cb9de10 feat: add ollama as first-class provider · zut

M README.md +42 -2


@@ -282,15 +282,55 @@ Place a `models.json` in `$ZOT_HOME` (macOS: `~/Library/Application Support/zot/
282	282		}
283	283		```
284	284
285		-	Supported fields per model: `id` (required), `name`, `reasoning`, `contextWindow`, `maxTokens`, `priceInput`, `priceOutput`, `priceCacheRead`, `priceCacheWrite`.
	285	+	Supported fields per model: `id` (required), `name`, `reasoning`, `contextWindow`, `maxTokens`, `baseUrl`, `priceInput`, `priceOutput`, `priceCacheRead`, `priceCacheWrite`.
286	286
287	287		Provider keys are normalized: `openai-codex` and `openai-responses` map to `openai`, `anthropic-messages` maps to `anthropic`.
288	288
289	289		User-defined models show `source: user` in `--list-models` and take precedence over both the baked-in catalog and live-discovered models. Missing or invalid files are silently ignored.
290	290
	291	+	### Local models with ollama
	292	+
	293	+	zot works with [ollama](https://ollama.com) out of the box. Ollama serves an OpenAI-compatible API locally, so any model you have pulled works with zot.
	294	+
	295	+	Quick start:
	296	+
	297	+	```bash
	298	+	ollama pull qwen3.5:4b
	299	+	zot --provider ollama --model qwen3.5:4b
	300	+	```
	301	+
	302	+	That's it. No API key needed for local models. zot defaults to `http://localhost:11434`.
	303	+
	304	+	For a remote ollama instance or one behind auth:
	305	+
	306	+	```bash
	307	+	zot --provider ollama --model llama3 --base-url https://my-server.com/v1 --api-key my-token
	308	+	```
	309	+
	310	+	You can also add models to your `models.json` so you don't need flags every time:
	311	+
	312	+	```json
	313	+	{
	314	+	"providers": {
	315	+	"ollama": {
	316	+	"models": [
	317	+	{
	318	+	"id": "qwen3.5:4b",
	319	+	"name": "Qwen 3.5 4B",
	320	+	"contextWindow": 32768,
	321	+	"maxTokens": 8192
	322	+	}
	323	+	]
	324	+	}
	325	+	}
	326	+	}
	327	+	```
	328	+
	329	+	The `ollama` provider uses the OpenAI chat completions protocol internally, so it also works with any OpenAI-compatible server (vLLM, LM Studio, LocalAI, etc.).
	330	+
291	331		## Inline images
292	332
293		-	When a tool returns an image (for example `read` on a PNG), zot renders it inline on terminals that support it: iTerm2, WezTerm, Kitty, Ghostty. On other terminals you see a text placeholder with MIME type, pixel dimensions, and byte size. Control with the `ZOT_INLINE_IMAGES` env var:
	333	+	When a tool returns an image (for example `read` on a PNG), zot renders it inline on terminals that support it: Ghostty, Kitty, iTerm2, WezTerm. On other terminals you see a text placeholder with MIME type, pixel dimensions, and byte size. Control with the `ZOT_INLINE_IMAGES` env var:
294	334
295	335		\| Value \| Effect \|
296	336		\|---\|---\|


