Last updated: September 27, 2026 · model list checked against NVIDIA’s live /v1/models endpoint (82 IDs) and each tool’s official docs on the same day
Your free NVIDIA API key works in almost any AI coding tool that accepts an OpenAI-compatible endpoint: set the base URL to https://integrate.api.nvidia.com/v1, paste your nvapi- key and pick a model ID such as z-ai/glm-5.3. Continue and OpenCode have native NVIDIA providers, Aider reaches it through LiteLLM’s nvidia_nim/ prefix, Cline and Kilo Code use their “OpenAI Compatible” provider, Cursor needs its base-URL override, and Claude Code only works through a translating proxy because NVIDIA’s hosted API has no Anthropic /v1/messages endpoint. The free tier is for testing and evaluation, and its rate limits vary by model and are not published.
Which coding tools work with the free NVIDIA NIM API?
Every mainstream coding tool can use it, but only some can use it without workarounds. NVIDIA’s hosted catalog at build.nvidia.com speaks the OpenAI Chat Completions format. So the question for each tool is simple: does it let you point an OpenAI-style client at a custom URL, and does it handle the model’s tool calls correctly once you do?
If you do not have a key yet, follow our step-by-step guide on how to get a free NVIDIA API key first. It takes a few minutes. Everything below assumes you have an nvapi- key in an environment variable.
The table sums up what each tool’s own documentation says as of September 27, 2026. “Native” means the tool ships a named NVIDIA provider. Everything else goes through a generic OpenAI-compatible setting.
| Tool | Native NVIDIA provider? | How you connect | Agent / tool calls | Main gotcha |
|---|---|---|---|---|
| Continue | Yes (provider: nvidia) | config.yaml | Yes, after adding capabilities: [tool_use] | Agent mode stays greyed out without the capability flag |
| OpenCode | Yes (built-in NVIDIA provider) | /connect or NVIDIA_API_KEY | Yes | Brand-new models may need a manual config entry |
| Aider | Yes, via LiteLLM nvidia_nim/ | NVIDIA_NIM_API_KEY + --model | Uses edit formats, not tool calls | Unknown context size warning; pin a safe LiteLLM version |
| Cline | No (community PR closed unmerged) | “OpenAI Compatible” provider | Yes | Set the context window by hand; 429 bursts in long tasks |
| Kilo Code | No | “OpenAI Compatible” provider | Yes | Some models reject penalty parameters with HTTP 422 |
| Zed | No | openai_compatible in settings.json | Yes (tools on by default) | List each model and its max tokens yourself |
| Cursor | No | “Override OpenAI Base URL” | Chat yes; Agent mode unconfirmed | Override reroutes all OpenAI-path models; key passes through Cursor’s servers |
| Claude Code | No, and not supported by Anthropic | Translating proxy (LiteLLM, claude-code-router) | Depends on the proxy | NVIDIA has no /v1/messages; extra moving parts |
One tool is missing on purpose. Roo Code shut down on May 15, 2026 and its extension repository is archived. Several NVIDIA-plus-coding tutorials that still rank for these queries recommend it. If you used Roo Code, Cline is the closest replacement and the setup below applies.
Which NVIDIA NIM models are best for coding right now?
For agentic coding in September 2026, start with z-ai/glm-5.3, keep deepseek-ai/deepseek-v4.1-flash or z-ai/glm-5.3-flash as a fast fallback, and try moonshotai/kimi-k3 in tools that handle reasoning models well. All models in the table below appeared in NVIDIA’s live model list on September 27, 2026, and NVIDIA’s own reference pages list tool calling for each of them.
That check matters more than it sounds. Many popular guides still recommend Qwen3 Coder 480B, MiniMax M2.7 or GPT-OSS-120B for the free tier. None of those IDs were in the live /v1/models response on the day we checked, even though some still have pages on build.nvidia.com or in NVIDIA’s docs index. A model that is listed in the docs but missing from the API list will usually fail in your IDE with a “not found” error rather than a helpful message.
