Every app and framework below talks to Zalmir through its OpenAI-compatible settings: the base URL https://zalmir.ai/api/v1, a key from API keys in the Console, and any model ID from Models. Each setup follows the app's own documentation or source, and names the switch that turns off the app's own telemetry where it has any. We checked in code that each base URL, plus the path the app adds to it, reaches the right Zalmir endpoint; we have not run the apps themselves. These projects are independent of Zalmir.
n8n
Workflow automation. Setup from docs.n8n.io.
Credentials → Add credential → OpenAI
API Key: zal_...
Base URL: https://zalmir.ai/api/v1
AI Agent or Basic LLM Chain → OpenAI Chat Model
Credential: the one above
Model: deepseek/deepseek-v4.1-flash-abliterated
Use Responses API: off- n8n tests the credential by listing models from the base URL, then lists Zalmir's models in the Model field.
- Keep Use Responses API off, so n8n uses chat completions. Its built-in tools (web search, file search) run on OpenAI's side, and Zalmir refuses them.
- Self-hosted n8n shares anonymous telemetry with n8n by default: N8N_DIAGNOSTICS_ENABLED=false turns it off.
Sim
Workflow automation. Setup from docs.sim.ai.
# .env of your self-hosted Sim, then restart it
VLLM_BASE_URL=https://zalmir.ai/api/v1
VLLM_API_KEY=zal_...
# Agent block → Model: vllm/deepseek/deepseek-v4.1-flash-abliterated- Self-hosted Sim only: its docs list no custom base URL for the hosted app.
- Sim lists Zalmir's models under its vLLM provider with a vllm/ prefix and removes the prefix before sending.
- NEXT_TELEMETRY_DISABLED=1, the Docker and Helm default, turns Sim's telemetry off.
Browser Use
Browser agent. Setup from docs.browser-use.com.
# export ZALMIR_API_KEY=zal_... and export ANONYMIZED_TELEMETRY=false
import asyncio, os
from browser_use import Agent, ChatOpenAI
llm = ChatOpenAI(model="deepseek/deepseek-v4.1-flash-abliterated", base_url="https://zalmir.ai/api/v1",
api_key=os.environ["ZALMIR_API_KEY"])
async def main():
await Agent(task="Find the opening hours of the city library", llm=llm).run()
asyncio.run(main())- Browser Use sends anonymized usage telemetry by default, including the addresses its agent visits (without query strings) and the final result. ANONYMIZED_TELEMETRY=false turns it off.
- Pick a model whose page lists tools and image input when the agent should read screenshots.
Pi
Coding agent. Setup from pi.dev.
// ~/.pi/agent/models.json (export ZALMIR_API_KEY=zal_...)
{
"$schema": "https://pi.dev/schemas/models.schema.json",
"providers": {
"zalmir": {
"baseUrl": "https://zalmir.ai/api/v1",
"api": "openai-completions",
"apiKey": "$ZALMIR_API_KEY",
"models": [{ "id": "deepseek/deepseek-v4.1-flash-abliterated" }]
}
}
}
// In Pi: /model, then pick deepseek/deepseek-v4.1-flash-abliterated- Pi fills in $ZALMIR_API_KEY from your environment when it sends a request.
- Add one entry under models for each Zalmir model you want in /model.
Hugging Face Chat UI
Chat app. Setup from github.com/huggingface/chat-ui.
# .env.local
OPENAI_BASE_URL=https://zalmir.ai/api/v1
OPENAI_API_KEY=zal_...- Chat UI lists every Zalmir model from /models.
- Chat UI keeps your conversations in its own database, not at Zalmir.
AnythingLLM
Chat app. Setup from docs.anythingllm.com.
Settings → AI Providers → LLM → Generic OpenAI
Base URL: https://zalmir.ai/api/v1
API Key: zal_...
Model: deepseek/deepseek-v4.1-flash-abliterated
Model context window: 1048576
Max Tokens: 4096
# Docker .env instead:
LLM_PROVIDER='generic-openai'
GENERIC_OPEN_AI_BASE_PATH='https://zalmir.ai/api/v1'
GENERIC_OPEN_AI_API_KEY=zal_...
