Zed with an uncensored model

Where the settings are

Zed keeps the provider in its user settings file: ~/.config/zed/settings.json on macOS and Linux, %APPDATA%\Zed\settings.json on Windows. Open it with zed: open settings file from the command palette. The project file .zed/settings.json takes editor and language options only, so the provider goes in the user file.

The same provider can be added from the settings UI: agent: open settings (also in the Agent Panel’s top-right menu) opens Settings → AI → LLM Providers, where Add Provider asks for a provider name, an API URL, a model ID and a context window. The key is entered on that page too, or read from an environment variable. It never goes in settings.json.

Steps

1. Add the provider to settings.json. Zed needs max_tokens for every entry, so the three ids under available_models are the first of the builds below whose context window the catalog publishes; any id on the models page goes in the same way, with the window its page states. The agent.default_model block selects the first one for new threads:

2. Create a key in the console and give it to Zed one of two ways. In the UI: agent: open settings, find uncensoredllm under LLM Providers and paste the key; Zed stores it in the system keychain. Or export it in the environment Zed starts from, under the name Zed derives from the provider id:

3. Open the Agent Panel (agent: new thread, or the sparkle icon in the status bar) and send a message. The model selector on the message editor (cmd-alt-/ on macOS, ctrl-alt-/ on Linux and Windows) switches to another id under uncensoredllm; the choice is saved back to agent.default_model.

A non-empty environment variable wins over a keychain entry. To stop using a key that came from the variable, unset it and restart Zed.

What each setting does

openai_compatible.uncensoredllm

the provider id

the name in the model selector, the provider value in agent.default_model and the stem of the env var name

api_url

https://api.uncensoredllm.ai/v1

where the requests go; the /v1 at the end matters, and nothing after it

available_models[].name

deepseek-v4.1-flash-1017-uncensored

the model id sent in each request, exactly as the models page shows it; Zed lists only the ids written here

available_models[].display_name

the name in the picker

optional; the id shows when it is missing

available_models[].max_tokens

the context window

required; Zed compacts a thread at 90% of it by default, so the catalog’s number keeps compaction honest

available_models[].max_output_tokens

the output limit

optional; sent as max_completion_tokens on each request

available_models[].capabilities

{ tools, images, parallel_tool_calls, prompt_cache_key }

tools defaults to true and is what the agent needs for edits and shell commands; images follows the catalog; the other two stay false here

agent.default_model

{ provider: uncensoredllm, model: deepseek-v4.1-flash-1017-uncensored }

the model new threads start on; inline_assistant_model, commit_message_model, thread_summary_model and compaction_model take the same shape

UNCENSOREDLLM_API_KEY

the whole sk-ullm- key

the env var Zed reads for this provider id; the keychain entry from the settings UI is the alternative

Picking a model

Zed sends OpenAI-format chat completions to the api_url, and every build on the models page answers in that format. An agent turn carries the thread, the files Zed read and every tool result so far, so input tokens are most of the bill and a build with a low input price costs the least per turn. The last column says whether the build takes tool calls; a model without them gets a No tools label in the selector, and the agent cannot edit files or run commands with it.

The context window matters twice. Zed compacts a thread when it reaches 90% of the entry’s max_tokens (the agent.auto_compact threshold), so a build with a long window goes longer between compactions and keeps more of the thread verbatim. A window under 80,000 tokens cannot auto-compact at all; Zed shows a banner near the limit and offers a new thread seeded with a summary instead. Compaction itself is a request to the thread’s model, billed like any other, unless agent.compaction_model points it at a smaller build.

40,000 input and 1,200 output tokens at the catalog prices on 2026-11-09. Tool calls as the catalog publishes them; where it does not, the model page says.

For coding specifically, the uncensored coding AI guide compares the code builds and costs a full coding day on each. The same ids work in Cursor, OpenCode and, through a router, Claude Code.

Check it

List the models your key can call; every id you put under available_models should be in the answer:

Then send one request the way Zed does. -i prints the headers: x-uplink-request-id names the request, and on a non-streamed answer x-uplink-charge is what it cost in USD.

In Zed, send a message in a new thread with one of your ids selected. The request shows on the console’s usage page with its tokens and its charge. The thread’s token count shows near the profile selector in the panel’s message editor and climbs against the max_tokens you set.

Errors and fixes

• 401 invalid_api_key: the key is wrong or revoked, or an old value of UNCENSOREDLLM_API_KEY is still set in Zed’s environment and is overriding the keychain entry. Copy the key again from the console, or unset the variable and restart Zed.
• 404 model_not_found: a name under available_models is not on the models page. Check the spelling against the page; the id is name, not display_name.
• 402 insufficient_balance: the balance does not cover the request. Top up in the console. 402 monthly_limit_reached means the key hit the monthly spend limit you set on it.
• 403 scope_denied: the key is limited to certain models and this is not one of them.
• 429 rate_limited: the key hit its rate limit. The response carries retry-after.
• LLM provider is not configured or does not support the configured model: the provider in agent.default_model (or inline_assistant_model) is not the key under openai_compatible, the model is not a name in that provider’s available_models, or the provider has no key yet. Make the three match and enter the key. Zed issue #63878 (open, September 2026) reports the message from the inline assistant while the same provider and model work in the Agent Panel.
• No tools next to the model: capabilities.tools is false for that entry. Set it to true for a build the model page lists with tool calls.
• The provider’s settings cannot be edited from the UI: as of Zed issue #35461 (open), the LLM Providers page changes a custom provider’s key only. Change the URL, the ids or the context windows in settings.json.
• Compaction or thread summaries fail with an error about reasoning_effort: Zed sends "reasoning_effort": "none" on those requests when a model entry declares a reasoning_effort (Zed issue #60411, open, November 2026). Leave the field out of these entries.
• Two providers on the same api_url share one key: Zed stores the key per URL, so a second block pointing at https://api.uncensoredllm.ai/v1 with a different key keeps asking for it (issue #41585). One provider block with all the ids is enough; a key limited to certain models is a scope on the key, not a second provider.

Questions