Tiger CoWork v0.7.2
A self-hosted AI workspace with chat, code execution, parallel multi-agent orchestration, cross-machine agent connection, and a skill marketplace. Mix different AI providers in the same agent team — OpenAI-compatible APIs, Claude Code CLI, and Codex CLI. Connect agents across machines on your network so distributed teams can collaborate in real time. Connect external MCP servers to extend the AI's toolbox. Built with 16 built-in tools and designed for long-running sessions with smart context compression and checkpoint recovery.
What's New in v0.7.2
- Crash-Safe Data Storage — every JSON store (settings, chat history, tasks, projects, skills) is now written atomically (temp file + rename) behind a per-file lock, so a crash mid-write can no longer truncate or wipe a store. Corrupt files are quarantined to
*.corrupt-<timestamp>and logged instead of being silently overwritten with defaults, and concurrent writers use read-modify-write updates so they can't clobber each other's changes. - Server Crash Prevention — the socket chat handlers are wrapped so an error thrown before their inner try/catch reports back to the client and keeps the server running instead of taking down the whole process; added top-level
unhandledRejection/uncaughtExceptionguards and a fatal-startup catch. - Correct Checkpoint Resume — resuming a long run from a checkpoint now appends your new message (it previously answered the previous question), expires stale checkpoints after 30 minutes, and clears checkpoints on terminal errors so they can't hijack the next turn.
- Session & Resource Cleanup — killing a task now tears down its realtime agent session (no more follow-up messages hanging on a dead session), TCP protocol channels are actually released on cleanup, an idle sweeper reclaims abandoned realtime sessions, and the web-search temp files that once filled
/tmpare removed on every path. - Hardened Path Validation — workspace path checks reject sibling-directory escapes, agent-config deletion and skill-zip extraction are guarded against path traversal / zip-slip, and log endpoints validate the session id before touching the filesystem.
- Client Reliability — API error responses are now surfaced as errors instead of crashing callers; the project chat can no longer get stuck in a permanent "loading" state after a missed event; and the app shares a single websocket connection instead of opening one per view.
What's New in v0.7.1
- Context Compaction Reliability — fixed a class of
"chat content is empty (2013)"errors that surfaced asContext overflow after 3 compression retriesin long pipeline runs. Emergency retries now bypass the 60-second compaction cooldown;trimConversationContextis tool-pair aware (assistanttool_callsand their matchingtoolresults are dropped as a unit so nothing gets orphaned); a newtruncateLargestToolResulthelper targets the single biggest tool message on retries 2–3 instead of halving the whole transcript; and a structuralvalidateMessageStructurepass runs after every compression/trim to drop any orphan that slipped through. - Auto-Skill Run Visibility — when the synthesiser returned
{"proposals":[]}, the run summary previously showedcreated=0 updated=0 skipped=0 candidates=Nwith no indication of why nothing happened. Empty-proposal outcomes now surface inskillAutoUpdateLastRunSummaryasLLM returned no proposals for N session(s), and the LLM's actual reply head is logged to the server console for debugging. - Stricter
looksEmptyHeuristic — the prose-fallback that coerces a refusal-shaped reply (e.g. "no skills worth capturing") intoproposals=[]now requires the whole trimmed reply to be under 200 characters. Previously a long, malformed proposal that happened to contain "no skill" inside its rationale could be silently swallowed.
What's New in v0.7.0
- Automatic Skill Generation — successful chats are mined on a cron (or via Run Now) to synthesise reusable
SKILL.mdworkflows. The synthesiser proposes new skills or refines existing auto-generated ones; updates land as a side-by-sideSKILL.md.proposeddiff for review before going live. See Automatic Skill Generation for the full algorithm and settings. - Skill Approval Workflow — pending auto skills surface in the Skills page with an Approve/Reject control. Approval renames the proposed file to
SKILL.mdand enables it; the next user turn in any chat picks it up automatically (the system prompt is rebuilt per turn — no restart needed). - Per-Message Feedback Loop — opt-in 👍 / 👎 + comment on every assistant reply. 👍 tells the synthesiser "capture this procedure"; 👎 with a comment triggers a focused remediation pass that picks the auto-skill most likely responsible (chat-log-driven targeting first, token-overlap fallback) and rewrites just that one skill. 👎 alone is treated as "do not distil from this chat."
- Sub-Agent–Aware Auto-Skills — the pipeline now parses
data/chat_logs/<sessionId>.logto surface a per-agent workflow trace (task, tools used, skills loaded). For multi-agent chats this fixes a previous gap where only the orchestrator's merged reply was visible to the synthesiser; remediation can now identify and rewrite the exact auto-skill a sub-agent loaded, and synthesis can capture the agent topology in the SKILL.md body — not just the outcome.