| Model ID (paste as-is) | Context | Tool calling | Best use in a coding tool |
|---|---|---|---|
z-ai/glm-5.3 | 1M | Yes (OpenAI format) | Default agent model. NVIDIA positions it for complex coding and long-horizon agent work; thinking depth via reasoning_effort. The hosted endpoint serves a quantized checkpoint. |
moonshotai/kimi-k3 | 1M | Yes | Heavy multi-step tasks. Reasoning model: clients must send back the full assistant message, including reasoning_content and tool_calls. |
deepseek-ai/deepseek-v4.1-flash | 1M | Yes | Newest addition (September 18, 2026). Good fallback when your main model returns 429. |
z-ai/glm-5.3-flash | up to 1M | Yes | Lighter, faster sibling of GLM-5.3 for quick edits and background tasks. |
moonshotai/kimi-k2.6 | 256K | Yes | Mature option; NVIDIA describes it as tuned for production-level coding tasks. |
nvidia/nemotron-3-super-120b-a12b | up to 1M | Yes | OpenCode’s own documentation example; balanced speed and quality. |
nvidia/nemotron-3-ultra-550b-a55b | up to 1M | Yes | NVIDIA’s largest open model; toggle reasoning with enable_thinking. See our Nemotron 3 Ultra breakdown. |
poolside/laguna-xs-2.1 | 262K | Yes | Built for agentic coding and terminal-style tasks. |
mistralai/mistral-nemotron | 128K | Yes | Smaller context but a long track record on the catalog (since June 2025). |
openai/gpt-oss-20b | 128K | Yes | Small and fast; fine for commit messages and simple refactors. |
You do not have to trust this table, or anyone else’s. NVIDIA’s model list is public and needs no key, so you can check what exists before you paste an ID into your IDE:
# List every model ID the hosted API currently exposes (no key required)
curl -s https://integrate.api.nvidia.com/v1/models | jq -r '.data[].id' | sort
# Narrow it down to the coding-relevant families
curl -s https://integrate.api.nvidia.com/v1/models \
| jq -r '.data[].id' | grep -Ei 'glm|kimi|deepseek|nemotron-3|laguna|gpt-oss'Listed does not always mean callable. The public list also includes embedding, safety and vision models, and it does not mark which endpoints are free for your account. NVIDIA’s developer forum has reports of listed models returning 404 “Not found for account”. If a model fails, try the next one in the table before debugging your tool.
Our selection logic, rather than a leaderboard: for agent loops, tool-call reliability beats raw benchmark scores. A model that scores two points higher on a coding benchmark but occasionally emits malformed tool calls will waste more of your rate limit on retries than it saves. That is why the default pick is a model NVIDIA documents with OpenAI-format tool calls and a controllable thinking budget, not simply the biggest model in the catalog. For broader context on how models are compared, see our AI agent benchmarks guide.
How do you use NVIDIA NIM in Cline?
Choose “OpenAI Compatible” as the API provider, enter https://integrate.api.nvidia.com/v1 as the base URL, paste your key and type the model ID. Cline has no dedicated NVIDIA option. A community pull request to add one (#8721) was closed without merging in January 2026, and a later feature request was also closed. Search snippets that say “Cline added an NVIDIA NIM provider” are describing that unmerged PR.
- Open Cline’s settings (gear icon) and set API Provider to OpenAI Compatible.
- Base URL:
https://integrate.api.nvidia.com/v1 - API Key: your
nvapi-…key. - Model ID: for example
z-ai/glm-5.3. - Open Model Configuration and set the Context Window explicitly. Cline cannot read it from NVIDIA’s API, and a wrong value either wastes context or triggers errors on long tasks. For the 1M-token models, a working value of 128,000–200,000 tokens is usually more practical than the maximum, because every extra token in each request adds latency.
- Leave image support off unless the model is a vision model.
The most common failure in Cline is not configuration but HTTP 429. Every tool step in an agent loop (read a file, run a command, edit, re-read) is a separate API request. A single “fix this failing test” task can add up to dozens of requests in a few minutes. One developer’s published Cline pull-request review workflow on NIM hit 429 simply by sending several chunked review calls within a minute. Keep one Cline task running at a time and keep a second model configured as a fallback profile.
How do you set up NVIDIA NIM in Continue?