@@ -282,15 +282,55 @@ Place a `models.json` in `$ZOT_HOME` (macOS: `~/Library/Application Support/zot/
282		}	282		}
283		```	283		```
284			284
285	-	Supported fields per model: `id` (required), `name`, `reasoning`, `contextWindow`, `maxTokens`, `priceInput`, `priceOutput`, `priceCacheRead`, `priceCacheWrite`.	285	+	Supported fields per model: `id` (required), `name`, `reasoning`, `contextWindow`, `maxTokens`, `baseUrl`, `priceInput`, `priceOutput`, `priceCacheRead`, `priceCacheWrite`.
286			286
287		Provider keys are normalized: `openai-codex` and `openai-responses` map to `openai`, `anthropic-messages` maps to `anthropic`.	287		Provider keys are normalized: `openai-codex` and `openai-responses` map to `openai`, `anthropic-messages` maps to `anthropic`.
288			288
289		User-defined models show `source: user` in `--list-models` and take precedence over both the baked-in catalog and live-discovered models. Missing or invalid files are silently ignored.	289		User-defined models show `source: user` in `--list-models` and take precedence over both the baked-in catalog and live-discovered models. Missing or invalid files are silently ignored.
290			290
			291	+	### Local models with ollama
			292	+
			293	+	zot works with [ollama](https://ollama.com) out of the box. Ollama serves an OpenAI-compatible API locally, so any model you have pulled works with zot.
			294	+
			295	+	Quick start:
			296	+
			297	+	```bash
			298	+	ollama pull qwen3.5:4b
			299	+	zot --provider ollama --model qwen3.5:4b
			300	+	```
			301	+
			302	+	That's it. No API key needed for local models. zot defaults to `http://localhost:11434`.
			303	+
			304	+	For a remote ollama instance or one behind auth:
			305	+
			306	+	```bash
			307	+	zot --provider ollama --model llama3 --base-url https://my-server.com/v1 --api-key my-token
			308	+	```
			309	+
			310	+	You can also add models to your `models.json` so you don't need flags every time:
			311	+
			312	+	```json
			313	+	{
			314	+	"providers": {
			315	+	"ollama": {
			316	+	"models": [
			317	+	{
			318	+	"id": "qwen3.5:4b",
			319	+	"name": "Qwen 3.5 4B",
			320	+	"contextWindow": 32768,
			321	+	"maxTokens": 8192
			322	+	}
			323	+	]
			324	+	}
			325	+	}
			326	+	}
			327	+	```
			328	+
			329	+	The `ollama` provider uses the OpenAI chat completions protocol internally, so it also works with any OpenAI-compatible server (vLLM, LM Studio, LocalAI, etc.).
			330	+
291		## Inline images	331		## Inline images
292			332
293	-	When a tool returns an image (for example `read` on a PNG), zot renders it inline on terminals that support it: iTerm2, WezTerm, Kitty, Ghostty. On other terminals you see a text placeholder with MIME type, pixel dimensions, and byte size. Control with the `ZOT_INLINE_IMAGES` env var:	333	+	When a tool returns an image (for example `read` on a PNG), zot renders it inline on terminals that support it: Ghostty, Kitty, iTerm2, WezTerm. On other terminals you see a text placeholder with MIME type, pixel dimensions, and byte size. Control with the `ZOT_INLINE_IMAGES` env var:
294			334
295		\| Value \| Effect \|	335		\| Value \| Effect \|
296		\|---\|---\|	336		\|---\|---\|

M internal/agent/args.go +1 -1


@@ -328,7 +328,7 @@ func PrintHelp(version string) {
328	328		row{"zot tg ...", "short alias for telegram-bot"},
329	329		)
330	330		section("provider and model flags",
331		-	row{"--provider", "provider to use (anthropic\|openai)"},
	331	+	row{"--provider", "provider to use (anthropic\|openai\|ollama)"},
332	332		row{"--model ID", "model id (see --list-models)"},
333	333		row{"--api-key KEY", "api key for this run (env / auth.json fallback)"},
334	334		row{"--base-url URL", "override provider api base url"},


@@ -328,7 +328,7 @@ func PrintHelp(version string) {
328		row{"zot tg ...", "short alias for telegram-bot"},	328		row{"zot tg ...", "short alias for telegram-bot"},
329		)	329		)
330		section("provider and model flags",	330		section("provider and model flags",
331	-	row{"--provider", "provider to use (anthropic\|openai)"},	331	+	row{"--provider", "provider to use (anthropic\|openai\|ollama)"},
332		row{"--model ID", "model id (see --list-models)"},	332		row{"--model ID", "model id (see --list-models)"},
333		row{"--api-key KEY", "api key for this run (env / auth.json fallback)"},	333		row{"--api-key KEY", "api key for this run (env / auth.json fallback)"},
334		row{"--base-url URL", "override provider api base url"},	334		row{"--base-url URL", "override provider api base url"},