What's New in v0.6.1
- Per-Project Agent Mode Override — each project can override the global sub-agent mode (Auto Spawn, Auto Create, Manual, Realtime, Auto Swarm) and pick its own YAML config, architecture type, agent count, and connection protocols. The active override is shown as a clickable purple tag in the project header and chat banner.
- Auto Architecture — AI-Decided Settings — new "Auto (AI decides)" option for architecture type and agent count (3–8 default). Connection protocols are now multi-select toggle buttons instead of a single dropdown.
- Full Chat Log with Agent Reasoning — every chat session now records a complete log file capturing user messages, tool calls (with arguments), sub-agent reasoning text, and final responses. New Log button next to Activity opens a live-updating panel; new Export button downloads the log as
.txt. - Finished Tasks History — Tasks page now shows the last 100 completed/cancelled/errored tasks with status, duration, agents used, and tools called. Open Chat button on each finished task jumps directly to that session.
- Project List Sorting — sort projects by A–Z or Recent (most recently updated). Sort preference persists across reloads.
- Sub-Agent Reasoning in Chat Log — orchestrators and worker agents stream their intermediate thinking text to the chat log between tool calls, giving full visibility into the decision-making chain.
- AsyncLocalStorage Settings Override — project agent overrides now propagate correctly through every async call in the backend, ensuring
getSettings()returns the project-scoped configuration throughout the entire chat lifecycle.
Warning: This app executes AI-generated code and shell commands. Run it inside Docker or a sandboxed environment. See Security & Docker Setup.
Screenshots
AI Chat with tool-calling — generates React/Recharts visualizations rendered in the output panel.
Visual Agent Editor — drag-and-drop multi-agent design with mesh networking and YAML export.
Minecraft Task Monitor — live pixel-art agents with speech bubbles, walking animations, and inter-agent interactions.
Key Features
- AI Chat with 16 Built-in Tools — web search, Python, React, shell, files, skills, sub-agents
- Mix Any Model per Agent — assign different AI providers per agent (API, Claude Code CLI, Codex CLI)
- Parallel Multi-Agent System — 7 orchestration topologies, 4 communication protocols, P2P swarm governance
- Cross-Machine Agent Connection — connect agents running on different machines over the network, enabling distributed multi-agent collaboration across your infrastructure
- Minecraft Task Monitor — live pixel-art characters (Steve, Creeper, Enderman, etc.) with speech bubbles showing agent activity, walking animations when agents interact
- Long-Running Session Stability — sliding window compression, smart tool result handling, checkpoint recovery
- MCP Integration — connect any Model Context Protocol server (Stdio, SSE, StreamableHTTP)
- Output Panel — renders React components, charts, HTML, PDF, Word, Excel, images, and Markdown
- Skills & ClawHub — install AI skills from the marketplace or build your own
- Automatic Skill Generation — distil successful chats into reusable
SKILL.mdprocedures with an LLM synthesiser, approval gate, and proposed-diff review - Projects — dedicated workspaces with memory, skill selection, and file browser
Installation
One-Click Installers
Mac:
- Download
TigerCoWork.zip - Unzip, right-click
TigerCoWork.appand select Open
Windows:
- Download
TigerCoWorkInstaller.zip - Unzip and run
TigerCoWorkInstaller.bat
Prerequisite: Docker Desktop must be installed and running.
Terminal Install
Mac/Linux:
curl -fsSL https://raw.githubusercontent.com/Sompote/tiger_cowork/main/install.sh | bash
Windows (PowerShell):
irm https://raw.githubusercontent.com/Sompote/tiger_cowork/main/install.ps1 | iex
Manual Install
Log in to your Linux server directly or via SSH:
ssh root@<your-server-ip>
⚠️ Security Warning: AI agents can execute arbitrary code and shell commands that may modify or delete files on the host system. It is strongly recommended to run Tiger CoWork on a VPS or dedicated machine that contains no important data. Do not run it on a machine with sensitive or irreplaceable information.
Prerequisites: Node.js >= 18, npm, Python 3 (optional)
git clone https://github.com/Sompote/tiger_cowork.git
cd tiger_cowork
bash setup.sh # installs deps, prompts for ClawHub token
npm run build && npm start # → http://localhost:3001
Running in background (recommended): Use PM2 to keep Tiger CoWork running after you close the terminal.
npm install -g pm2 # install PM2 globally npm run build # build production bundle pm2 start npm --name "tiger-cowork" -- start # start in background pm2 save # save process list for auto-restart pm2 startup # enable auto-start on system bootUseful PM2 commands:
pm2 status # check running processes pm2 logs tiger-cowork # view logs pm2 restart tiger-cowork # restart pm2 stop tiger-cowork # stop
First-time token: The default token is
your-secret-token-herein the UI. Please change it later in.envfor your security.