GENERIC_OPEN_AI_MODEL_PREF='deepseek/deepseek-v4.1-flash-abliterated'
GENERIC_OPEN_AI_MODEL_TOKEN_LIMIT=1048576
GENERIC_OPEN_AI_MAX_TOKENS=4096
DISABLE_TELEMETRY="true"- Use the base URL above, not the full /chat/completions address.
- Max Tokens caps each answer (AnythingLLM's default is 1024). Keep AnythingLLM's own embedder for your documents.
- Recent AnythingLLM versions show a reasoning level picker for models whose page lists reasoning levels: Zalmir's /models gives them as reasoning_efforts.
- AnythingLLM collects anonymous telemetry by default: turn it off under Settings → Privacy & Data, or set DISABLE_TELEMETRY="true" in Docker.
Jan
Chat app. Setup from jan.ai.
Settings → Model Providers → Add Provider
API format: OpenAI-compatible
Provider name: Zalmir
Base URL: https://zalmir.ai/api/v1
API key: zal_...- Jan loads Zalmir's models from /models when you save the provider.
- Jan cannot detect a custom provider's model features: turn on Tools or Vision per model under its Capabilities when the model's page lists them.
crewAI
Agent framework. Setup from github.com/crewAIInc/crewAI.
# export ZALMIR_API_KEY=zal_...
# export CREWAI_DISABLE_TELEMETRY=true and export CREWAI_TRACING_ENABLED=false
import os
from crewai import LLM, Agent, Crew, Task
llm = LLM(model="deepseek/deepseek-v4.1-flash-abliterated", custom_openai=True,
base_url="https://zalmir.ai/api/v1", api_key=os.environ["ZALMIR_API_KEY"])
writer = Agent(role="Writer", goal="Answer briefly", backstory="A careful writer", llm=llm)
task = Task(description="Name three uses of a hash map.", expected_output="Three bullet points", agent=writer)
print(Crew(agents=[writer], tasks=[task]).kickoff())- custom_openai=True sends the full model ID to the base URL above through crewAI's own OpenAI client, using chat completions.
- crewAI sends anonymous usage telemetry by default: CREWAI_DISABLE_TELEMETRY=true turns it off. CREWAI_TRACING_ENABLED=false also refuses its trace upload.
Agno
Agent framework. Setup from github.com/agno-agi/agno.
# export ZALMIR_API_KEY=zal_... and export AGNO_TELEMETRY=false
import os
from agno.agent import Agent
from agno.models.openai.like import OpenAILike
agent = Agent(model=OpenAILike(id="deepseek/deepseek-v4.1-flash-abliterated", base_url="https://zalmir.ai/api/v1",
api_key=os.environ["ZALMIR_API_KEY"]))
agent.print_response("Name three uses of a hash map.")- OpenAILike is Agno's class for any OpenAI-compatible API.
- Agno agents and teams send run telemetry to Agno by default: AGNO_TELEMETRY=false turns it off (or pass telemetry=False to each Agent).
LlamaIndex
Agent framework. Setup from github.com/run-llama/llama_index.
# export ZALMIR_API_KEY=zal_...
import os
from llama_index.llms.openai_like import OpenAILike
llm = OpenAILike(model="deepseek/deepseek-v4.1-flash-abliterated", api_base="https://zalmir.ai/api/v1",
api_key=os.environ["ZALMIR_API_KEY"], is_chat_model=True,
is_function_calling_model=True, context_window=1048576)
print(llm.complete("Name three uses of a hash map."))- Set is_chat_model=True, or LlamaIndex sends plain completions. is_function_calling_model=True turns on its tool use.
- context_window is the model's context_length from /models; LlamaIndex's own default is much smaller.
- LlamaIndex has no usage telemetry to turn off; it sends traces only to an observability tool you add yourself.
smolagents
Agent framework. Setup from github.com/huggingface/smolagents.
