Comparisons
Several products now offer an agent that works on its own for one person or a team. This page compares Valor with the ones closest to him, on the two things Valor is built around: independence and data sovereignty. Facts about other products were checked on 2026-10-21 and link to their sources.
Inspiration
Projects Valor took ideas from
Valor was built from other people's work as well as his own. Each project below links to its home, with what Valor took from it and the file in Valor's repository that says so.
- Hermes AgentNous Research
- One long-lived session per chat, the way chat sessions compact, and parts of the Slack bridge (Socket Mode, the ping watchdog, the catch-all ack).
docs/harnesses.md · docs/plans/team-chat.md · Slack plan - OpenClawOpenClaw Foundation
- The Slack bridge's crash safety: each event recorded before it is acknowledged, and a marker on every post so a send isn't repeated after a crash.
Slack plan · bridges/slack/send.py - coriTom Counsell
- Valor's core idea, authority in a deterministic kernel outside the model's reach, and the corrigibility commitments. It is also one of the four places a component is taken from before one is written.
CLAUDE.md, rule 10 · cori setup plan - Claude CodeAnthropic
- The harness every turn runs in, and the first place to take a component from. Chat compaction is Claude Code's own.
docs/harnesses.md · CLAUDE.md, rule 10 - Piearendil-works
- Valor's second harness: a turn can run in Pi instead of Claude Code, behind the same port. Pi's durable package was also studied for orchestration ideas.
docs/pi.md · docs/harnesses.md · docs/plans/pi-durable-research.md - popotoTom Counsell
- The memory layer, and the measurement that raw turns beat model extraction, which is why memory stores raw entries.
REFERENCES.md · docs/memory.md - PaperclipPaperclip
- The ideas of budget hard-stops, goal-aligned tracing and an immutable audit log, adopted without the platform.
old research page - OpenHumanTinyHumans
- The case for a memory store people can read as files.
old research page
Studied, not adopted
- Firstmate: The closest outside system to Valor's shape: one session handing work to agents in isolated checkouts. See the comparison below.
- Claude Managed Agents: Anthropic's hosted agent loop, studied as a place to move execution later; not adopted.
- Perplexity Computer: A visual digital worker; it showed why an orchestrator needs an API it can call.
- Omnigent: An open-source agent control plane; it was a reminder to check maturity claims at the source.
- jcode: A fast single-author coding runtime with no integration seam; an example of a project to watch rather than adopt.
Sources for the studied list: the old research page
Terms
Independence and data sovereignty
- Independence
- Valor runs on his own machines and acts as himself on his own accounts: his own email and Google Workspace seat, his own GitHub and chat accounts. He never acts as Tom, who built Valor and runs Yudame, and holds none of Tom's accounts.
- Data sovereignty
- Valor's record of his work stays on those machines: the ledger of every call, turn and effect, the turn transcripts and the memory store, all in Postgres. The code is public under the GPL.
What still leaves Valor's machines, such as model calls, is listed on Trust and safety. Four of the products below run the agent on the vendor's computers, and five act under your account.
Source: docs/mission.md · docs/data.md · LICENSE
At a glance
Where each agent runs, keeps data, and whose name it uses
| Agent | Runs on | Data kept | Acts as | Price |
|---|---|---|---|---|
| Valor | his own machines | Postgres on those machines | his own accounts, as Valor Engels | works for Yudame's clients; code public |
| Firstmate | your machine (macOS or Linux) | files on your disk | your GitHub login | free, MIT license |
| OpenClaw | your machine or server | your machine | the chat accounts you connect | free, MIT; you pay for the model |
| Claude (Cowork) | Anthropic's servers by default | your Claude account | you, through your connectors | Pro or Max, $20 to $200 a month |
| Meta Muse | a virtual machine in Meta's cloud | that virtual machine | its own name and email | free, with paid plans |
| Grok Bot | Cursor-hosted computers in the US | xAI's service | you, the signed-in member | included with SuperGrok and Cursor plans |
| OpenAI dots | its own cloud computer, or your laptop with permission | OpenAI, for as long as you keep the dot (reported) | your connected accounts | Pro, from $100 a month (reported) |
Each linked name opens a page comparing that product with Valor. "Reported" means the figure comes from press coverage, not the vendor's own page.
