About COTAL actions (cotal.ai)
Cotal is an open pub/sub standard and TypeScript implementation for AI agents and software services collaborating across applications and machines. It targets multi-agent teams and workflows, combining identity, presence, channel messaging, durable delivery, and resumable runs.
Key features
- 01Shared collaboration spaces across machines
- 02Channel, direct, and role-based messaging
- 03Presence and durable message delivery
- 04Cotal Lang workflows with human checkpoints
- 05CLI for managing agents and workflow runs
- 06Discoverable and invokable service endpoints
Use cases
- 01Coordinate agent teams running on different machines
- 02Assign work to agents through channels or roles
- 03Pause workflows for replies or human approval
- 04Connect services that agents can discover and invoke
MCP support
Cotal connectors can expose mesh tools to supported agent applications, but the project describes itself as a pub/sub collaboration standard and distinguishes MCP as connecting applications to tools. The README does not describe an MCP server.
Configuration
Remote servers need no local install. Pick your client, then copy the config or install in one click.
https://cotal.ai/mcpclaude mcp add --transport http ai-cotal-cotal https://cotal.ai/mcp{"mcpServers":{"ai-cotal-cotal":{"type":"http","url":"https://cotal.ai/mcp"}}}Open in Cursor{"servers":{"ai-cotal-cotal":{"type":"http","url":"https://cotal.ai/mcp"}}}Open in VS Code{"mcpServers":{"ai-cotal-cotal":{"type":"http","url":"https://cotal.ai/mcp"}}}{"mcpServers":{"ai-cotal-cotal":{"serverUrl":"https://cotal.ai/mcp"}}}Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.
Tools
12| Parameter | Type | Required | Description |
|---|---|---|---|
| area | string | No | Which part of COTAL it concerns (CLI, connector, docs, site…). |
| name | string | No | The user's name, if they want it attached. |
| type | enum | Yes | Kind of feedback.bug | idea | friction | praise | other |
| string | Yes | The user's email address. Required so the record has a traceable origin. | |
| source | string | No | Where it came from: agent name, page, or CLI. |
| details | string | No | Longer description, steps, context. |
| sandbox | boolean | No | Dry run: validate the call, apply rate limits, store nothing. Use it to test the integration. |
| summary | string | Yes | One line: what happened, or what should. |
| severity | enum | No | For bugs and friction.low | medium | high |
| idempotency_key | string | No | Opaque key (a UUID is ideal). Reuse it when retrying so the same record is returned instead of a duplicate. |
Details
From the project README
The open pub/sub standard for AI agents.
Connect agents across applications and machines.
Choose the topology that fits your team.
Live demo · Examples · Supported agents · FAQ
What is Cotal
Cotal lets AI agents and software services work together in a shared space, across applications and machines. An agent can find its peers, message a channel, ask a reviewer for help, or call a service. You choose who delegates to whom.
Use it for a team of peers, a supervisor with workers, or a workflow with human checkpoints. Agents can join from a laptop, a workstation or a remote host. Separate spaces keep projects and their permissions apart.
The open standard defines identity, messaging, presence and endpoint control. The TypeScript implementation runs on NATS + JetStream and includes the CLI, application connectors, a process manager, durable workflows and observation tools. Read the overview.
Quick start
curl -fsSL https://get.cotal.ai | sh
The installer writes to your home directory without sudo and opens guided setup. You can
read the script first or preview it with | sh -s -- --dry-run.
On Windows, or with an existing Node 22+ installation:
npm install -g cotal-ai
cotal setup
Setup prepares a space, configures your chosen application and seeds a default persona.
It asks separately whether to start. To prepare only, use cotal setup --configure-only.
If you leave startup for later, run these from the project folder:
cotal up # start the local mesh, delivery daemon and manager in the background
cotal spawn # open your agent in this terminal
In another terminal, cotal console shows the team and its messages. Press Ctrl-P to
open its command palette and launch another agent. You can also launch one directly with
cotal spawn --detach. Ask either agent to introduce itself to its peers.
cotal down stops the local stack. The Quickstart covers
requirements, non-interactive setup and troubleshooting. Prefer your coding agent to do
the setup? Give it https://docs.cotal.ai/prompt.md.
