IDE Integration

EdgeCoder provides two ways to write code with AI assistance: the EdgeCoder Desktop App (the primary, built-in experience) and external IDE integration through an OpenAI-compatible endpoint that works with VS Code, Cursor, Zed, JetBrains, Windsurf, and any other editor that supports custom OpenAI providers.

EdgeCoder Desktop App (Primary)

The desktop application is a Tauri + Svelte app that ships as the primary development surface for EdgeCoder. It connects directly to all local services on ports 4301 through 4304 and provides a fully integrated IDE experience without any external configuration.

Chat-first interface

The default tab is ChatView -- a conversational interface with streaming responses and full conversation history. Conversations are persisted in IndexedDB, so your history survives restarts.

Code blocks in chat responses include "Open in Editor" buttons that send the snippet directly into the built-in editor.

Built-in Monaco editor

The EditorView tab embeds a Monaco editor (the same engine behind VS Code) with:

An integrated chat panel for inline code assistance
A file explorer for navigating your project
Syntax highlighting for all common languages

You can work entirely within the desktop app -- ask questions in chat, open generated code in the editor, and iterate without switching windows.

Additional pages

Beyond Chat and Editor, the desktop app includes:

Page	Purpose
Dashboard	Node status, task metrics, mesh health
ModelManager	Install, remove, and configure Ollama models
MeshTopology	Visual map of connected mesh peers
TaskQueue	View and manage queued, running, and completed tasks
Credits	Credit balance and transaction history
ActiveWork	Currently executing tasks and progress
Settings	Configuration for services, models, and network
LogViewer	Streaming logs from all services
Account	User profile, wallet, passkeys

Storage

IndexedDB for conversation persistence
localStorage for user settings and UI state

Launching the desktop app

bash

cd "/path/to/EdgeCoder"
npm run desktop

The app connects to the local services automatically. Make sure the backend services are running (see Runtime modes and the README quickstart for details).

External IDE Integration

If you prefer to work in your own editor, EdgeCoder exposes an OpenAI-compatible endpoint on port 4304. Any IDE or tool that can talk to the OpenAI API can use EdgeCoder as a backend -- no plugin or extension required, just a base URL.

How it works

Your IDE  --> POST /v1/chat/completions
                      |
                      v
          EdgeCoder Provider Server :4304
                      |
                      v
          IntelligentRouter waterfall
         +------------+-------------+
         v            v             v
    BT-local     ollama-local    swarm
    (free)       (free)         (credits)

The provider server at :4304 accepts standard OpenAI chat/completions requests and routes them through IntelligentRouter automatically:

Priority	Route	Cost	When
1	Bluetooth-local	Free	iPhone in BT Local mode + connected
2	Ollama-local	Free	Local Ollama healthy, <2 concurrent, p95 <8s
3	Swarm	Credits	Mesh token set, local overloaded or slow
4	Edgecoder-local	Free	Always-on stub safety net

Available model IDs

When adding EdgeCoder in your IDE, you can choose which routing mode to use by picking a model name:

Model ID	Behaviour
`edgecoder-auto`	IntelligentRouter picks the best route automatically (recommended)
`edgecoder-local`	Force local Ollama only -- never touches swarm
`edgecoder-swarm`	Force swarm network -- costs credits

Step 1 -- Start the provider server

Open a terminal in the EdgeCoder project root:

bash

cd "/path/to/EdgeCoder"
npm run dev:ide

Keep this running. You should see:

{"level":"info","msg":"Server listening at http://127.0.0.1:4304"}

Security: token, bind address and shell approval

The provider server can read and write your project and run commands, so it is locked down by default (#4).

Bind address. It binds to 127.0.0.1 only. Set IDE_PROVIDER_HOST to change this. A non-loopback host is refused unless IDE_PROVIDER_TOKEN is set explicitly.
Bearer token. Every route except GET /health needs Authorization: Bearer <token>.
- The token comes from IDE_PROVIDER_TOKEN, or from ~/.edgecoder/ide-token.
- If neither exists, the server creates the file on first start (mode 0600, 32 random bytes, base64url).
- The VS Code extension reads the file automatically, or uses the edgecoder.providerToken setting.
Browser and DNS-rebinding protection. Requests carrying a web Origin (http(s)://…) get 403. When the server is bound to loopback, so do requests whose Host is not 127.0.0.1:4304, localhost:4304 or [::1]:4304.
- The desktop webview origins (tauri://localhost, http(s)://tauri.localhost) are allowed.
- Add more origins with IDE_ALLOWED_ORIGINS (comma-separated).
- Desktop in Vite dev mode runs at http://localhost:1420, so start the IDE server with IDE_ALLOWED_ORIGINS=http://localhost:1420 npm run dev:ide.
- The desktop app reads ~/.edgecoder/ide-token through the Tauri command get_ide_token, so the token file is the single source of truth.
run_shell. Every shell command needs your approval in the IDE.
- Commands are parsed into argv and run with shell: false, with variables like *_TOKEN, *_SECRET, *_PASSWORD, *_API_KEY and *_PRIVATE_KEY stripped from the environment.
- Shell syntax (;, &&, |, backticks, $(, redirects) is refused unless approved.
- Commands get an allowlisted environment only: PATH, HOME, USER, SHELL, locale, terminal and temp variables, and LC_*. Add names with IDE_SHELL_ENV_PASS (comma-separated). Everything else, including every token and secret, is dropped.
- Without approval, the following are refused:
  - node/python with inline-code flags (-e, -c, -p, -m, --eval, …) or with no arguments (REPL)
  - npx, npm exec, cargo run and go run
  - binaries given as a path
  - any path argument outside the project root (absolute, ~ or ..)
  - git subcommands that modify anything
- Warning: IDE_SHELL_AUTO_APPROVE=1 still lets the model run allowlisted tools such as npm test or node script.js, and those run project code as you. Only enable it for projects you trust.
- IDE_SHELL_AUTO_APPROVE=1 (local opt-in) skips the prompt only for allowlisted binaries. The default allowlist is npm, npx, node, python3, pytest, git (read-only subcommands), ls, cat, rg and similar. Override it with IDE_SHELL_ALLOW (comma-separated).