Firstmate
Firstmate
Firstmate is an open-source project by Kun Chen, started in June 2026, with about 8,100 GitHub stars on 2026-10-21. It describes itself as "an agent distro for running a crew of agents": a set of instructions, skills and scripts that a coding agent follows. You talk to one agent, and it sends worker agents into their own git worktrees, watches them, and brings back pull requests, approved merges or reports.
What Valor and Firstmate share
- One agent you talk to, which hands the work to others.
- Each piece of work in its own isolated checkout.
- Both run on machines their operator controls, not in a vendor's cloud.
Where they differ
- Firstmate uses your GitHub login; Valor has his own accounts and name.
- Firstmate keeps state in files on disk; Valor keeps an append-only Postgres ledger.
- Firstmate leaves merging to you, or to authority you configure; Valor merges his own work when its checks pass, inside each task's effect ceiling.
- Valor works in Telegram, Slack and email; Firstmate works in your terminal.
Sources: github.com/kunchenguid/firstmate (README, 2026-10-21) · GitHub API (stars, created date) · Valor's engineering page
Hosted agents
The hosted agents from Anthropic, Meta, xAI and OpenAI
All four run the agent on the vendor's computers by default, and keep your sessions with that vendor. They differ in whose name the agent uses: Muse has its own, and the other three act as you.
Claude (Cowork), Anthropic
Anthropic has no feature called "cloud mode"; the nearest is Cowork, the agent in the Claude app, which now runs on Anthropic's servers by default for Pro and Max, in a sandbox destroyed when the session ends. Sessions and files are saved to your Claude account. Anthropic's Cowork pages give no retention period; its data page for Claude Code says consumer data is kept for 30 days, or for 5 years if you allow it to be used to improve the models. Cowork acts under your account through connectors, and Anthropic says you remain responsible for what it does. In September 2026 Anthropic merged Cowork into the main Claude app.
Sources: Cowork architecture · Cowork is now Claude (2026-09-16) · Claude Code data usage
Muse, Meta
Meta launched Muse on 2026-09-08 as "the world's first personal AI agent built for everyone". Each agent "runs on its own dedicated computer in the cloud", a virtual machine that holds both the agent and your data, and has its own name and email address. Meta says it doesn't share your conversations or that data with its ad systems, and that you can opt out of training; press reports say training is on by default. Meta says it's free for most uses, with subscription plans.
Sources: Meta newsroom (2026-09-08) · Wikipedia
Grok Bot, xAI
Grok Bot went into beta on 2026-08-11. Each bot works on a persistent cloud computer with a browser, files and a terminal; xAI says those computers are hosted by Cursor and run in the United States. xAI's security page says "a Bot has no identity or credentials of its own" and acts as the signed-in member. The same page says that after the service ends, data is deleted within 30 days of a written request, and action recordings are kept for 90 days. It's included with SuperGrok and Cursor's paid plans.
Sources: Grok Bot security · xAI news (2026-08-26)
dots, OpenAI
OpenAI announced dots on 2026-09-29: always-on agents in ChatGPT, each working from a cloud computer of its own that you can open to watch, and able to work on your laptop with permission. A dot keeps context from your conversations and connected apps for as long as you keep it, and can add to your ChatGPT memory. It signs in to services with your saved credentials. Press reports put the Pro plan, which includes one dot, at $100 a month or more. OpenAI's own pages couldn't be opened directly when checked, so these details come from press coverage and search summaries of OpenAI's help center.
Sources: TechCrunch (2026-09-29) · SiliconANGLE (2026-09-29) · OpenAI help center
OpenClaw
OpenClaw
OpenClaw is an open-source personal agent, first published in November 2025 and renamed OpenClaw in January 2026, with about 393,000 GitHub stars on 2026-10-21. You run it on your own machine or server and talk to it through WhatsApp, Telegram, Slack, Signal and other apps you connect. It's free under the MIT license, and you pay your model provider directly. Its data stays on the machine it runs on, as Valor's does, but it works through your accounts where Valor uses his own.
Sources: github.com/openclaw/openclaw · OpenClaw docs · Wikipedia