Continue.dev on a security model that answers

Point Continue at this API with a single models entry in config.yaml: provider openai, apiBase, a secret for the key and tool_use so Agent mode works. Covers VS Code, JetBrains and the cn CLI, what an Agent turn costs, and the errors you may hit.

Steps checked November 2, 2026 · prices as of November 17, 2026

You are mid-engagement, you ask your IDE assistant why a deserialization path in the client’s code is exploitable, and it declines. Continue fixes that with one entry in config.yaml: provider: openai, this API’s base URL, a key and the roles the entry fills. After that a security-tuned or abliterated build answers in Chat, edits inline and drives Agent mode in VS Code, JetBrains and the cn CLI. The same ids also work in Cursor, OpenCode and, via a router, in Claude Code.

What you need

Three things: Continue installed in VS Code or JetBrains (or the cn CLI), a key from the console, and write access to ~/.continue/config.yaml. On Windows the file is %USERPROFILE%\.continue\config.yaml. Continue writes it on first launch, so it exists before you touch it.

Continue’s docs have no dedicated “OpenAI-compatible” provider. Anything that speaks the OpenAI chat format is declared as provider: openai plus an apiBase override. This API speaks that format, so that is the whole trick. If you only want to see how a security build answers before you edit any config, the prompt box on the home page gives a few free turns and the playground runs the same catalog with no install.

The shortest path to a first answer

  1. Drop the key into ~/.continue/.env as a single unquoted line:
~/.continue/.env
1PENTESTLLM_API_KEY=sk-ptl-live-…

That file is where ${{ secrets.PENTESTLLM_API_KEY }} resolves from. Want a key scoped to one engagement folder instead? Put it in .env at the workspace root or in .continue/.env below it. Lookup order per the docs: those two, then the global file, then the process environment. The process environment only reaches the CLI; the IDE extensions cannot see shell variables at all.

  1. Paste these entries under models: in config.yaml. Leave the name, version and schema lines at the top alone. The three ids below are the ones Continue users on this API call most. Any other id from the models page takes the same shape: its display name from the catalog and, where the catalog lists one, its context length.
config.yaml
1models:
2 - name: GLM-5.3 Uncensored
3 provider: openai
4 model: glm-5.3-uncensored
5 apiBase: https://api.pentestllm.ai/v1
6 apiKey: ${{ secrets.PENTESTLLM_API_KEY }}
7 roles: [chat, edit, apply]
8 capabilities: [tool_use]
9 - name: GLM-5.3 Flash Uncensored
10 provider: openai
11 model: glm-5.3-flash-uncensored
12 apiBase: https://api.pentestllm.ai/v1
13 apiKey: ${{ secrets.PENTESTLLM_API_KEY }}
14 roles: [chat, edit, apply]
15 capabilities: [tool_use]
16 - name: DeepSeek V4.1 Flash Abliterated
17 provider: openai
18 model: deepseek-v4.1-flash-abliterated
19 apiBase: https://api.pentestllm.ai/v1
20 apiKey: ${{ secrets.PENTESTLLM_API_KEY }}
21 roles: [chat, edit, apply]
22 capabilities: [tool_use]
  1. Save the file. The three names show up in the model picker above the chat input. Choose one, then set the mode below the input to Chat, Plan or Agent (Cmd/Ctrl + . steps through them). Edit and Apply pick up the same entries through the roles line.

