panes with layout presets; the sizes you drag are remembered.
- ThemesDark, light, or a theme you build yourself in the custom theme editor.
- Environment panelOne machine-level list of variables, so a key entered once works in every project instead of per project.
Where your data lives
- Local filesProjects are folders on your disk. Nothing is uploaded to us — the only thing that leaves is the model call you asked for.
- No accountThere is nothing to sign up for, and nothing to cancel.
SARAH, shown not told
01 — Command CenterLive runs, approvals, queue, steering and stop.
02 — ProjectsFiles, chats, agents, skills, plugins, tasks, notes, memory.
03 — SkillsMarketplace + project skills that teach agents your ways.
04 — Tasks & Time MachineTrack work, snapshot files, revert anything.
06 - Multi-ChatMultiple agents working on a single project together at the same time.
Click any shot to open it large. Replacements go in through Admin → Screenshots and appear here and in the gallery window at once.
Asked often, answered plainly
What is SARAH, in one sentence?
A desktop home for AI agents: chat with any provider with an API key, organise work into projects, and let agents use tools, skills and plugins under your control.
How is it different from Cursor or Claude Code?
Those live inside code. SARAH lives around it — files, memory, publishing, computer control, timers, backups — so agents can do non-code jobs too.
Which models does it use?
Any provider that speaks the OpenAI API — OpenAI, OpenRouter, Groq, xAI, Google, DeepSeek, Mistral, Together, Fireworks, Cerebras, local Ollama, and the built-in Opencode Go/Zen. Add your own custom endpoint whenever you like. You bring keys; SARAH never locks you to one lab.
Is my work safe?
Chat, Plan (read-only) and Build (full access) modes, plus approvals, steering, stop, Time Machine snapshots and a full activity log.
What does it cost?
SARAH costs nothing to download or use; you pay your chosen providers for usage.
Where do I get plugins and skills?
You can get Plugins and Skills inside the SARAH app in the Skills and Plugins bubbles.
Where do I get it?
Open the Download window on the home page and choose Windows or Linux.
One-off, any amount
SARAH is free — free to download, free to keep, nothing held back behind a paywall. If it saved you an hour this week, a few quid keeps the next version coming.
Card payments go straight to Stripe. No account to create, no subscription, and SARAH never sees your card details.
What it actually pays for
- Hosting and the domainThis site, and the bandwidth for both builds every time somebody downloads one.
- Model credits for testingEvery provider integration is run against a real API before it ships, and that is paid for out of pocket.
- Hours between releasesSARAH is built by one person. A donation buys time, not features.
What you get
Nothing extra, and that is deliberate. The free build is the whole build: every provider, Chat, Plan and Build, the Command Center, Time Machine, skills and plugins. A donation buys no licence, no priority support and no badge.
If it went wrong
Donated by mistake, or the wrong amount? Email support@sarahai.co.uk and it gets returned — no argument.
Thank you. Genuinely — it is the only thing paying for the next one.
Desktop app · Your keys · Your projects
SARAH is the home your agents have been missing
SARAH is a desktop app that gives AI agents a proper place to work: chat through many providers, organise everything into projects, hand them tools, skills and plugins — and keep control with approvals, steering and a full activity record.
What SARAH actually is
Most agent tools live inside a code editor. SARAH lives outside it. It is a workshop: projects hold your files, chats, agents, skills, plugins, tasks, notes, memory and environment in one place. Agents run shell commands, read and write your project files, call plugin bridges, package new tools, and even edit SARAH itself when you ask. Chat, Plan and Build modes decide how much they may touch.
An app on your machine, not a tab
SARAH is a desktop application with a small server that runs alongside it. Both sit on your computer. A project is an ordinary folder you can open in any file manager, and agent runs happen locally — the only thing that leaves the machine is the model call you asked for. Close the app and nothing keeps working in a warehouse.
Built around projects, not files
A project is the unit of everything. Open one and you get its folder, its chats, its personas, the skills and plugins it uses, its tasks, its notes, its memory, its environment and a record of every run that touched it. Switch project and all of that switches with you, so two clients, two codebases or two hobbies never share a memory by accident.
Three modes, one dial
The only thing that changes between them is how much the agent may touch. Plan mode reads and cannot write. Chat mode uses tools but does not go editing your project. Build mode does the work — and even then, anything reckless waits at an approval until you say yes.
What it is not
- Not an editor pluginIt sits beside your editor rather than inside it, so the job can be a spreadsheet, a game server or a website as easily as code.
- Not a chat wrapperAgents get a shell, a file system, the web, plugins and the ability to drive another app. Talking is the interface, not the product.
- Not a hosted serviceNo account, no seats, no upload. Your keys, your folders, your machine.