Quick Start
- Open
http://localhost:3001 - Go to Settings → enter your API Key, API URL, and Model
- Click Test Connection to verify
- Start chatting — the AI can search the web, run code, generate charts, and more
Automatic Skill Generation
Tiger CoWork can automatically synthesise reusable SKILL.md workflows from your past chat sessions. After a chat finishes, an LLM-driven pipeline reviews recent conversations and either proposes a brand-new skill, refines an existing auto-generated one, or — when 👎 feedback was left — surgically rewrites the auto-skill most likely responsible. Future chats then match a stored procedure instead of re-improvising from scratch.
How it works
- Schedule. A cron job (default 60 min, configurable) wakes the pipeline. You can also fire it manually from the Skills page (Run Auto-Update Now). Manual runs ignore the cursor and re-evaluate the newest sessions against current skills.
- Candidate selection. Sessions whose
updatedAtis newer than the saved cursor are summarised: first 6 user prompts (≤600 chars each), the last assistant reply (≤3000 chars), any per-message 👍 / 👎 feedback with comments, and a sub-agent workflow trace parsed fromdata/chat_logs/<sessionId>.log(per-agent task, tools used, skills loaded viaload_skill, completion/error status). The trace is necessary because in multi-agent modes only the orchestrator's merged final reply lands insession.messages— the chat log is the only place that knows which sub-agent did what. Sessions whose final reply looks like an error (rate limit, missing API key, connection error) are silently dropped, as are sessions with no user or assistant turn yet. - Pass 1 — Remediation (👎 with a comment). For each session containing an assistant message marked 👎 with a non-empty comment, the pipeline picks a target auto-skill in this priority order: (a) chat-log-unique — exactly one auto-skill was loaded during the session → use it directly; (b) chat-log-subset — multiple auto-skills loaded → restrict the token-overlap heuristic to that subset; (c) token-overlap fallback — no
load_skillevents found → match against all auto-skills using the disliked excerpt + comment (minimum score = 2 distinct content tokens). The model is then asked to rewrite that one skill to address the complaint, or{"skip": true}if it judges the wrong target was picked. The remediation prompt also includes the parsed sub-agent workflow so the model can correct orchestration-layer failures (wrong sub-agent, wrong order, missing prerequisite). Capped at 5 remediations per run; the rest defer to the next tick. Successful remediations consume the session, so Pass 2 will not also distil from it. The run summary records which strategy picked the target (e.g.target picked via chatlog-unique) for debuggability. - Pass 2 — Synthesis. The remaining candidates are sent to the TigerBot model with the current skill list and the per-session sub-agent workflow block. The synthesiser returns strict JSON proposals, each either:
create— a new skill (must not collide with any existing name across all sources), orupdate— refines an existing skill only if itssource: "auto". Human-curated skills (, , , ) are write-protected. The synthesiser is told to skip casual chats / one-off Q&A that don't generalise — quality > quantity. 👍 reinforces "capture this"; 👎 explicitly suppresses distillation. When is non-empty, the prompt instructs the synthesiser to capture the agent topology (which roles to spawn, in what order, with which skills loaded) in the SKILL.md body — not just the outcome.
Per-message human feedback
When skillAutoUpdateHumanFeedbackEnabled is on, every assistant message in chat picks up a 👍 / 👎 control plus an optional comment box. Submitting feedback bumps the session's updatedAt, so the auto-skill loop will re-consider that session on its next pass.
Settings (in data/settings.json)
Toggling any of these from the Settings page reconciles the cron job immediately — no restart needed.
REST endpoints
POST /api/skills/auto/run-now— trigger the pipeline immediately (ignores cursor)POST /api/skills/:id/approve— accept a pendingcreateorupdatePOST /api/skills/:id/reject— drop a pending proposal (deletes the folder for acreate, deletes only.proposedfor anupdate)GET /api/skills/:id/proposed-diff— return current vs. proposedSKILL.mdfor reviewGET /api/skills/:id/content— read the liveSKILL.mdplus its supporting-file listGET /api/skills/:id/download— download a singleSKILL.mdor a zipped folder of the whole skillPOST /api/chat/sessions/:id/messages/:index/feedback— set/clear 👍/👎 + comment on one assistant message ({rating?: "up"|"down", comment?: string, clear?: true})
When does a new skill become callable?
Skills are advertised to every agent in the topology — orchestrator, spawn_subagent workers, and realtime/auto-swarm/auto-create agents — by buildEnabledSkillsBlock, which re-scans the skills directory and registry on every prompt build. Once a skill flips to enabled: true (i.e. after approval), the next user turn in any chat sees it, including any sub-agents the orchestrator spawns. No server restart needed.
A skill cannot be applied within the same chat that produced it: synthesis only runs after a session settles, and that session is the pipeline's input, not its consumer. The earliest a skill distilled from chat A can be invoked is the next user turn that follows the next cron tick (and approval click, if required).
Documentation
License
This project is licensed under the MIT License.