Prefer the key in the file? Replace the secret reference on apiKey with the full sk-ptl- key; the docs allow either. Still on the deprecated config.json? The one-model equivalent is below, with title where YAML has name and no roles:

config.json
1{
2 "models": [
3 {
4 "title": "GLM-5.3 Uncensored",
5 "provider": "openai",
6 "model": "glm-5.3-uncensored",
7 "apiBase": "https://api.pentestllm.ai/v1",
8 "apiKey": "sk-ptl-live-…"
9 }
10 ]
11}

The terminal client cn loads the same config.yaml by default. Hand it another YAML file for a single session with cn --config, or type /config in the TUI to flip between the local configs it has found:

shell
1cn --config ~/.continue/config.yaml

Where to find the file from the IDE

Open the Continue panel (Cmd/Ctrl + L in VS Code, Cmd/Ctrl + J in JetBrains). Click the agent selector over the chat input, hover your agent and hit the gear. The configs dropdown in the top right corner of the input reaches the same file through the cog next to Local Config. Because cn reads that file too, one edit configures the extension and the terminal at once. config.json next to it is the legacy format and the docs mark it deprecated; use the YAML.

Each line of the entry

name
the display name
what the picker shows; no two entries in the file may share one
provider
openai
the wire format; Continue talks OpenAI chat completions to whatever apiBase sits under it
model
glm-5.3-uncensored
the id on the wire, spelled the way the models page spells it
apiBase
https://api.pentestllm.ai/v1
the server that replaces OpenAI; ends at /v1, with no /chat/completions after it
apiKey
${{ secrets.PENTESTLLM_API_KEY }}
the full sk-ptl- key, pasted or named as a secret; goes out as the bearer token
roles
[chat, edit, apply]
the places this entry is selectable: chat and Agent, inline edits, writing a diff. autocomplete, embed and rerank are the remaining roles
capabilities
[tool_use]
declares tool-call support, which Plan and Agent require; it adds to Continue’s own detection and never removes anything
defaultCompletionOptions
contextLength
the window Continue fits prompts into; maxTokens, temperature and stop share this map
requestOptions
timeout, caBundlePath, headers
HTTP settings for this entry alone; a corporate certificate authority is handled here

Which build goes first

Under provider: openai Continue emits OpenAI-format requests, and every build in the catalog accepts them. What drives cost is the shape of an Agent turn: it resends the whole conversation, every file Continue opened and every tool result so far, then gets a short edit back. Input tokens dominate, so the build with the lowest input price costs least per turn. The right-hand column records whether the build accepts native tool calls. Continue documents a fallback to system-message tools for models that do not, which is why capabilities: [tool_use] stays on each entry regardless.

BuildCost of one Agent turnNative tool calls
GLM-5.3 Uncensored$0.01model page
GLM-5.3 Flash Uncensored$0.0047model page
DeepSeek V4.1 Flash Abliterated$0.04model page
DeepSeek V4 Pro Uncensored$0.04model page
GLM-5.3 Abliterated$0.09model page
GLM-5.3 Flash Code Uncensored$0.25model page

A turn here is 40,000 tokens in and 1,200 out, priced from the catalog as of 2026-11-17. Tool-call support is whatever the catalog states; when it states nothing, the model page has the answer.

Uncensored models and their prices per million tokens
GLM-5.3 Uncensoredglm-5.3-uncensoredUncensored$0.282$0.846
GLM-5.3 Flash Uncensoredglm-5.3-flash-uncensoredUncensored$0.1053$0.3685
DeepSeek V4.1 Flash Abliterateddeepseek-v4.1-flash-abliteratedAbliterated$0.8488$1.6975
DeepSeek V4 Pro Uncensoreddeepseek-v4-pro-uncensoredUncensored$0.97$1.94
GLM-5.3 Abliteratedglm-5.3-abliteratedAbliterated$2.1825$4.365
GLM-5.3 Flash Code Uncensoredglm-5.3-flash-code-uncensoredUncensored$6.00$6.00

Prices per million tokens, read from the live catalog on 2026-11-17.

The roles chat, edit and apply cover Chat, Plan and Agent, the inline edit, and the step that lands a proposed diff in the file. autocomplete is a different animal: its own prompt template, its own request shape. Continue’s example config pairs a chat model with a separate autocomplete model, and that is the right split here too. Do not add autocomplete to these entries; keep whatever serves it now. embed and rerank are indexing roles and also stay as they are.