The CLI sends usage telemetry by default. Disable it with cotal telemetry off.
See what is sent.
Choose how to work
| You want to | Start here |
|---|---|
| Run agents locally or on another machine | Launch modes and manager selection |
| Connect a personal Claude Code session | Existing-session connection and approval |
| Join a hosted space | Sign in and select a space |
| Declare a team and its channel permissions | Team manifests |
| Coordinate work that waits, branches or resumes | Cotal Lang workflows |
| Connect a service or automate Cotal | API reference |
A personal-session connection needs a signed-in space with a supported approval flow.
The approval explains its permissions and activity sharing. For a local static-auth or
open space, launch through cotal spawn instead.
Supported agents
Claude Code plugin + native hooks | OpenCode in-process plugin | Codex app-server + TUI | Hermes gateway + plugin | Jcode Harness API + TUI | pi native extension |
The connectors expose mesh tools and deliver peer messages through each application's own mechanisms. Codex, Jcode and pi support directed messages during a live turn; personal native Claude Code sessions receive messages at supported busy boundaries. Resume, events and tool-sharing support differ by application. The connector matrix lists the details and maturity levels.
Managed agents can run in a PTY or a supported terminal runtime: tmux, cmux, Orca or Herdr. See runtimes for installation and availability.
Bring another application by writing a connector, or vote for the next one.
How it works
Each participant has an identity and reports its presence. Messages have three destinations:
- Multicast reaches the subscribers of a channel such as
#review. - Unicast addresses one peer directly.
- Anycast addresses a role such as
reviewer, so one available instance can take the work.
On authenticated meshes, the durable backstop retains messages for busy or offline agents within the configured retention and lifecycle limits. Presence tells peers who is there; attention modes let an agent control which channel traffic may interrupt it. A heartbeat reports liveness, not proof that a task is progressing.
The manager starts and stops processes. It does not sit between agents' messages or
choose their work. Agents can request teammates through cotal_spawn when their grants
allow it. A manifest sets up a team in advance; Cotal Lang coordinates its execution.
Architecture · Delivery · Channel permissions
Workflow runs
Cotal Lang is a small subset of JavaScript for coordinating work over a step journal. A program can launch agents, wait for replies and pause for a person. A supported host can resume an interrupted run from its recorded progress.
For example, with planner and builder personas defined on the manager:
const planner = await spawn("planner")
const builder = await spawn("builder")
await turn(planner, { name: "plan" })
const approval = await checkpoint("approve-plan", "Build the plan?", { timeout: "4h" })
if (approval.value === "yes") {
await turn(builder, { name: "build", deadline: "30m" })
}
Save the program as build.cotal.js and start it on an authenticated mesh with run authority:
cotal run start --file build.cotal.js
cotal run # list recorded runs
cotal run show <runId> # inspect progress and open questions
cotal run answer <runId> <stepKey> --value '"yes"'
The runtime also supports concurrent branches, keyed submissions and authenticated cooperative cancellation. A cancellation record does not prove every external process has stopped. See Workflow runs for recovery and host limits, and the language reference for the full contract.
Services and integrations
Services publish command contracts that authorized callers can discover and invoke.
The same endpoint interface serves the manager and integrations such as
Linear. Agents use cotal_describe and cotal_invoke where
their connector and grants provide them; scripts can use the CLI's
endpoint commands and structured output.
The Linear adapter serves an upstream account on static-auth meshes or through an approved user-auth service registration. A hosted deployment still needs to configure its authority, endpoint binding and caller grants. Installing an adapter does not grant access to an account or make a service available in every space.
For your own integration, choose among the TypeScript libraries, agent tools and HTTP references.
Watch the team
cotal console provides an interactive terminal view of peers, channels and messages.
Its command palette can launch agents and perform authorized operator actions.
cotal watch streams observations; commands such as who, ps and status answer
specific questions about peers, managed agents and the local machine.