M internal/agent/build.go +53 -18


@@ -127,19 +127,32 @@ func Resolve(args Args, requireCred bool) (Resolved, error) {
127	127
128	128		// User-requested provider (explicit > config > default).
129	129		provName := firstNonEmpty(args.Provider, cfg.Provider, "anthropic")
130		-	if provName != "anthropic" && provName != "openai" {
131		-	return Resolved{}, fmt.Errorf("provider must be anthropic or openai (got %q)", provName)
	130	+	if provName != "anthropic" && provName != "openai" && provName != "ollama" {
	131	+	return Resolved{}, fmt.Errorf("provider must be anthropic, openai, or ollama (got %q)", provName)
132	132		}
133	133
134		-	// Try the requested provider first.
135		-	cred, method, accountID, credErr := ResolveCredentialFull(provName, args.APIKey)
	134	+	var (
	135	+	cred string
	136	+	method string
	137	+	accountID string
	138	+	credErr error
	139	+	)
	140	+	if provName == "ollama" {
	141	+	if args.BaseURL == "" {
	142	+	args.BaseURL = "http://localhost:11434"
	143	+	}
	144	+	cred = firstNonEmpty(args.APIKey, "ollama")
	145	+	method = "apikey"
	146	+	} else {
	147	+	cred, method, accountID, credErr = ResolveCredentialFull(provName, args.APIKey)
	148	+	}
136	149
137	150		// If the user did NOT explicitly pick a provider and the default one
138	151		// has no credentials, auto-fall-back to whichever provider is actually
139	152		// logged in. That way running plain `zot` after `/login` (any provider)
140	153		// never shows a "not logged in" banner.
141	154		userPickedProvider := args.Provider != ""
142		-	if credErr != nil && !userPickedProvider {
	155	+	if credErr != nil && !userPickedProvider && provName != "ollama" {
143	156		other := "openai"
144	157		if provName == "openai" {
145	158		other = "anthropic"

@@ -152,25 +165,41 @@ func Resolve(args Args, requireCred bool) (Resolved, error) {
152	165
153	166		model := firstNonEmpty(args.Model, cfg.Model)
154	167		if model == "" {
155		-	if provName == "openai" {
	168	+	switch provName {
	169	+	case "openai":
156	170		model = "gpt-5"
157		-	} else {
	171	+	case "ollama":
	172	+	return Resolved{}, fmt.Errorf("ollama requires --model (e.g. --model llama3)")
	173	+	default:
158	174		model = provider.DefaultModel.ID
159	175		}
160	176		}
161	177		// If the resolved model belongs to a different provider (e.g. config
162	178		// says gpt-5 but we fell back to anthropic), pick that provider's default.
163		-	if m, err := provider.FindModel("", model); err == nil && m.Provider != provName {
164		-	if provName == "openai" {
165		-	model = "gpt-5"
166		-	} else {
167		-	model = provider.DefaultModel.ID
	179	+	if provName != "ollama" {
	180	+	if m, err := provider.FindModel("", model); err == nil && m.Provider != provName {
	181	+	if provName == "openai" {
	182	+	model = "gpt-5"
	183	+	} else {
	184	+	model = provider.DefaultModel.ID
	185	+	}
168	186		}
169	187		}
170	188		resolvedModel, err := provider.FindModel(provName, model)
171		-	if err != nil {
	189	+	if err != nil && provName != "ollama" {
172	190		return Resolved{}, err
173	191		}
	192	+	if err != nil && provName == "ollama" {
	193	+	resolvedModel = provider.Model{
	194	+	Provider: "ollama",
	195	+	ID: model,
	196	+	DisplayName: model,
	197	+	ContextWindow: 32768,
	198	+	MaxOutput: 8192,
	199	+	BaseURL: args.BaseURL,
	200	+	Source: "ollama",
	201	+	}
	202	+	}
174	203
175	204		// If the model defines a base URL (e.g. local ollama) and the
176	205		// user didn't pass --base-url, use the model's URL.

