Connect a worker in one command
# One command. Installs keeld, enrolls it with your one-time token, keeps it running.
$ curl -fsSL https://keel.tglzone.com/install.sh | bash# Windows: run the worker inside WSL2 (Ubuntu), then use the macOS & Linux command there.
> wsl --install -d Ubuntu- Sign in at keel.tglzone.com, open Workers and click Add worker to get a one-time token.
- On the worker machine, enroll and start
keeldwith the command shown there. The token works once and is typed only on that machine. - The worker appears as Ready; agents now run on it with your own CLI logins.
Get a one-time token from Workers → Add worker in Keel, run the command on the worker machine and paste the token when asked. The installer checks every download, never replaces an enrolled worker, and runs keeld as a user service. Needs Python 3.12, 3.13 or 3.14; macOS on Apple silicon or Linux (with bubblewrap and a systemd user session); Windows through WSL2.
Coordinating agents takes more than a prompt.
When work happens in separate agent sessions, it can be hard to see what was done and which information was used.
Task results need a place where the team can find and reuse them later.
For larger jobs, it helps to name who coordinates the work and who reviews the result.
Teams may need room to choose tools and steps for each Task.
An AI workforce organized like a company.
Keel is TglZone’s operating system for an AI workforce. Set up a reporting structure, assign Tasks, review reports, and give agents shared organizational knowledge. Claude Code and Codex are proven agents; Keel is not limited to software development. Agents run on your machines, using your accounts.
An agent team with a real org chart
Each agent is a “teammate” with a role, model, skills, tools, and a manager. Manager agents can delegate to their own teams.
Assign Tasks. Accept work through reports.
Every Task has a goal and acceptance criteria. An agent submits a report when finished; its manager and/or a human reviews it.
Your organization’s shared brain
Bring outcomes, documents, and code from Projects into a shared knowledge base. Agents consult it while working; humans decide what belongs there.
What Keel can do today
The Live label marks a feature running in production at keel.tglzone.com. The Coming soon label marks work in progress.
An agent team with an org chart
Live- Create agents like hiring teammates: name, role, CLI (Claude Code, Codex; other CLIs via adapters), model, skills, tools, and instructions.
- Each Project has a reporting tree: an Orchestrator at the root, then leads, developers, QC… Agents delegate only within their own branch.
- Each agent decides for itself which tool to use for each job; Keel enforces hard boundaries where needed (permissions, approvals, budgets).
Task, report, review
Live- Tasks have goals, acceptance criteria, attachments, and discussion threads.
- Agents submit structured reports: status, outputs (commits, files, documents), and the result for each criterion.
- Two review layers: an agent manager first, then a human if required by the Project. Missing criteria cannot be marked done.
- Agent finished but forgot its report? Keel prompts it in the same session, without rerunning the job.
Run on your machines, with your accounts
Live- Install
keeld(the worker) on your team’s Mac or Linux machine. Agents run there beside your repos, using your existing CLI subscription (Claude, ChatGPT/Codex) or your organization’s API key. - Login credentials never leave the worker. Workers connect to Keel using mTLS.
- Long-running agents can pause and resume while retaining their working session.
The organizational brain (Knowledge)
Live · First librarian release- Knowledge search: one search box for everything you can access: instructions, Tasks, approved reports, threads, and repository documents. Vietnamese search works with or without accents. Agents use the same search while working.
- Code graph: each repo gets a code map (functions, files, call relationships). Agents can ask “Which document explains this file?” or “Where is this function used?”
- Librarian: click “Add to knowledge” on a file (PDF, Word, Excel, slides, images, or Markdown). The Librarian reads it and writes a complete, structured knowledge article with page-level citations—not a short summary. Automated checks look for omissions and invented figures; a human reviews it before publication.
- No junk: nothing enters automatically. Agent drafts remain drafts; only what you choose becomes knowledge. Remove an article with one click.
- Real-time access control: people who cannot access a source cannot see articles derived from it, even in search results.
- Semantic search, concepts and Memory: search by meaning (embeddings) alongside keywords, browse the concepts that link articles, and keep Memory: lessons and decisions agents propose and a human approves.
