Scout
Repository researcher
Finds new projects, clones the promising ones and writes a sourced weekly report.
- Tools
- Web search, scraping, shell
- Runs
- Mondays 09:00
Tell your chief of staff what you need. It hands the work to your bots, each in its own cloud computer, and they ask before making changes.
01The run
Every step streams as it happens: what the model says, which tool it calls, what came back. You can follow a run in the CLI, the web app or over the HTTP API.
The agent loop sends your task and its tools to the model, runs each tool call, and feeds the results back until the job is done.
Before a write, a send or a risky command, it shows you the exact arguments and waits for a yes.
Each run ends with the step count, tokens used, cache hit rate and the cost the provider reported.
02How it works
A name and what it's for. It gets its own sandbox, memory and skills, and learns the rest by chatting with you.
Type it in the web app, run it from the CLI or send it over HTTP. With no task, it follows its saved instructions.
Say yes to the changes you want. It comes back with the result and can save what it learned.
edit_file reports/2026-10-01.md
03Chief of staff
Every account starts with a chief of staff. Tell it what you need. It splits the work between your bots, runs them side by side and brings their answers back to you.
Get me ready for Thursday's renewal call with Acme.
On it. Inbox Manager will go through the Acme threads while Scout checks their news.
BriefSummarise every Acme renewal thread since July, list the open questions and find 30 free minutes before Thursday's call.
BriefFind Acme news from the last 90 days, such as funding, leadership changes and launches. Cite sources.
Your brief is ready. Acme still wants a three-year price lock and SSO in the base plan, and it raised a Series C in August. You're free 9:30 to 10:00 on Thursday to prepare.
It reads your bots' descriptions and writes each one a self-contained brief. If no bot fits, it suggests one and creates it once you agree.
Several handoffs in one reply run at the same time. Each bot works in its own computer, with its own memory, skills and tools.
A handoff window streams each brief, step and reply. The work stays in that bot's chat, so you can carry on with it directly.
Ask the chief to make it a weekly job, and it schedules the handoff itself.
04The crew
Each bot keeps its own workspace, memory and skills, so the researcher's notes never end up in the inbox manager's files. The chief of staff decides who gets what.
Repository researcher
Finds new projects, clones the promising ones and writes a sourced weekly report.
Email assistant
Finds threads that need you, summarises them, drafts replies and finds time for the follow-up call.
Maintainer's helper
Reads new issues, tries to reproduce bugs in its sandbox and comments with what it found.
Knowledge keeper
Reads the contracts and scans in your Drive, keeps your Notion pages up to date and sends you a summary on Telegram.
05Features
Bots that remember, keep their own files, read your documents, run on a schedule and never act behind your back.
Each bot works in its own persistent Linux sandbox on Daytona. It stops after 15 idle minutes and starts again with its files intact.
Reads run on their own. Writes, sends and risky commands wait for one-time approval of the exact arguments, and expire after five minutes.
gmail_send · to dana@acme.com
Two Markdown files per bot: your preferences and what it has learned. It fills them in as you chat, and you can edit them any time.
reports/, one per weekSay "every Monday at 9" and the bot schedules itself. Pause, resume or cancel from the Scheduled page.
Teach a workflow once as a skill folder. Bots see a short list and load the full instructions only when a task calls for them.
Give each bot a character for how it talks to you. The work stays exact whatever the voice.
Drive files and Gmail attachments come back as Markdown, scanned pages and photos included. Parsing and OCR run on your server, with no cloud service.
When a task fills half the model's context window, Pekka summarises the earlier steps and carries on. Your request and the latest step stay word for word.
Deleting files, force-pushing, sudo and piping a download into a shell are blocked outright, whether or not anyone is watching.
06Plugins
Connect a tool once and all your bots can use it. Pick one to see exactly what bots can do with it.
Off until you say soConnecting a tool leaves access off. Turn it on, or off again, any time.
Tokens stay on the serverEncrypted in your database, never shown to the model or the bot's computer.
Reads run, changes askLooking things up is automatic. Anything that writes or sends waits for your yes.
07Architecture
Pekka is a Node.js app you run yourself. It calls the services you choose and keeps every credential on the server.
Talk to your bots from wherever suits you.
The API and scheduler you run. They hold the keys.
The model, any one with tool calling
One persistent sandbox per bot
Bots, memory, skills and schedules
GitHub, Google, Notion, Telegram, AgentMail
Turn on Google sign-in to share one server with your team. Everyone gets their own bots, memory, schedules and plugin connections, and nobody can see anyone else's.
08Early access
Pekka isn't open to everyone yet. Join the waitlist and we'll email you when your spot opens.
A chat assistant tells you what to do. A Pekka bot does it: it runs the commands, writes the files and uses your connected tools itself, then reports back. It keeps its workspace and memory between tasks and can run on a schedule.
Yes. Each named bot has its own persistent Daytona sandbox, plus its own memory and skills. Sandboxes stop after 15 idle minutes and start again on the next task with their files intact.
Yes, through your chief of staff. It sends self-contained briefs to your other bots, runs several at once and brings their answers together. Each bot does the work in its own computer and keeps the brief and its answer in its chat, so you can follow up with it. Bots that receive a handoff can't pass it on.
Read: files, memory, skills, web searches, read-only plugin calls, and a short allowlist of shell commands such as ls and cat. Everything else, including other commands, file and memory writes, schedule changes and anything that sends or publishes, needs one-time approval of the exact arguments.
Any OpenRouter model that supports tool calling. Set PEKKA_MODEL to pick one.
Yes. Turn on Google sign-in and list the addresses or domains allowed in. Each person gets their own bots, memory, schedules and plugin connections.
Pekka is free and open source. You pay OpenRouter for model usage, Daytona for sandboxes, and any optional providers you turn on.
Bots can't drive a web browser yet, and a bot's own email address is send-only for now. Word, Excel and PowerPoint files can't be read yet. Scheduled tasks run while the pekka scheduler process is running.
Your chief of staff is almost ready for you.