Extension endpoints

The VS Code extension uses these endpoints (#36). The full contract is in src/apps/ide/routes.ts.

GET /health (no token)
GET /status (alias of /v1/router/status)
POST /run, POST /run/local and POST /run/swarm with { task, language, maxTokens? }

/run/swarm returns 503 { "error": "swarm_unavailable" } when MESH_AUTH_TOKEN is not set.

Step 2 -- Add as a custom model in your IDE

Cursor

Open Cursor Settings (Cmd+,)
Go to Models
Under the OpenAI section, enable "Override OpenAI Base URL"
Set the base URL to:
```
http://127.0.0.1:4304/v1
```
Set the API key to your IDE token (cat ~/.edgecoder/ide-token). It is sent as Authorization: Bearer <token> and is required.
In the model name field, type edgecoder-auto and press Enter to add it
Select edgecoder-auto as your active model

Cursor calls POST /v1/chat/completions -- EdgeCoder handles it natively.

Zed

Add to ~/.config/zed/settings.json:

json

{
  "language_models": {
    "openai": {
      "api_url": "http://127.0.0.1:4304/v1",
      "available_models": [
        {
          "name": "edgecoder-auto",
          "display_name": "EdgeCoder (auto-route)",
          "max_tokens": 8192
        },
        {
          "name": "edgecoder-local",
          "display_name": "EdgeCoder (local only)",
          "max_tokens": 8192
        }
      ],
      "version": "1"
    }
  },
  "agent": {
    "default_model": {
      "provider": "openai",
      "model": "edgecoder-auto"
    }
  }
}

Or use the UI: Agent Panel -> settings icon -> Add Provider -> OpenAI-compatible -> paste the URL.

Continue.dev (VS Code / JetBrains plugin)

Add to ~/.continue/config.json:

json

{
  "models": [
    {
      "title": "EdgeCoder (auto-route)",
      "provider": "openai",
      "model": "edgecoder-auto",
      "apiKey": "<contents of ~/.edgecoder/ide-token>",
      "apiBase": "http://127.0.0.1:4304/v1"
    },
    {
      "title": "EdgeCoder (local only)",
      "provider": "openai",
      "model": "edgecoder-local",
      "apiKey": "<contents of ~/.edgecoder/ide-token>",
      "apiBase": "http://127.0.0.1:4304/v1"
    }
  ]
}

Or use the in-app config: click the model selector in the Continue sidebar -> Add Model -> OpenAI-compatible -> fill in the URL and model name.

JetBrains (IntelliJ, PyCharm, WebStorm, etc.)

Go to Settings -> Tools -> AI Assistant -> Models & API keys
Click Add Model
Select provider: OpenAI compatible
Set the API endpoint URL to:
```
http://127.0.0.1:4304/v1
```
Enter any API key (e.g. edgecoder)
Enter model name: edgecoder-auto
Click Test Connection -- should return model info
Click Apply

Windsurf (Codeium)

Windsurf's primary AI is hosted. For custom models, use the MCP (Model Context Protocol) server support:

Go to Settings -> Cascade -> MCP Servers
Add an OpenAI-compatible MCP entry pointing to http://127.0.0.1:4304/v1

Alternatively, install Continue.dev alongside Windsurf -- it works in any VS Code fork.

Claude Code (Anthropic CLI)

Claude Code is a terminal agent, not a visual editor, so it does not add custom models the same way. Instead, you can call the EdgeCoder provider directly from scripts or use it as a tool endpoint.