Continue has a native nvidia provider, so you only need the model ID and your key in config.yaml. The one extra line most guides miss is the capability flag: without it, Continue may not recognise newer model IDs as tool-capable, and Agent mode stays unavailable.
name: NVIDIA NIM config
version: 0.0.1
schema: v1
models:
- name: GLM-5.3 (NVIDIA)
provider: nvidia
model: z-ai/glm-5.3
apiKey: nvapi-xxxxxxxxxxxxxxxxxxxx # keep this file out of git
roles: [chat, edit, apply]
capabilities: [tool_use]
defaultCompletionOptions:
contextLength: 131072
maxTokens: 8192
- name: DeepSeek V4.1 Flash (NVIDIA fallback)
provider: nvidia
model: deepseek-ai/deepseek-v4.1-flash
apiKey: nvapi-xxxxxxxxxxxxxxxxxxxx
roles: [chat, edit]
capabilities: [tool_use]Two practical notes. First, the native provider already points at https://integrate.api.nvidia.com/v1/, so you do not set apiBase. If you prefer the generic route, provider: openai with apiBase: https://integrate.api.nvidia.com/v1 works the same way. Second, do not give a NIM model the autocomplete role. Autocomplete fires on keystrokes, so it would burn through a per-minute limit that your chat and agent requests need. Use a small local model for autocomplete instead. Our Ollama vs LM Studio vs llama.cpp guide covers the local side.
How do you use NVIDIA NIM with Aider?
Aider talks to NVIDIA through LiteLLM, so you set NVIDIA_NIM_API_KEY and prefix the model ID with nvidia_nim/. Aider edits code with its own edit formats (whole-file or diff) rather than native tool calls, so tool-calling quirks matter less here than in Cline or Continue.
# macOS / Linux (Windows: setx NVIDIA_NIM_API_KEY nvapi-... then open a new terminal)
export NVIDIA_NIM_API_KEY=nvapi-xxxxxxxxxxxxxxxxxxxx
# Provider prefix + NVIDIA model ID
aider --model nvidia_nim/z-ai/glm-5.3
# Alternative: treat NVIDIA as a generic OpenAI-compatible endpoint
export OPENAI_API_BASE=https://integrate.api.nvidia.com/v1
export OPENAI_API_KEY=nvapi-xxxxxxxxxxxxxxxxxxxx
aider --model openai/z-ai/glm-5.3For new models Aider warns “Unknown context window size and costs, using sane defaults”. The fix is a small metadata file in your home directory or repository root. It tells Aider how much context to use, which also keeps your requests smaller:
{
"nvidia_nim/z-ai/glm-5.3": {
"max_input_tokens": 131072,
"max_output_tokens": 8192,
"input_cost_per_token": 0,
"output_cost_per_token": 0,
"litellm_provider": "nvidia_nim",
"mode": "chat"
}
}Check your LiteLLM version. Aider depends on LiteLLM, and LiteLLM releases 1.82.7 and 1.82.8 on PyPI were compromised with a credential stealer in March 2026. Run pip show litellm in the environment where Aider is installed. If you see either version, upgrade and rotate every API key that machine could read, including your NVIDIA key.
How do you use NVIDIA NIM in OpenCode?
OpenCode ships a built-in NVIDIA provider: run /connect, search for NVIDIA, paste your key, then choose a model with /models. OpenCode’s own docs describe build.nvidia.com as giving access to Nemotron and many other open models for free, and use nemotron-3-super-120b-a12b as the example. You can also skip /connect and export NVIDIA_API_KEY before starting OpenCode.
OpenCode reads its model catalog from the community-maintained models.dev registry, so a model NVIDIA added last week may not appear yet. On September 27, 2026, DeepSeek V4.1 Flash was live on NVIDIA but not yet in that registry. The documented custom-provider pattern fixes that:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"nim": {
"npm": "@ai-sdk/openai-compatible",
"name": "NVIDIA NIM (custom)",
"options": {
"baseURL": "https://integrate.api.nvidia.com/v1",
"apiKey": "{env:NVIDIA_API_KEY}"
},
"models": {
"deepseek-ai/deepseek-v4.1-flash": {
"name": "DeepSeek V4.1 Flash",
"limit": { "context": 1048576, "output": 32768 }
}
}
}
}
}The output limit is a client-side cap you choose, not an NVIDIA figure. Keep it modest; agents rarely need more than a few thousand output tokens per step. Credentials added via /connect are stored in ~/.local/share/opencode/auth.json, so treat that file like any other secret.
Can you use the NVIDIA API in Cursor?