@@ -303,16 +332,20 @@ func (r Resolved) NewClient() provider.Client {
303	332		if !r.HasCredential() {
304	333		panic("NewClient called without credential; check HasCredential first")
305	334		}
306		-	if r.Provider == "openai" {
	335	+	switch r.Provider {
	336	+	case "ollama":
	337	+	return provider.NewOpenAI(r.Credential, r.BaseURL)
	338	+	case "openai":
307	339		if r.AuthMethod == "oauth" {
308	340		return provider.NewOpenAICodex(r.Credential, r.AccountID, r.BaseURL)
309	341		}
310	342		return provider.NewOpenAI(r.Credential, r.BaseURL)
	343	+	default:
	344	+	if r.AuthMethod == "oauth" {
	345	+	return provider.NewAnthropicOAuth(r.Credential, r.BaseURL)
	346	+	}
	347	+	return provider.NewAnthropic(r.Credential, r.BaseURL)
311	348		}
312		-	if r.AuthMethod == "oauth" {
313		-	return provider.NewAnthropicOAuth(r.Credential, r.BaseURL)
314		-	}
315		-	return provider.NewAnthropic(r.Credential, r.BaseURL)
316	349		}
317	350
318	351		// UseSandbox replaces the sandbox pointer that every tool in r's

@@ -395,6 +428,8 @@ func envVarName(provider string) string {
395	428		switch provider {
396	429		case "openai":
397	430		return "OPENAI"
	431	+	case "ollama":
	432	+	return "OLLAMA"
398	433		default:
399	434		return "ANTHROPIC"
400	435		}

 	// User-requested provider (explicit > config > default).
 	provName := firstNonEmpty(args.Provider, cfg.Provider, "anthropic")
-	if provName != "anthropic" && provName != "openai" {
+	if provName != "anthropic" && provName != "openai" && provName != "ollama" {
-		return Resolved{}, fmt.Errorf("provider must be anthropic or openai (got %q)", provName)
+		return Resolved{}, fmt.Errorf("provider must be anthropic, openai, or ollama (got %q)", provName)
 	}
-	// Try the requested provider first.
+	var (
-	cred, method, accountID, credErr := ResolveCredentialFull(provName, args.APIKey)
+		cred      string
+		method    string
+		accountID string
+		credErr   error
+	)
+	if provName == "ollama" {
+		if args.BaseURL == "" {
+			args.BaseURL = "http://localhost:11434"
+		}
+		cred = firstNonEmpty(args.APIKey, "ollama")
+		method = "apikey"
+	} else {
+		cred, method, accountID, credErr = ResolveCredentialFull(provName, args.APIKey)
+	}
 	// If the user did NOT explicitly pick a provider and the default one
 	// has no credentials, auto-fall-back to whichever provider is actually
 	// logged in. That way running plain `zot` after `/login` (any provider)
 	// never shows a "not logged in" banner.
 	userPickedProvider := args.Provider != ""
-	if credErr != nil && !userPickedProvider {
+	if credErr != nil && !userPickedProvider && provName != "ollama" {
 		other := "openai"
 		if provName == "openai" {
 			other = "anthropic"
 	model := firstNonEmpty(args.Model, cfg.Model)
 	if model == "" {
-		if provName == "openai" {
+		switch provName {
+		case "openai":
 			model = "gpt-5"
-		} else {
+		case "ollama":
+			return Resolved{}, fmt.Errorf("ollama requires --model (e.g. --model llama3)")
+		default:
 			model = provider.DefaultModel.ID
 		}
 	}
 	// If the resolved model belongs to a different provider (e.g. config
 	// says gpt-5 but we fell back to anthropic), pick that provider's default.
-	if m, err := provider.FindModel("", model); err == nil && m.Provider != provName {
+	if provName != "ollama" {
-		if provName == "openai" {
+		if m, err := provider.FindModel("", model); err == nil && m.Provider != provName {
-			model = "gpt-5"
+			if provName == "openai" {
-		} else {
+				model = "gpt-5"
-			model = provider.DefaultModel.ID
+			} else {
+				model = provider.DefaultModel.ID
+			}
 		}
 	}
 	resolvedModel, err := provider.FindModel(provName, model)
-	if err != nil {
+	if err != nil && provName != "ollama" {
 		return Resolved{}, err
 	}
+	if err != nil && provName == "ollama" {
+		resolvedModel = provider.Model{
+			Provider:      "ollama",
+			ID:            model,
+			DisplayName:   model,
+			ContextWindow: 32768,
+			MaxOutput:     8192,
+			BaseURL:       args.BaseURL,
+			Source:        "ollama",
+		}
+	}
 	// If the model defines a base URL (e.g. local ollama) and the
 	// user didn't pass --base-url, use the model's URL.
 	if !r.HasCredential() {
 		panic("NewClient called without credential; check HasCredential first")
 	}
-	if r.Provider == "openai" {
+	switch r.Provider {
+	case "ollama":
+		return provider.NewOpenAI(r.Credential, r.BaseURL)
+	case "openai":
 		if r.AuthMethod == "oauth" {
 			return provider.NewOpenAICodex(r.Credential, r.AccountID, r.BaseURL)
 		}
 		return provider.NewOpenAI(r.Credential, r.BaseURL)
+	default:
+		if r.AuthMethod == "oauth" {
+			return provider.NewAnthropicOAuth(r.Credential, r.BaseURL)
+		}
+		return provider.NewAnthropic(r.Credential, r.BaseURL)
 	}
-	if r.AuthMethod == "oauth" {
-		return provider.NewAnthropicOAuth(r.Credential, r.BaseURL)
-	}
-	return provider.NewAnthropic(r.Credential, r.BaseURL)
 }
 // UseSandbox replaces the sandbox pointer that every tool in r's
 	switch provider {
 	case "openai":
 		return "OPENAI"
+	case "ollama":
+		return "OLLAMA"
 	default:
 		return "ANTHROPIC"
 	}