For agentic workflows, the REST API at :4304 is fully scriptable:

bash

curl -s http://127.0.0.1:4304/v1/chat/completions \
  -H "Authorization: Bearer $(cat ~/.edgecoder/ide-token)" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "edgecoder-auto",
    "messages": [{"role": "user", "content": "Write a Python function to debounce a callback"}]
  }' | jq '.choices[0].message.content'

Verifying the connection

Test that the server is up and returning models:

bash

# Health check
curl http://127.0.0.1:4304/health

# Model list (OpenAI format)
curl -H "Authorization: Bearer $(cat ~/.edgecoder/ide-token)" http://127.0.0.1:4304/v1/models | jq '.data[].id'

# Quick inference test
curl -s http://127.0.0.1:4304/v1/chat/completions \
  -H "Authorization: Bearer $(cat ~/.edgecoder/ide-token)" \
  -H "Content-Type: application/json" \
  -d '{"model":"edgecoder-auto","messages":[{"role":"user","content":"say hello"}]}' \
  | jq '.choices[0].message.content, .edgecoder'

The response includes an edgecoder field with route and latencyMs so you can see which backend handled the request.

System-aware chat

All chat requests (both portal and IDE provider) receive a server-injected system prompt that makes the chatbot aware of its own model, installed models, swarm state, routing behavior, and download progress. You can ask "What model are you running?" or "What models are on the network?" and get accurate, real-time answers.

For the IDE provider path, any file-context system messages sent by the editor (e.g. "The user is editing main.py") are preserved and merged after the EdgeCoder system prompt. See the Model Management guide for details.

Download progress

When a model download is in progress, the desktop app displays a slim progress banner in both the Chat and Editor panels. The banner shows the model name, a progress bar, and percentage. You can also query the progress programmatically:

bash

curl http://127.0.0.1:4301/model/pull/progress

Routing behaviour

When does traffic go to swarm?

The router spills to swarm when local Ollama is overloaded (>2 concurrent), too slow (p95 >8s), or unhealthy. Swarm tasks cost credits from your agent account; the fulfilling agent earns them.

When does Bluetooth-local activate?

When your iPhone is running EdgeCoder with Bluetooth Local mode enabled and is connected to your Mac over BT. The inference runs on the phone's llama.cpp model -- free, offline, no credits involved.

Forcing a specific route

Pick the model ID:

Want	Model ID
Always local	`edgecoder-local`
Always swarm	`edgecoder-swarm`
Smart routing	`edgecoder-auto`

Troubleshooting

IDE says "connection refused" or "API error"

The provider server is not running. Start it:

bash

cd "/path/to/EdgeCoder" && npm run dev:ide

IDE says "model not found"

Make sure you typed the model name exactly: edgecoder-auto, edgecoder-local, or edgecoder-swarm. Verify they appear:

bash

curl -H "Authorization: Bearer $(cat ~/.edgecoder/ide-token)" http://127.0.0.1:4304/v1/models

Responses are slow

Check which route was used -- look at the X-EdgeCoder-Route response header or the edgecoder.route field in the JSON. If it is hitting swarm, the local model may be overloaded or unhealthy. Verify Ollama is running:

bash

curl http://127.0.0.1:11434/api/tags

Cursor keeps asking for an OpenAI API key

Enter your IDE token (~/.edgecoder/ide-token) in the API key field. The server rejects requests without it (401).

Desktop app cannot connect to services

Make sure the backend services are running. The desktop app expects the coordinator on port 4301, inference on 4302, control plane on 4303, and the IDE provider on 4304. Start all services:

bash

cd "/path/to/EdgeCoder" && npm run dev

IDE Integration ​

EdgeCoder Desktop App (Primary) ​

Chat-first interface ​

Built-in Monaco editor ​

Additional pages ​

Storage ​

Launching the desktop app ​

External IDE Integration ​

How it works ​

Available model IDs ​

Step 1 -- Start the provider server ​

Security: token, bind address and shell approval ​

Extension endpoints ​

Step 2 -- Add as a custom model in your IDE ​

Cursor ​

Zed ​

Continue.dev (VS Code / JetBrains plugin) ​

JetBrains (IntelliJ, PyCharm, WebStorm, etc.) ​

Windsurf (Codeium) ​

Claude Code (Anthropic CLI) ​

Verifying the connection ​

System-aware chat ​

Download progress ​

Routing behaviour ​

When does traffic go to swarm? ​

When does Bluetooth-local activate? ​

Forcing a specific route ​

Troubleshooting ​

IDE says "connection refused" or "API error" ​

IDE says "model not found" ​

Responses are slow ​

Cursor keeps asking for an OpenAI API key ​

Desktop app cannot connect to services ​

IDE Integration

EdgeCoder Desktop App (Primary)

Chat-first interface

Built-in Monaco editor

Additional pages

Storage

Launching the desktop app

External IDE Integration

How it works

Available model IDs

Step 1 -- Start the provider server

Security: token, bind address and shell approval

Extension endpoints

Step 2 -- Add as a custom model in your IDE

Cursor

Zed

Continue.dev (VS Code / JetBrains plugin)

JetBrains (IntelliJ, PyCharm, WebStorm, etc.)

Windsurf (Codeium)

Claude Code (Anthropic CLI)

Verifying the connection

System-aware chat

Download progress

Routing behaviour

When does traffic go to swarm?

When does Bluetooth-local activate?

Forcing a specific route

Troubleshooting

IDE says "connection refused" or "API error"

IDE says "model not found"

Responses are slow

Cursor keeps asking for an OpenAI API key

Desktop app cannot connect to services