# export ZALMIR_API_KEY=zal_...
import os
from smolagents import CodeAgent, OpenAIServerModel
model = OpenAIServerModel(model_id="deepseek/deepseek-v4.1-flash-abliterated", api_base="https://zalmir.ai/api/v1",
api_key=os.environ["ZALMIR_API_KEY"])
agent = CodeAgent(tools=[], model=model)
print(agent.run("What is the 20th Fibonacci number?"))- OpenAIServerModel passes api_base to the OpenAI client, so every call goes to Zalmir.
- smolagents sends no telemetry by default. Its tracing is an extra you install and point somewhere yourself; skip it to keep runs on your machine.
- CodeAgent runs the Python the model writes on your machine; give it tools and inputs you trust.
OpenAI Agents SDK
Agent SDK. Setup from github.com/openai/openai-agents-python.
# export ZALMIR_API_KEY=zal_... and export OPENAI_AGENTS_DISABLE_TRACING=1
import asyncio, os
from agents import Agent, AsyncOpenAI, OpenAIChatCompletionsModel, Runner, set_tracing_disabled
set_tracing_disabled(True) # on by default: every run's trace is uploaded to OpenAI
client = AsyncOpenAI(base_url="https://zalmir.ai/api/v1", api_key=os.environ["ZALMIR_API_KEY"])
agent = Agent(name="Assistant", instructions="Answer briefly.",
model=OpenAIChatCompletionsModel(model="deepseek/deepseek-v4.1-flash-abliterated", openai_client=client))
print(asyncio.run(Runner.run(agent, "Name three uses of a hash map.")).final_output)- Tracing is on by default and uploads every run to OpenAI. set_tracing_disabled(True), or OPENAI_AGENTS_DISABLE_TRACING=1 before the first run, turns it off.
- If you set Zalmir as the default client with set_default_openai_client, the SDK also uses that client's key for trace uploads unless you pass use_for_tracing=False. With tracing off, nothing is uploaded.
- OpenAIChatCompletionsModel uses chat completions, which every Zalmir model supports.
Google ADK
Agent SDK. Setup from github.com/google/adk-python.
# export ZALMIR_API_KEY=zal_... then once: adk telemetry disable
# my_agent/agent.py (pip install "google-adk[extensions]" for LiteLLM)
import os
from google.adk.agents import Agent
from google.adk.models.lite_llm import LiteLlm
root_agent = Agent(name="assistant", instruction="Answer briefly.",
model=LiteLlm(model="openai/deepseek/deepseek-v4.1-flash-abliterated", api_base="https://zalmir.ai/api/v1",
api_key=os.environ["ZALMIR_API_KEY"]))
# adk run my_agent- LiteLlm with the openai/ prefix and api_base sends chat completions to Zalmir. Zalmir receives the model ID without the prefix.
- On first use, the adk command asks to send usage data to Google, and pressing Enter says yes. adk telemetry disable records no.
- LiteLLM downloads its model price list when it loads; LITELLM_LOCAL_MODEL_COST_MAP=True uses the copy it ships with.
CopilotKit
Agent UI kit. Setup from github.com/CopilotKit/CopilotKit.
// .env of the server that runs the CopilotKit runtime:
// ZALMIR_API_KEY=zal_...
// COPILOTKIT_TELEMETRY_DISABLED=true
import { BuiltInAgent } from "@copilotkit/runtime/v2";
import { createOpenAI } from "@ai-sdk/openai";
const zalmir = createOpenAI({ baseURL: "https://zalmir.ai/api/v1", apiKey: process.env.ZALMIR_API_KEY });
const agent = new BuiltInAgent({ model: zalmir.chat("deepseek/deepseek-v4.1-flash-abliterated") });- Use .chat(...) so the AI SDK sends chat completions to Zalmir.
- The CopilotKit runtime sends telemetry by default: COPILOTKIT_TELEMETRY_DISABLED=true (or DO_NOT_TRACK=1) in the runtime's environment turns it off, and CopilotKit says it also stops the Inspector's events.