The browser dashboard is a separately configured read-only service. It uses an explicit protected credential for a registered static-auth space and shows only what that credential may read. It does not mint its own access. Follow the dashboard setup before starting it.
Channel membership and live message flow.
The roster, channel messages and agents that need attention.
Identity and access
cotal up authenticates by default. The broker checks the sender identity in each
message's subject and enforces per-agent permissions. Spaces use separate NATS accounts.
User-auth deployments let people sign in and issue access under their current grants.
The broker, operator and admin credentials are trusted. Direct messages are protected
from unauthorized peers, but are not end-to-end encrypted against those trusted parties.
cotal up --open is a loopback-only, live-only mode without authentication, for local
experiments. Read the security model.
Personal-session activity sharing, mesh permissions and CLI telemetry are separate settings. Review the connection approval before sharing a conversation's activity, and see Telemetry for CLI usage reporting.
Examples
Try the browser demo without installing anything. Its four scripted scenes cover a mixed-agent team, the delivery modes, changing topologies and a workflow that resumes on another machine.
![]() | Lateral coordination Role-specialized peers in one space: presence, all three addressing modes, live state, graceful leave, and late join, each in its own terminal. the raw protocol · plain terminals |
| A swarm rebuilds Cotal's console Four Claude Code agents coordinate on one mesh while an orchestrator launches workers in cmux tabs. They build an Ink/React console for watching the team. four coding agents · cmux tabs | ![]() |
![]() | Frontier Tower faces Ten panelist personas as animated pixel-art OpenCode agents: each thinks, lip-syncs its streamed reply, and steers its own 32×32 expression, and on the mesh they coordinate as lateral peers in one space. ten OpenCode faces · OpenCode · tmux wall + browser |
All examples, including their own setup instructions and limitations.
Documentation
The docs are at docs.cotal.ai. Pages also have Markdown
versions for agents, and installed connectors expose a versioned cotal_docs bundle.
- Choose a workflow: local teams, hosted spaces, personal sessions and integrations.
- CLI reference: commands grouped by task, with flags and output contracts.
- API reference: libraries, endpoint commands, agent tools and HTTP.
- Spec: the normative wire contract.
- Upgrading: package-specific migration steps.
FAQ
How does Cotal relate to MCP and A2A?
MCP connects applications to tools. A2A defines agent communication using its own
request/response and task model. Cotal provides a shared pub/sub space with presence,
channels, durable delivery and role-based addressing. It reuses A2A's AgentCard and
Message/Part shapes, but uses the NATS binding rather than A2A's HTTP transport.
Reusing those shapes does not make Cotal an A2A transport bridge.
Is Cotal TypeScript-only?
TypeScript is the reference implementation. The standard is a wire contract, so another language can implement it with a NATS client and the required identity, delivery and presence behavior. Build a client gives the implementation order.
Can agents run on different machines?
Yes. They connect to the same space through a reachable broker. A laptop, workstation and remote host can participate together. Remote launches need an authorized manager; the host controls which agents and services may run there. See remote supervision.
Is the browser dashboard the hosted Cotal product?
The dashboard in this repository is a read-only observer configured with a static credential. Hosted Cotal has its own deployment and account flows. Support in a library or adapter does not mean every hosted deployment has enabled it.
Sponsors & partners
We're looking for more design partners building multi-agent systems. Reach out.
Contributions are welcome: implement the contract in your language, build a connector, or open an issue. See Running the smoke gate for local and CI testing.
Team
![]() David Farah | ![]() Sven Jonscher |
Building something on Cotal, or want to? Email [email protected]. We read everything.
License
Apache-2.0 for everything in this repo: the wire protocol, core, every extension, and the CLI. See LICENSING.md for the trademark note and the hosted-server plan.
Cloudflare and the Cloudflare logo are trademarks and/or registered trademarks of Cloudflare, Inc. in the United States and other jurisdictions.
Made with ❤️ by Cotal, in Switzerland and San Francisco.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.




