Model Context Protocol

Connect icōd to the AI ecosystem

Enable Claude, ChatGPT, and any AI agent to read your projects and take real actions in your icōd apps.

What is MCP?

A standard port for AI

MCP is an open protocol that lets AI applications connect to external systems. Think of it as USB-C for AI applications — just as USB-C gives every device one standardized way to connect, MCP gives every AI one standardized way to reach your tools and data.

Open, versioned standardNot tied to any one vendor or model.
Any AI client connectsOne implementation, every compatible client.
Scoped secure accessPer-token roles and per-tool permissions.
Bidirectional by designServers can read from and act on your behalf.
Possibilities

What it unlocks

Once icōd speaks MCP, every AI client you already use becomes an operator of your infrastructure.

AI-Powered Automation

Deploy, roll back and reconfigure with natural language. A single sentence triggers a real, audited server action.

Intelligent Code Generation

Claude writes code that calls icōd APIs correctly, because MCP hands it your real project structure and conventions.

Context-Aware Assistance

Agents pull deployment history, environment configs and live metrics before they answer — so suggestions match reality.

How icōd + MCP works

One prompt, five hops

Follow a real request end to end. Select any step to inspect the message on the wire.

Capabilities

What can icōd MCP do?

Three primitives cover the whole surface — read state, change state, and standardize the way you ask.

Resources

Read-only data
  • List all apps and projects
  • Get deployment history
  • Read performance metrics
  • Access environment configs
  • View error logs

Tools

Perform actions
  • Deploy app to production
  • Create a new project
  • Update configuration
  • Roll back a deployment
  • Manage environment variables

Prompts

AI-optimized templates
  • Code generation workflow
  • Debugging helper
  • Performance analysis
  • Migration guide
Quick start

A working server in 15 lines

Pick your language. Every SDK exposes the same three primitives.

server.py
from mcp.server import Server
from icod_sdk import ICodTools

server = Server("icod-server")
tools = ICodTools()

@server.tool()
async def deploy_app(project_id: str, version: str = "latest"):
    """Deploy an icōd app to production."""
    return await tools.deploy(project_id, version)

@server.resource("icod://apps")
async def list_apps():
    return await tools.apps()

server.run()
API reference

icōd MCP API

Every capability the server exposes, with parameters and return shapes.

METHODWHAT IT RETURNSRETURN SHAPE
list_apps()Returns every app in the authenticated workspace.App[] · id, name, env, status
get_deployment_status(app_id)Current deployment state, commit and health check result.Deployment · state, sha, url, duration
read_logs(app_id, limit=100)Streams recent log lines, newest last. Supports follow mode.LogLine[] · ts, level, message
get_metrics(app_id, metric_type)Performance and usage series for latency, errors, traffic or cost.Series · points[], unit, window
get_config(app_id, env)Environment configuration with secret values redacted.Config · keys[], env, updated_at
Examples

See it in action

Three reference implementations you can clone and run today.

Claude + icōd

Deployment Agent

Ask Claude to ship your latest changes; it uses MCP to run the deploy and report back with the live URL.

Deploy my latest changes to production
ChatGPT + icōd

Project Creator

Generate a whole project structure — routes, tests and config — with MCP wiring the scaffold into your workspace.

Create a TypeScript API endpoint with tests
Custom Agent

Monitor & Alert

Build your own long-running agent on the icōd MCP server — polling metrics, opening issues and paging you on errors.

Monitor my apps and alert on errors
Patterns

Building reliable MCP servers

Best practices

Do this

  • One responsibility per server — deploys, config and monitoring stay separate.
  • Return structured errors with a code, a cause and a suggested fix.
  • Validate and sanitize every input against a typed schema.
  • Rate-limit resource-heavy operations and long-running tools.
  • Version your capabilities so old clients keep working.
  • Document each tool as if the model were a new teammate.
Avoid

Not this

  • Exposing credentials or tokens through resources.
  • Blocking operations with no timeout or cancellation path.
  • Vague error strings the model cannot recover from.
  • Unsafe file operations outside an allow-listed root.
  • Unvalidated outbound requests triggered by model input.
  • Destructive tools that run without explicit confirmation.
Security

Safe by default

Every MCP session is authenticated, scoped and logged. The model sees only what your access policy allows — never raw credentials.

Encrypted transport

TLS on HTTP, isolated pipes on stdio.

Scoped access control

Per-token roles, per-tool permissions.

Credentials never exposed

Secrets are redacted before responses.

Full audit trail

Every tool call recorded and exportable.

Ecosystem

Part of something bigger

icōd is one server in a growing MCP ecosystem. Discover others, contribute, or bring your questions to the community.

MCP RegistryGitHub RepositoryCommunity BuildsDiscussionsJSON-RPC SpecArchitecture Deep Dive
Questions

Common questions.

Any client that speaks MCP — Claude, ChatGPT, editor integrations, and custom agents built on the official SDKs. The protocol is open and versioned, so new clients can connect without icōd shipping anything new on its side.

No. Every MCP session is authenticated and scoped to a token's roles and per-tool permissions. Secret values are redacted before they're ever included in a response — the model sees configuration keys, never the underlying values.

That depends on your client's settings, not icōd's server — icōd returns structured errors and typed schemas so the client can ask for confirmation before anything destructive runs. We recommend requiring explicit confirmation for deploys, rollbacks, and config changes.

Yes. Resource-heavy operations and long-running tools are rate-limited per token, and every call — successful or not — is written to an exportable audit trail.

Yes — the SDKs in Quickstart are a starting point, not the only option. Anything that implements the MCP specification and calls icōd's APIs correctly will work; see the Quickstart section for a 15-line reference server.

Ready to connect icōd to AI?

Stand up a server in minutes and let your agents deploy, debug and build alongside you.