Reviewing code for vulnerabilities all day? The coding guide compares the code builds and prices a full day on each. The ids here also drop into Cursor and OpenCode with their own config shapes, and the coding AI page explains why stock assistants bail on security code review in the first place.

Confirm it works

Ask the API which models the key may call. Each id you added under models: should come back:

shell
1curl https://api.pentestllm.ai/v1/models -H "Authorization: Bearer sk-ptl-live-…"

Now fire a single request shaped like Continue’s. The -i flag shows the response headers, which carry a request id and, on a non-streamed reply, the charge in USD.

shell
1curl -i https://api.pentestllm.ai/v1/chat/completions \
2 -H "Authorization: Bearer sk-ptl-live-…" \
3 -H "Content-Type: application/json" \
4 -d '{"model": "glm-5.3-uncensored", "messages": [{"role": "user", "content": "Reply with the single word ok."}]}'

Back in the IDE, pick one of your entries and send a Chat message. Then switch to Agent and give it a small, in-scope job, say “add a CVSS vector line under each finding in findings.md”. Agent pauses for approval before each tool run by default; the tool policies in the docs can auto-approve or exclude individual tools. Plan behaves the same way with read-only tools, which makes it the right mode for watching a new build walk a target repository before you let it change anything. Every request lands on the console’s usage page with token counts and charge.

When something fails

  • 401 invalid_api_key: bad or revoked key, or the secret never resolved. Remember that the extensions cannot read a key exported in your shell; it has to sit in ~/.continue/.env under the name ${{ secrets.PENTESTLLM_API_KEY }} refers to. Or copy a fresh key from the console.
  • 404 model_not_found: the value on the model line is not an id from the models page. Compare the spelling; the id belongs in model, never in name.
  • A bare 404 that names no model: apiBase lost its /v1, or gained a /chat/completions. Continue appends the route on its own; stop apiBase at /v1.
  • 402 insufficient_balance: the balance cannot cover this call; add credit in the console. A 402 monthly_limit_reached instead means the key reached the monthly spend ceiling you gave it.
  • 403 scope_denied: this key is restricted to a model list and the requested id is outside it.
  • 429 rate_limited: the key is over its rate limit. Read retry-after on the response.
  • Plan or Agent shows Not Supported and will not switch on: Continue found no tool support on the entry. The documented remedy is capabilities: [tool_use], already present in the block above. A build whose model page says it takes no native tool calls runs through Continue’s system-message tools instead.
  • Your entries are absent from the picker: the configs dropdown is pointed at a different config. Set it back to Local Config, the one that maps to ~/.continue/config.yaml.
  • fetch failed on a client network with its own CA, common on an internal engagement: set requestOptions.caBundlePath on each entry to the exported certificate, as Continue’s FAQ describes.
  • Requests hitting /completions rather than /chat/completions: the entry has useLegacyCompletionsEndpoint: true. Delete it; this API serves chat completions.

Questions

provider: openai, with apiBase pointing at the endpoint’s /v1 URL. There is no separate OpenAI-compatible provider in Continue; the docs send every server that answers the OpenAI chat format through openai plus an apiBase override, and this API qualifies.

On the model entry’s apiKey line: either the full key, or ${{ secrets.NAME }} with NAME set in ~/.continue/.env, a workspace .env or .continue/.env. Only the cn CLI can read a shell variable; the VS Code and JetBrains extensions rely on those .env files.

No. This API already answers the OpenAI format under its own key, so the entry above connects Continue to it directly. LiteLLM earns its place only when you want several providers behind a single local URL. See the LiteLLM guide.

Yes. Plan and Agent depend on tool calls, so every entry declares capabilities: [tool_use]. Builds whose model page lists native tool support take them directly; for the others Continue’s docs describe a system-message fallback. The table above prices one Agent turn on each build at today’s catalog rate. See the models page.

Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Help