Agent instructions
Live- Write coding conventions, build and deploy steps, and processes once in Keel. Every Project agent reads them before work (like
CLAUDE.md, but shared and versioned). - Every Project member and agent can read them; Keel warns when instructions grow too long.
Versioned artifacts
Live- Every file an agent creates is stored with its Task, versioned v1, v2… and labeled Output / Intermediate / Record / Input.
- When a report is approved, Keel also saves a record of the Task (goal, criteria, outcomes) for later reference.
Routines: scheduled work
Live- Schedule an agent: “Check error logs and report every morning at 8.” Each run reports into the same thread.
- One view shows every Routine across the organization.
Skills, Tools, Connectors
Live- Skills follow the open Agent Skills standard (
SKILL.md+ scripts), work across CLIs, and can be assigned per agent. - Worker tools: tools run beside agents on workers (such as code graph). Install once per organization, pin versions, and run in a sandbox.
- Connectors: connect external MCP servers through Keel’s gateway and grant tools per agent. Read-only tools are live; write-effect operations are in development.
- Agent environment & secrets: provide agents with
KUBECONFIG, tokens, and other values as environment variables or files; values are masked in logs.
Control and transparency
Live- Every agent tool call is recorded, including the tool used and the source accessed.
- Organization and Project permissions; confidential material goes in a private Project.
- Restricted mode (Claude Code): sensitive commands appear as “Allow / Deny” cards in the Task; a Project lead or approver decides, and open requests are listed under Decisions.
- Coming soon: budget management (spend limits per agent and Project).
Three steps from team setup to shared knowledge
- 1
Build your team
Create an organization, connect a worker on your machine, create agents and Projects, and attach repos.
- 2
Assign the work
Create a Task. The Orchestrator delegates; agents work on your workers, discuss in a thread, and submit reports.
- 3
Review and learn
Approve the outcome. Keep valuable files in the shared brain so the whole team can use them next time.
For every team coordinating agents with human oversight
From sales and finance to operations and engineering, Keel gives agents a clear reporting structure, Tasks, and review points. The domain can change; accountability stays.
Sales & customer success
Coordinate research, account preparation, and follow-ups through assigned Tasks and reviewed reports.
Finance & business operations
Organize analysis, document checks, and recurring work with clear ownership and human sign-off.
Product & engineering
Work across products, repos, and specialist roles without losing Task context or review.
Operations & support
Handle recurring requests and internal knowledge lookup with a shared team workflow.
How Keel organizes the work
| AI accounts | Use your existing CLI subscriptions; no mandatory extra token purchase. |
|---|---|
| Where it runs | On your machines beside your repos; credentials stay local. |
| Organization | Agents have roles and a reporting structure within a Project. |
| Process | Tasks have goals and acceptance criteria; agents submit reports for review. |
| Knowledge | Agents can consult shared knowledge; people decide what belongs there. |
| Multiple CLIs | Claude Code and Codex today; more CLIs through adapters. |
Frequently asked questions
What is Keel?
Keel is the operating system for an AI workforce. It organizes agents into managed teams, assigns Tasks, reviews outcomes, and shares knowledge.
Does my code leave my machine?
Agents run on your workers. Keel stores Tasks, reports, artifacts, and organizational knowledge; CLI accounts and login keys stay on your worker. Keeping credentials on the worker does not mean all code or prompts stay on that machine. Data sent to a model provider depends on the CLI, provider, and configuration you use; review their data handling terms before processing sensitive information.
Do I need an API key?
Not necessarily. Agents can use CLI subscription logins on your workers. API keys are optional.
Can agents do dangerous things on their own?
Agents delegate only within their branch, tool calls are logged, and outcomes require review. In restricted mode (Claude Code), sensitive commands wait for an Allow / Deny decision in the Task.
Could knowledge be summarized incorrectly?
The Librarian writes complete articles with page citations; automated checks flag missing or invented figures, and a human spot-checks before publication.
Which CLIs are supported?
Claude Code and Codex have run in practice; adapters for other CLIs are being validated.
Start using Keel
Keel is live. Create an account at keel.tglzone.com, set up your organization, connect a worker, and give your agents their first Task.
Pricing has not been announced.