Yes, for chat, through Settings → Models: add the model name, paste your nvapi- key in the OpenAI API key field and enable “Override OpenAI Base URL” with https://integrate.api.nvidia.com/v1. It works, but it is the least clean setup in this guide, for four reasons that come from Cursor’s own help pages and staff answers on its forum:
- Requests go through Cursor’s servers. Cursor builds the final prompt on its backend, so your NVIDIA key and your code pass through Cursor on the way to NVIDIA. Cursor states that Zero Data Retention does not apply to requests made with your own keys.
- The override is global. Cursor staff confirmed in August 2026 that the base-URL override reroutes all OpenAI-path models, including Cursor-managed ones. People toggle it on for NIM and off again for Cursor’s models.
- Tab completion stays on Cursor’s models. Custom keys only work with chat models.
- Agent mode is not guaranteed. Older forum threads show Agent and Edit refusing custom models billed to an API key. We could not confirm the September 2026 behaviour for free-plan users, so test it before you depend on it.
If Cursor’s agent is the reason you use Cursor, the free NVIDIA key is a poor fit. A VS Code setup with Cline or Continue gives you a direct connection and full agent features. Our Cursor vs Claude Code comparison covers what you would give up.
Can Claude Code use the free NVIDIA NIM API?
Only through a proxy that translates Anthropic’s Messages API into OpenAI format, and Anthropic explicitly does not support it. Claude Code sends requests to /v1/messages. We sent a request to https://integrate.api.nvidia.com/v1/messages on September 27, 2026 and got HTTP 404. Some guides confuse this with NVIDIA’s self-hosted NIM containers, which document different endpoints. The hosted free API has no Anthropic-compatible route.
Claude Code’s documentation is equally clear from the other side: it “does not support non-Claude models”, and setting ANTHROPIC_BASE_URL changes where requests go, not which model answers. So this is a community setup for experimentation, not a supported configuration. If you still want to try it, LiteLLM’s official tutorial for non-Anthropic models (which requires Claude Code v2.1.129 or later) is the most documented route:
model_list:
- model_name: glm-5.3
litellm_params:
model: nvidia_nim/z-ai/glm-5.3
api_key: os.environ/NVIDIA_NIM_API_KEY
- model_name: glm-5.3-flash
litellm_params:
model: nvidia_nim/z-ai/glm-5.3-flash
api_key: os.environ/NVIDIA_NIM_API_KEY
general_settings:
master_key: os.environ/LITELLM_MASTER_KEY# 1) Run the proxy (listens on port 4000). Avoid LiteLLM 1.82.7 / 1.82.8.
export NVIDIA_NIM_API_KEY=nvapi-xxxxxxxxxxxxxxxxxxxx
export LITELLM_MASTER_KEY=sk-local-$(openssl rand -hex 12)
uv tool install 'litellm[proxy]'
litellm --config config.yaml
# 2) In a second terminal, point Claude Code at the proxy
export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=$LITELLM_MASTER_KEY
export ANTHROPIC_DEFAULT_HAIKU_MODEL=glm-5.3-flash # background tasks
export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 # avoids 400s from beta fields
claude --model glm-5.3Known rough edges: Claude Code’s /model picker only keeps gateway models whose ID contains “claude” or “anthropic”, so pass --model directly. An open LiteLLM issue documents a backend rejecting Claude Code’s output_config parameter. Remote Control and voice dictation are unavailable with a custom base URL. claude-code-router (@musistudio/claude-code-router) is the other popular translator and accepts any OpenAI-compatible endpoint, though its docs have no NVIDIA-specific example.
Our view: if you want Claude Code specifically, the proxy route adds a local service, a supply-chain dependency and a translation layer for tool calls, all to reach a trial API you are not allowed to use in production. For most people, OpenCode with its native NVIDIA provider gives a similar terminal-agent experience with none of that. Our guide to using Claude Code for free compares the local-model and OpenRouter routes.
What about Kilo Code and Zed?
Both work through a generic OpenAI-compatible setting. Kilo Code’s official blog walks through exactly this setup: API Provider “OpenAI Compatible”, base URL https://integrate.api.nvidia.com/v1, your key and a model ID, then “Make Active Profile”. One reported issue is worth knowing: some NVIDIA endpoints reject presence_penalty and frequency_penalty with HTTP 422 “Extra inputs are not permitted”. If you see that, remove those parameters from your profile.
