Connect a model

中文 · Install and quick start

The desktop app puts the common connection routes inside Connect a model. Normal use does not require entering an endpoint or editing .env: choose a route, authorize or provide a key, then select the models you want.

Connect inside the app

1. Select Connect a model in the toolbar.
2. Choose a provider, coding plan, Ollama, or a custom connection.
3. Follow the in-app authorization flow, or paste the key issued for that route.
4. Select Fetch models, enable the models you want to see, and save.

ThoughtDAG supplies the preset endpoint, reads the available catalog, keeps your selection, and exposes the default model in the toolbar. Use the edit button beside a connection later to refresh its catalog or change the enabled models.

Choose a route

Your situation	Choose in the app	What remains for you
You want the shortest start	OpenRouter	Authorize in the browser, then choose models in the app
You already have a provider API key	That provider preset	Paste the key; the endpoint is already filled in
You have a coding plan	Its coding preset	Use the key issued by that plan
You run local models	Ollama	Start Ollama; the app reads its model catalog
You use another compatible service	Custom endpoint	Provide its base URL and any credentials it requires

Current presets include OpenRouter, Zhipu GLM / Z.ai GLM, GLM Coding, Kimi Code, MiniMax, DeepSeek, OpenAI, Google AI Studio, Kimi / Moonshot, Ollama, and custom OpenAI-compatible endpoints. The models shown in the interface depend on the catalog returned by the service and the access available to your account.

Subscriptions and coding plans

When a plan issues a key for a compatible endpoint, select its preset and paste that key; you do not need to enter its dedicated endpoint manually. GLM Coding, Kimi Code, and MiniMax routes are already included in the app.

A consumer subscription does not necessarily include third-party API access. The ChatGPT plan entry is an optional local compatibility route: ThoughtDAG can connect to an adapter that is already running on the same machine, but the current app does not install or start that adapter for you. Without that local setup, OpenRouter, a provider API key, or local Ollama is the more direct route.

Local models with Ollama

Start Ollama, then choose Ollama → Fetch models in Connect a model. The default local address is already supplied and no key is needed. Other local runtimes can be added through Custom endpoint.

Model requests sent to a local endpoint remain on that route. If web search, scholarly search, or another remote tool is enabled, that tool still receives its corresponding query.

Models, vision, and tools

• The toolbar model is the default for new nodes.
• A node can pin another model, and each answer version records the model actually used.
• Image input needs a vision-capable model or companion text extracted for a text model.
• Web search, scholarly search, and MCP tools are separate capabilities; switching models does not change which graph paths enter context.

Continue with Models, tools, and memory or Privacy and storage.

Connection mechanism for developers

An in-app connection is not a separate model protocol. It turns a compatible endpoint into a runtime model catalog for ThoughtDAG:

Provider preset / custom endpoint
        ↓ Fetch models
Compatible /models catalog
        ↓ User selects and saves
Browser connection data and model selection
        ↓ Generate
Runtime model registry → selected endpoint

Presets and model discovery

A preset mainly carries a stable base URL, the route for obtaining a key, and a small recommended set. It does not hard-code the complete model catalog into the interface or documentation. Fetch models asks the endpoint's /models route through the local or desktop proxy; only connections without that route use a preset catalog. Refreshing a connection reads the catalog again while retaining user selections that remain valid.

Runtime registration and request path

Confirmed connections and model selections are stored in the current browser configuration. At generation time, ThoughtDAG maps the model ID recorded on the node back to its connection:

• Desktop and local runs normally use the proxy running with the app before reaching the selected endpoint.
• On the hosted build, a small set of compatibility-tested connections can use a browser-direct route; other capabilities depend on the deployment path.
• Runtime connections build a request-scoped model mapping without rewriting server providers registered through .env.
• Keys are not written into normal canvas backups. See Privacy and storage for the data boundary.

.env and custom extensions

A source deployment can still connect providers entirely through the interface. Copy .env.example to .env only when the server should register providers at startup. Interface connections are added to that startup catalog for the current user; an already registered server model wins if IDs collide.

A custom service must expose compatible model-list and generation interfaces. Its catalog, capabilities, limits, pricing, and data handling are determined by that service. Additional MCP servers are configured separately in mcp.config.json.

Decision model

The decision model is the fast-thinking half of ThoughtDAG: it never writes an answer, it decides — is this excerpt relevant, is this line a decision, does this upstream change bear on that answer — in about half a second, with calibrated probabilities. Recall, map badges, staleness and memory admission all rest on it; with it off, each falls back to a rule.

It speaks the System One contract (POST /v1/systemone: a state plus typed questions, probabilities back), which Jev-class models implement. Open the model picker (top right) → the Decision model · fast thinking row → the switch is on by default; click the row to configure:

Route	What to enter	Notes
OpenRouter	nothing, if OpenRouter is already among your model connections; else an OpenRouter key	The saved key is used automatically. Model typesafe/jev-1.13. Billed by input tokens, roughly a fraction of a cent per decision.
TypeSafe	a TypeSafe API key	The official endpoint; requests go through the local proxy.
Cloudflare Workers AI	account id and an API token with Workers AI access	Model typesafe/jev; through the local proxy.
Self-hosted	the URL of any /v1/systemone server, key optional	Open reproductions such as Kev, Laya or Von, on your own machine or LAN.
Chat model	nothing	Your current answering model answers the same questions as JSON. Works, but slower, costs output tokens, and the numbers are opinions, not calibrated probabilities.

Test sends one tiny decision and reports the model, the latency and whether it is calibrated. A decision that does not answer within six seconds trips a one-minute breaker: the app says so, and every decision uses its rule until then.

Below the switch, amount sets how much recall may bring into an ask: lean 4k, standard 12k, generous 40k input tokens (a small context window caps at two fifths of itself).

What leaves the machine: the question and the candidate excerpts of that one decision, to the route you chose. Nothing else, and nothing when the decision model is off.

Connect local agent sessions

Model connections answer “which endpoint generates this response.” Session Atlas brings local agent sessions into a mirror canvas that you can inspect and reorganize. In the desktop app, Session Atlas → Sources can re-detect directories and Enable/Update the command that opens a session in ThoughtDAG.

Open the Session Atlas guide