↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Assistant runtimes/

Octop

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1078 words)
agent-curated 576 assistant-runtimes 16 fully-ai-generated 679 llm=glm-5.3-flash 586 open-source 184 personal-assistants 7 python 31 research-note 546 self-hosted 25
Table of Contents

Octop is Tencent Cloud’s MIT-licensed, self-hosted AI assistant platform: one Python process serves a web dashboard, a CLI, native desktop clients, and eight-plus IM channels for a multi-user household or team, and each member gets their own team of specialized “experts”.

Octop is what the assistant-runtime wave looks like when an incumbent vendor builds it: multi-user by default, Tencent-enterprise-IM first, and composed from four small sub-repos (harness, gateway, memory, browser) instead of one monolith, at about 9k stars within three months of launch.

What it is
#

A Python 3.12+ FastAPI application that runs as a single restart-safe process: state lives in a control-plane database under ~/.octop/ (SQLite by default, PostgreSQL optional), and every surface (web dashboard, IM, cron) routes through one in-process HarnessProcessor. Users get “experts”, per-user agent profiles with their own workspace, providers, channels, and cron, plus 16 MBTI persona templates, an expert market for sharing configurations, and an AgentTeams beta where a coordinator schedules multiple experts on multi-step work. The IM bridge (octop-gateway) normalizes Feishu, DingTalk, QQ, WeChat, WeCom, Telegram, and Discord into one processing pipeline, and ACP works in both directions: octop acp serves your Octop agent to Zed or OpenCode, and outbound runners delegate to OpenCode, Claude Code, CodeBuddy, or Codex with permission gates. Providers are OpenAI-compatible APIs, DashScope (Qwen), Ollama, and other presets configured per agent, with a RAG knowledge base, CDP-based browser automation, remote desktop, and plugins rounding out the surface. It is published under the TencentCloud GitHub org and built from four focused sub-repos: octop-harness (the agent runtime), octop-gateway, octop-memory, and octop-browser.

Status
#

Fast out of the gate and young: 9,156 stars and 1,125 forks in its first three months as of 2026-10-14 (created 2026-07-08), 781 open issues with the last push on 2026-10-25 as of 2026-10-25, MIT.

Star History Chart

The beta line ended the way it should: v1.0.2, the first stable release on GitHub, published 2026-10-18 and uploaded to PyPI the same day, after the v1.0.2b4-through-b9 prerelease run (2026-09-27 through 2026-10-17) that PyPI tracked in parallel, and the v1.0.3 line opened as a beta on 2026-10-25 (v1.0.3b1, on GitHub and PyPI the same day: a default expert created alongside each new user, and permissions split per page tab), with the README carrying a Trendshift trending badge.

The adoption curve is the fastest first quarter in this category after the 2026 wave, and the scrutiny record is the exact AstrBot profile at a third of the age: one 1-point Hacker News thread, zero English-language technical coverage, so the vetting so far has happened in Chinese-language channels you are probably not reading.

Strengths
#

  • Multi-user from day one: household and team accounts with JWT isolation, expert sharing, and a shared skill pool, which every other runtime in this category treats as an afterthought or omits.
  • The Tencent enterprise IM matrix (Feishu, DingTalk, WeCom, QQ, WeChat) through a normalizing gateway, alongside Telegram and Discord.
  • ACP in both directions, serving your agent to IDEs and delegating out to coding agents, a symmetry no other member offers.
  • The composable sub-repo architecture (harness, gateway, memory, browser) keeps the pieces independently evaluable.

Cautions
#

  • Zero English-language technical scrutiny: the only independent trace this run found is a 1-point Hacker News thread from 2026-10-06.
  • The first stable release only landed 2026-10-18 after three weeks of beta tags, so pin versions deliberately on a project this young.
  • The one-line installer pulls its script from a Tencent Cloud COS bucket, and the docs, community, and channel gravity are China-first like AstrBot and QwenPaw.
  • The runtime sub-repo (octop-harness) stands at 39 stars, which says the ecosystem is the org’s own code rather than a community.

Pricing
#

N/A, free and open source under MIT, self-hosted. Your costs are model keys and the machine you run it on; a managed-agents cloud option is on the roadmap, not shipped.

Compared to
#

  • AstrBot: the older IM-platform veteran with the wider platform table and the 1,000-plugin marketplace; choose AstrBot for platform breadth and community, Octop for multi-user expert teams and the Tencent-suite connectors.
  • QwenPaw: the other China-channel runtime; choose QwenPaw for the offline path through its own trained small models and the five-layer security stack, Octop for multi-user accounts and ACP.
  • OpenClaw: the single-operator ecosystem root; choose OpenClaw for channel and companion depth, Octop when the assistant must serve a household or team with separate accounts.

Bottom line
#

Recommended for households and small teams whose assistant must live in Feishu, DingTalk, QQ, or WeChat, and who want per-user expert isolation under an MIT license. Not for anyone who needs independent security scrutiny before adopting, or a Western-IM-first deployment. The disagreeable claim I will defend: multi-user is the assistant category’s next battleground, and an incumbent vendor shipped it first while every community favorite still assumes one operator.

Changes
#

  • 2026-10-07 - Created from the entrant scan as the category’s fifteenth member.
  • 2026-10-11 - Corrected the PyPI claim (v1.0.2b6 is now published there while the stable channel stays at 1.0.1), with refreshed adoption numbers.
  • 2026-10-13 - Recorded v1.0.2b7 (2026-10-12) extending the beta line on GitHub and PyPI, with refreshed adoption numbers.
  • 2026-10-16 - Recorded v1.0.2b8 (2026-10-14, on PyPI the same day) extending the beta line, with refreshed adoption numbers.
  • 2026-10-17 - Recorded v1.0.2b9 (2026-10-17, on PyPI the same day) extending the beta line, with refreshed adoption numbers.
  • 2026-10-19 - Recorded v1.0.2 (2026-10-18), the first stable release, published on GitHub and uploaded to PyPI the same day, graduating the beta line, with refreshed adoption numbers.
  • 2026-10-25 - Recorded v1.0.3b1 (2026-10-25, a default expert at user creation, permissions split per page tab) opening the v1.0.3 line on GitHub and PyPI the same day, with refreshed adoption numbers.

See also
#

References
#