M internal/provider/openai.go +12 -3


@@ -111,9 +111,18 @@ type oaiRequest struct {
111	111		// ---- request building ----
112	112
113	113		func (c openaiClient) buildRequest(req Request) (oaiRequest, error) {
114		-	m, err := FindModel("openai", req.Model)
	114	+	// Look up the model by id across all providers (not just openai)
	115	+	// because the OpenAI client is also used for ollama and other
	116	+	// OpenAI-compatible backends.
	117	+	m, err := FindModel("", req.Model)
115	118		if err != nil {
116		-	return nil, err
	119	+	// Unknown model: use sensible defaults so local/custom
	120	+	// models still work without a catalog entry.
	121	+	m = Model{
	122	+	ID: req.Model,
	123	+	ContextWindow: 32768,
	124	+	MaxOutput: 8192,
	125	+	}
117	126		}
118	127		out := &oaiRequest{
119	128		Model: req.Model,

@@ -311,7 +320,7 @@ func (c openaiClient) runStream(ctx context.Context, resp http.Response, req R
311	320		defer close(out)
312	321		defer resp.Body.Close()
313	322
314		-	model, _ := FindModel("openai", req.Model)
	323	+	model, _ := FindModel("", req.Model)
315	324		out <- EventStart{Model: req.Model, Provider: "openai"}
316	325
317	326		raw := make(chan sseEvent, 16)

 // ---- request building ----
 func (c *openaiClient) buildRequest(req Request) (*oaiRequest, error) {
-	m, err := FindModel("openai", req.Model)
+	// Look up the model by id across all providers (not just openai)
+	// because the OpenAI client is also used for ollama and other
+	// OpenAI-compatible backends.
+	m, err := FindModel("", req.Model)
 	if err != nil {
-		return nil, err
+		// Unknown model: use sensible defaults so local/custom
+		// models still work without a catalog entry.
+		m = Model{
+			ID:            req.Model,
+			ContextWindow: 32768,
+			MaxOutput:     8192,
+		}
 	}
 	out := &oaiRequest{
 		Model:         req.Model,
 	defer close(out)
 	defer resp.Body.Close()
-	model, _ := FindModel("openai", req.Model)
+	model, _ := FindModel("", req.Model)
 	out <- EventStart{Model: req.Model, Provider: "openai"}
 	raw := make(chan sseEvent, 16)