Langflow
Workflow automation. Setup from github.com/langflow-ai/langflow.
# Langflow's .env (or the shell that starts it), then restart Langflow
DO_NOT_TRACK=true
# In a flow: the OpenAI component
OpenAI API Key: zal_...
Model Name: deepseek/deepseek-v4.1-flash-abliterated (type it in)
OpenAI API Base: https://zalmir.ai/api/v1 (advanced field)- Langflow's releases send product telemetry by default: DO_NOT_TRACK=true turns it off. Its OpenTelemetry export stays off unless you set OTEL_ variables.
- Type the key into the component. Langflow refuses to send a key from its own server environment to an address other than OpenAI's unless LANGFLOW_PROVIDER_CREDENTIAL_ALLOWED_HOSTS=zalmir.ai is set.
OpenClaw
Agent gateway. Setup from github.com/openclaw/openclaw.
// ~/.openclaw/openclaw.json (export ZALMIR_API_KEY=zal_... where OpenClaw starts)
{
update: { checkOnStart: false }, // no automatic request to OpenClaw's update and telemetry server
telemetry: { enabled: false },
models: {
providers: {
zalmir: {
baseUrl: "https://zalmir.ai/api/v1",
apiKey: "${ZALMIR_API_KEY}",
api: "openai-completions",
models: [{
id: "deepseek/deepseek-v4.1-flash-abliterated", name: "DeepSeek V4.1 Flash (Refusals removed)",
reasoning: false, // its model page lists no reasoning levels
input: ["text", "image"], contextWindow: 1048576, maxTokens: 32768,
}],
},
},
},
agents: { defaults: { model: { primary: "zalmir/deepseek/deepseek-v4.1-flash-abliterated" } } },
}- OpenClaw fills in ${ZALMIR_API_KEY} from its environment, so the key stays out of the file. Pick the model as zalmir/ plus the full model ID.
- Mark reasoning: true only on models whose page lists reasoning levels (reasoning_efforts in /models): OpenClaw then sends reasoning_effort. supportedReasoningEfforts keeps it to levels the model lists. Leave reasoning: false on the others.
- OpenClaw checks for updates every day by default and reports update results and an anonymous install count to its own server. update.checkOnStart: false (or OPENCLAW_NO_AUTO_UPDATE=1) stops all of it; you then update by hand. telemetry.enabled: false (or DO_NOT_TRACK=1) stops the install count.
Dify
Workflow automation. Setup from github.com/langgenius/dify.
Settings → Model Provider → install OpenAI-API-compatible (Marketplace)
OpenAI-API-compatible → Add Model
Model Type: LLM
Model Name: deepseek/deepseek-v4.1-flash-abliterated
API Key: zal_...
API Base URL: https://zalmir.ai/api/v1
Completion mode: Chat
Model context size: 1048576
Upper bound for max tokens: 32768
Function Call Type: Function Call
Vision Support: Support
User Identity Support: Not Support
# Self-hosted Dify: docker/.env, then docker compose up -d
DISABLE_TELEMETRY=true
NEXT_TELEMETRY_DISABLED=1- Use Dify's official OpenAI-API-compatible plugin and paste the base URL above as the API Base URL.
- Dify's default context size is 4096: enter the model's context_length from /models. Function Call Type: Function Call turns on tools for models that list them.
- User Identity Support: Not Support keeps Dify from sending its end-user ID with each request.
- Self-hosted Dify reports an install event and a daily heartbeat (a random instance ID, version, OS) to Dify by default: DISABLE_TELEMETRY=true in docker/.env turns it off. NEXT_TELEMETRY_DISABLED=1 does the same for the web container's Next.js.
OpenWhispr
Dictation app. Setup from github.com/OpenWhispr/openwhispr.
Settings → AI Models → Dictation Cleanup
Provider: Custom
Endpoint URL: https://zalmir.ai/api/v1
API Key (Optional): zal_...