- Not a modelSARAH does not sell tokens. It connects to the providers you already pay for, or a model running on the same computer.
"Cursor edits code. SARAH runs the whole job — files, tools, memory, publishing — and shows its working."
A day with it
You ask for something in plain language. The agent plans, then reads the files it needs. Anything that writes or runs waits for an approval if it looks risky. The Command Center shows each step as it happens, the terminal shows the shell output, and the activity record keeps the whole run. If the result is wrong, Time Machine puts the project back the way it was. That loop — ask, watch, approve, revert — is the product.
Who it suits
Anyone whose work is more than code: research, operations, writing, admin, game servers, small websites, data clean-up. If the job is “look at this, work out what to do, do it, show me”, SARAH is built for it.
Download SARAH
One download, your providers, your projects. Bring API keys for the models you already use — any provider with an API key works. Timers only tick while the app is open. Your files stay yours.
Windows or Linux. The same app, with your providers and your projects.
Windows: run install.bat. Linux: run ./install.sh.
Like SARAH? Chip in a few quid — all voluntary.
Before you click
- Bring your own keysSARAH connects to the providers you already pay for. No new subscription to start.
- Start in Plan modeLet agents look and think first. Flip to Build once the plan looks right.
- Keep Time Machine onSnapshot before big jobs. Revert the whole project if a run goes sideways.
Timers only tick while the app is open. Your files stay yours.
Contact us
Give me a message! I will respond as soon as possible. Ask me anything, question about SARAH? Question about me or my business? Ask away and i will be as honest as possible.,
Usually Responds Within 5 Hours
Tell us what you think
Using SARAH? Say what works, what does not, and what you would change. Good or bad — both are useful.
What people have said
Nothing published yet. Yours could be the first.
Feedback is read before it appears here, and your email address is never published.
Where SARAH wins, and where it does not
Four tools people ask about. Each row below is the whole story — every point for them and every point for SARAH, the good and the bad, exactly as it appears on the full comparison. Nothing trimmed to flatter us.
What Cursor does well
- Theirs The fastest in-editor completions going — Tab is still the benchmark.
- Theirs Agent mode plus Cloud agents running in parallel.
- Theirs Full VS Code extension ecosystem, LSP and debugging built in.
Where SARAH stands
- SARAH Has its own built-in editor (CodeMirror, file tree + editor + chat in one window).
- SARAH Works with any provider that has an API key, so no model lock-in.
- SARAH Inline AI ghost completion (POST /api/complete), so there is Tab-style suggestion here too.
- SARAH LSP diagnostics and go-to-definition from server/lsp.js.
- SARAH An interactive terminal panel you can type in yourself (TerminalPanel.jsx + POST /api/terminal/run).
- SARAH Computer control, FTP projects, Time Machine and memory alongside the code.
What Opencode does well
- Theirs Open source under MIT — read it, fork it, audit it.
- Theirs 75+ providers out of the box.
- Theirs Terminal, desktop app and IDE extension, very cheap to run.
Where SARAH stands
- SARAH Same provider freedom — any OpenAI-compatible API, plus Opencode Go preconfigured.
- SARAH A visual home around the agent: projects, approvals, Command Center, Time Machine.
- SARAH Built-in editor, MCP support, computer control and FTP projects.
- SARAH Open source too — LICENSE at the repo root is MIT (Copyright 2026 Spectives), with EULA.md alongside.
What Claude Code does well
- Theirs Top-tier code quality on Anthropic models.
- Theirs 1M-token context window.
- Theirs Terminal, desktop, web, IDE, Slack and mobile — one engine everywhere.
Where SARAH stands
- SARAH Runs Claude and every other provider side by side, so no single-lab lock-in.
- SARAH Built-in editor, computer control, FTP, screenshots and Time Machine.
- SARAH Skills, plugins and a marketplace for know-how beyond code.
- SARAH MIT-licensed like the rest of the project (Copyright 2026 Spectives), with EULA.md for the desktop terms.
What Codex does well
- Theirs One ChatGPT seat covers CLI, IDE, web and iOS.
- Theirs Parallel cloud sandboxes, PR-ready diffs and a GitHub bot.
- Theirs Generous usage on Plus and Pro tiers.
Where SARAH stands
- SARAH Local-first — projects, notes, chats and backups sit on your machine.
- SARAH Built-in editor, 50 agent tools, computer control and MCP support.
- SARAH Use OpenAI models without being tied to a ChatGPT subscription.
- SARAH Parallel sandboxes run locally — isolated workspace copies on disk (server/sandboxes.js), not cloud VMs.
Written to be checked, not to be believed: every claim about SARAH was verified against the source code before it went up.