Zed has a first-class openai_compatible provider in settings.json. You define an api_url and list each model with its max_tokens, and Zed reads the key from the UI or from an environment variable named after the provider. Tool use is enabled by default for these providers, and parallel tool calls are off by default, which suits a rate-limited endpoint.
What are the free-tier limits for coding agents?
NVIDIA does not publish them. Limits are set per model, can change with overall traffic, and there is no official way to raise them on the free tier. This is where most guides are out of date. The old model of 1,000 free credits (expandable to 5,000) is gone. An NVIDIA staff member wrote on the developer forum in September 2025 that build.nvidia.com “no longer use[s] a credit-based system” and that “the rate limits vary for each model, and we do not publish those”. A forum moderator added in September 2026 that there is “no official way” to get a rate-limit increase on the same tier.
The figure you will see everywhere, about 40 requests per minute, comes from community reports and forum thread titles, not from an NVIDIA specification. Treat it as a rough baseline that may be lower for popular models at busy times. For the full picture of what paid access costs, see our NVIDIA NIM pricing and limits guide.
What this means for coding tools:
- One agent at a time. Parallel subagents or several open Cline tasks share the same per-minute budget.
- No NIM autocomplete. Keystroke-driven completion is the fastest way to hit 429.
- Configure a fallback model. Because limits are per model, switching from GLM-5.3 to DeepSeek V4.1 Flash is often faster than waiting.
- Keep context lean. Smaller requests finish faster, which matters when each agent step waits on the previous one.
Is it safe to send your code to the free NVIDIA API?
It is fine for personal projects and open-source code, but read the trial terms before you point an agent at a client’s or employer’s repository. The NVIDIA API Trial Terms of Service (version dated September 19, 2025) grant access “for limited trial purposes only and without use of the API Service or Generated Content in production”, and allow use “for internal testing and evaluation purposes, not in production” unless you buy a subscription.
On data, the same terms say NVIDIA will not store or use your content at the end of each session, but may “log and store User Content and Generated Content to monitor for security or to prevent fraud or abuse”. A coding agent sends large parts of your codebase, configuration files and sometimes terminal output with every request. That can include secrets you did not mean to share. Add .env files and credentials to your tool’s ignore list, and do not use the trial for proprietary code you are not authorised to send to a third party. Teams in the EU should also check that this fits their internal AI-use and GDPR policies, which increasingly reference the EU AI Act framework.
When you outgrow the trial: production use needs NVIDIA AI Enterprise or a paid provider. Our NVIDIA NIM alternatives and best inference APIs guides compare pay-as-you-go options that host many of the same open models, and hosted vs self-hosted NIM covers running the containers yourself.
How do you fix common NVIDIA NIM errors in coding tools?
| Symptom | Likely cause | Fix |
|---|---|---|
| 401 / 403 “Authorization failed” | Key missing, mistyped or not loaded in this shell | Check the variable name your tool expects (NVIDIA_API_KEY vs NVIDIA_NIM_API_KEY) and restart the IDE after setting it |
404 on /v1/messages | An Anthropic-format client (Claude Code) is calling NVIDIA directly | Put a translating proxy in between, or use an OpenAI-format tool |
| 404 “Not found for account” or model not found | Model retired, renamed or not enabled for your account | Re-check /v1/models; switch to another ID from the table |
| 422 “Extra inputs are not permitted” | Tool sends parameters the endpoint rejects (often penalty settings) | Remove presence_penalty / frequency_penalty from the profile |
| 429 Too Many Requests | Per-model rate limit reached | One task at a time, no NIM autocomplete, switch to a fallback model, retry after a pause |
| Continue: Agent mode unavailable | Continue does not recognise the model as tool-capable | Add capabilities: [tool_use] to the model entry |
| Agent loop breaks after a few steps with Kimi K3 or other reasoning models | Client does not send back reasoning_content with tool calls | Use a non-reasoning model or GLM-5.3 with a lower reasoning_effort in that tool |
| Cursor’s own models stop working | “Override OpenAI Base URL” also reroutes Cursor-managed models | Toggle the override off when you switch back |
Which setup should you choose?