Model: deepseek/deepseek-v4.1-flash-abliterated (picked from the list OpenWhispr loads)
Settings → Privacy & Data
Usage analytics: off (the default)
Sync personal Insights: off (the default)- OpenWhispr sends your dictated text to Zalmir for the clean-up step. Zalmir has no speech-to-text endpoint, so keep transcription on OpenWhispr's local model or another service.
- The endpoint above already ends in /v1, so OpenWhispr uses it as it is and lists Zalmir's models from /models. It tries /responses first, then /chat/completions; Zalmir answers both.
- OpenWhispr sends no usage analytics by default: Usage analytics and Sync personal Insights stay off until you turn them on.
Crush
Coding agent. Setup from github.com/charmbracelet/crush.
# ~/.config/crush/crushrc (export ZALMIR_API_KEY=zal_... first)
option metrics false
provider add zalmir --name Zalmir --type completions \
--base-url "https://zalmir.ai/api/v1" \
--api-key "$ZALMIR_API_KEY"
model add zalmir/deepseek/deepseek-v4.1-flash-abliterated --name "DeepSeek V4.1 Flash (Refusals removed)" \
--context-window 1048576 --default-max-tokens 32768
model large zalmir/deepseek/deepseek-v4.1-flash-abliterated
model small zalmir/deepseek/deepseek-v4.1-flash-abliterated- --type completions is Crush's setting for OpenAI-compatible APIs. Crush reads only model IDs from /models, so set the context window yourself (context_length in /models).
- Crush sends usage metrics to its makers by default: option metrics false turns them off, as does CRUSH_DISABLE_METRICS=1 or DO_NOT_TRACK=1.
- CRUSH_DISABLE_PROVIDER_AUTO_UPDATE=1 also stops Crush from downloading its built-in provider list.
Haystack
Agent framework. Setup from github.com/deepset-ai/haystack.
# export ZALMIR_API_KEY=zal_... and export HAYSTACK_TELEMETRY_ENABLED=False
from haystack.components.generators.chat import OpenAIChatGenerator
from haystack.dataclasses import ChatMessage
from haystack.utils import Secret
llm = OpenAIChatGenerator(model="deepseek/deepseek-v4.1-flash-abliterated", api_base_url="https://zalmir.ai/api/v1",
api_key=Secret.from_env_var("ZALMIR_API_KEY"))
reply = llm.run([ChatMessage.from_user("Name three uses of a hash map.")])["replies"][0]
print(reply.text)- OpenAIChatGenerator with api_base_url sends chat completions to Zalmir and asks for token usage when streaming.
- Haystack sends anonymous usage telemetry by default: HAYSTACK_TELEMETRY_ENABLED=False, set before Python imports Haystack, turns it off.
SWE-agent
Coding agent. Setup from github.com/SWE-agent/SWE-agent.
# zalmir.yaml (export ZALMIR_API_KEY=zal_... and export LITELLM_LOCAL_MODEL_COST_MAP=True)
agent:
model:
name: openai/deepseek/deepseek-v4.1-flash-abliterated
api_base: https://zalmir.ai/api/v1
api_key: $ZALMIR_API_KEY
per_instance_cost_limit: 0 # LiteLLM has no price for this model; 0 skips its cost check
total_cost_limit: 0
per_instance_call_limit: 100
max_input_tokens: 1048576
# sweagent run --config config/default.yaml --config zalmir.yaml \
# --env.repo.github_url=https://github.com/you/repo \
# --problem_statement.github_url=https://github.com/you/repo/issues/1- SWE-agent stops a run when LiteLLM cannot price the model, so set both cost limits to 0 and cap the run with per_instance_call_limit and the spending limit on your Zalmir key. A LiteLLM model registry file (litellm_model_registry) with the model's prices from /models works too and keeps SWE-agent's cost check.
- max_input_tokens is the model's context_length from /models; LiteLLM does not know it for this model.
- SWE-agent sends no telemetry. LITELLM_LOCAL_MODEL_COST_MAP=True keeps LiteLLM from downloading its price list when it loads.
Coding agents such as Claude Code and Codex: see Coding agents and Claude Code and Codex compatibility.