For most developers, the best free setup is VS Code with Continue (native NVIDIA provider) or Cline, plus OpenCode if you like working in the terminal. These tools connect directly to NVIDIA, support tool calling and need nothing beyond a config entry. Aider is the best choice if you prefer its git-centric, edit-format workflow and want the fewest tool-calling surprises. Cursor works for chat but not cleanly for its best features. Claude Code works only through a proxy that Anthropic does not support.
Treat the whole setup as a free testing ground, not free infrastructure. It is an excellent way to compare GLM, Kimi, DeepSeek and Nemotron on your own codebase before paying anyone. Once a model earns a permanent place in your workflow, move it to a paid endpoint with published limits. For a wider look at the tools themselves, see our ranking of the best AI coding agents in 2026 and our explainer on what the NVIDIA NIM API is. Polish-speaking readers new to the underlying concepts can start with LLM – co to jest.
FAQ
Can I use my free NVIDIA API key in Cline?
Yes. Set API Provider to “OpenAI Compatible”, Base URL to https://integrate.api.nvidia.com/v1, paste your nvapi- key and enter a model ID such as z-ai/glm-5.3. Cline has no native NVIDIA provider, so also set the context window manually in Model Configuration.
What is the base URL for the NVIDIA NIM API?
For NVIDIA’s hosted catalog it is https://integrate.api.nvidia.com/v1, used with the header Authorization: Bearer nvapi-… The endpoint follows the OpenAI Chat Completions format, and GET /v1/models lists available model IDs without a key.
Which NVIDIA NIM model is best for coding?
As of September 2026, z-ai/glm-5.3 is a strong default for agentic coding, with deepseek-ai/deepseek-v4.1-flash or z-ai/glm-5.3-flash as fast fallbacks and moonshotai/kimi-k3 for heavy tasks in tools that handle reasoning models. Check the live /v1/models list first, because the catalog changes often.
Does Claude Code work with NVIDIA NIM?
Not directly. NVIDIA’s hosted API has no Anthropic /v1/messages endpoint and returns 404, so you need a translating proxy such as LiteLLM or claude-code-router. Anthropic states that Claude Code does not support non-Claude models, so treat it as an unsupported experiment.
What is the NVIDIA NIM free tier rate limit?
NVIDIA does not publish it. Staff say limits vary by model and replaced the old credit system; community reports commonly cite about 40 requests per minute, but that is not an official figure and there is no official way to raise it on the free tier.
Can I use the free NVIDIA API for commercial or production work?
No. The NVIDIA API Trial Terms of Service allow limited trial, internal testing and evaluation use, not production. Production use requires a paid NVIDIA AI Enterprise subscription or another paid provider.
Why does Continue say Agent mode is unavailable with my NVIDIA model?
Continue has not detected tool support for that model ID. Add capabilities: [tool_use] to the model entry in config.yaml and reload. Continue’s docs recommend this for custom endpoints and newer models.
Sources
- NVIDIA. NIM LLM APIs reference (updated September 21, 2026), and live
GET https://integrate.api.nvidia.com/v1/modelsresponse, checked September 27, 2026. - NVIDIA. z-ai/glm-5.3 model reference.
- NVIDIA. moonshotai/kimi-k3 model reference.
- NVIDIA. DeepSeek V4.1 Flash model reference.
- NVIDIA. NVIDIA API Trial Terms of Service, version September 19, 2025.
- NVIDIA Developer Forums. Staff reply: credit system replaced by per-model rate limits, September 2025.
- NVIDIA Developer Forums. Moderator reply on free-tier rate-limit increases, September 2026.
- Cline. OpenAI Compatible provider configuration; PR #8721 (NVIDIA provider, closed unmerged).
- Continue. NVIDIA provider and model capabilities documentation.
- LiteLLM. NVIDIA NIM provider; Claude Code with non-Anthropic models; March 2026 security update.
- Aider. Other LLMs and model warnings.
- OpenCode. Providers documentation (NVIDIA, custom providers).
- Cursor. API keys help page; staff reply on base-URL override scope, August 2026.
- Anthropic. Claude Code LLM gateway and model configuration documentation.
- Kilo Code. Using NVIDIA NIM in Kilo Code, February 2026.
- Zed. Using your own API access.
- Roo Code. Sunset announcement, 2026.
- Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models, arXiv:2210.03629 (background on the tool-call loop used by coding agents).
- European Commission. AI Act regulatory framework.
