Works with your agent

tokdiet has no per-tool integration and no plugins to install. It's a local proxy that sits on the wire, so the only thing that matters is:

Does your tool let you override the model API base URL, and does it speak Anthropic Messages, OpenAI Chat Completions, or Gemini? If yes + yes → it works with tokdiet, zero tokdiet-specific code.

Point the tool at one of these and run it as usual:

Format	Set the base URL to
OpenAI Chat Completions	`http://localhost:7787/v1` (include the `/v1`)
Anthropic Messages	`http://localhost:7787` (no `/v1`)
Gemini	point the Gemini SDK base URL at `http://localhost:7787`

Three things that trip people up, true for almost every tool below:

The /v1 suffix belongs on OpenAI-format base URLs, not on Anthropic ones. A few tools append the path themselves — those are flagged.
A non-empty (dummy) API key is usually required even though tokdiet only forwards it upstream. Use any placeholder like sk-dummy.
Subscription / OAuth logins (Claude Pro/Max, ChatGPT, vendor cloud) are a different auth path locked to the vendor's host. There are no per-token charges to cut, so that's metering only, $0 saved — and usually not a clean proxy target. Use a pay-per-token API key to actually save.

By default tokdiet forwards to the real provider. To route an OpenAI/Anthropic-compatible client at a non-default upstream (e.g. MiniMax), set TOKDIET_OPENAI_UPSTREAM / TOKDIET_ANTHROPIC_UPSTREAM / TOKDIET_GEMINI_UPSTREAM on the tokdiet start process.

✅ Confirmed

opencode (sst/opencode)

Config file opencode.json, key provider.<id>.options.baseURL.

// ~/.config/opencode/opencode.json  (or ./opencode.json)
{ "$schema": "https://opencode.ai/config.json",
  "provider": { "tokdiet": { "npm": "@ai-sdk/openai-compatible", "name": "tokdiet",
    "options": { "baseURL": "http://localhost:7787/v1", "apiKey": "sk-dummy" },
    "models": { "gpt-4o": { "name": "gpt-4o" } } } } }
// Anthropic instead: "provider": { "anthropic": { "options": { "baseURL": "http://localhost:7787" } } }

options.baseURL must include /v1 for the OpenAI-compatible upstream; for the Anthropic override use the bare host (the SDK appends /v1/messages). Non-empty apiKey required. opencode caches provider config until restart. Verified: https://opencode.ai/docs/providers/

Aider

Env vars (passed through to LiteLLM). Prefix the model with openai/ or anthropic/ so LiteLLM routes to your custom base.

# OpenAI-compatible via tokdiet
export OPENAI_API_BASE=http://localhost:7787/v1
export OPENAI_API_KEY=sk-dummy
aider --model openai/gpt-4o
# --- OR Anthropic via tokdiet ---
export ANTHROPIC_API_BASE=http://localhost:7787    # no /v1; LiteLLM appends the path
export ANTHROPIC_API_KEY=sk-dummy
aider --model anthropic/claude-sonnet-4-5

OPENAI_API_BASE needs /v1; ANTHROPIC_API_BASE is the bare host. Non-empty key required. Verified: https://aider.chat/docs/llms/openai-compat.html

Continue.dev (IDE + `cn` CLI)

~/.continue/config.yaml, per-model apiBase with provider: openai.

models:
  - name: tokdiet-gpt4o
    provider: openai
    model: gpt-4o
    apiBase: http://localhost:7787/v1
    apiKey: sk-dummy

apiBase must include /v1. Use provider: openai (the generic OpenAI-compatible adapter) — its Anthropic block exposes base-URL less consistently. Same config serves the VS Code/JetBrains extension and the CLI. Verified: https://docs.continue.dev/customize/model-providers/top-level/openai

Cline (VS Code)

GUI settings (not env vars). Two routes:

OpenAI route:    API Provider = "OpenAI Compatible"
                 Base URL     = http://localhost:7787/v1   (include /v1)
                 API Key      = any non-empty dummy ; set Model ID
Anthropic route: API Provider = "Anthropic" → tick "Use custom base URL"
                 URL = http://localhost:7787 , API Key = dummy

Cline's separate "Claude Code"/subscription login is OAuth, locked to Anthropic's host — not a proxy target. Verified: https://docs.cline.bot/provider-config/openai-compatible

Roo Code

Same two GUI routes as Cline (Roo is a Cline fork).

OpenAI route:    API Provider = "OpenAI Compatible" , Base URL = http://localhost:7787/v1 , Key = dummy
Anthropic route: API Provider = "Anthropic" → "Use custom base URL" = http://localhost:7787 , Key = dummy

Note (issue #8488): the Anthropic custom-base-URL model list is restricted to claude-* ids — matters if you route a non-Claude upstream via TOKDIET_ANTHROPIC_UPSTREAM. Verified: https://docs.roocode.com/providers/openai-compatible

Kilo Code

GUI settings (Roo fork). Settings → Providers → "OpenAI Compatible", Base URL http://localhost:7787/v1, dummy key, set Model ID. Accepts either the /v1 base or a full .../v1/chat/completions URL. Verified: https://kilo.ai/docs/ai-providers/openai-compatible

Goose (Block / codename goose)

Env vars / ~/.config/goose/config.yaml. Goose's OPENAI_HOST is a host root — do NOT put /v1 on it (Goose appends the path itself).

export GOOSE_PROVIDER=openai
export OPENAI_HOST=http://localhost:7787      # host root, NO /v1
export OPENAI_API_KEY=dummy
export GOOSE_MODEL=<model>
# Anthropic alternative: GOOSE_PROVIDER=anthropic , ANTHROPIC_HOST=http://localhost:7787 , ANTHROPIC_API_KEY=dummy

No documented custom-Gemini host var was found (only GOOGLE_API_KEY), so Gemini-via-proxy on Goose is unconfirmed. Verified: https://block.github.io/goose/docs/getting-started/providers/

Zed (editor AI)

settings.json, language_models.openai_compatible.<name>.api_url (include /v1, no trailing slash). The key goes in the Agent Panel UI or an env var named after the provider id — not in settings.json.

"language_models": { "openai_compatible": { "tokdiet": {
  "api_url": "http://localhost:7787/v1",
  "available_models": [ { "name": "gpt-4o", "max_tokens": 128000 } ] } } }
// provider id "tokdiet" → set env var TOKDIET_API_KEY=sk-anything

Verified: https://zed.dev/docs/ai/use-api-access

JetBrains AI Assistant

Settings | Tools | AI Assistant | Providers & API keys → Third-party AI providers. Pick the OpenAI-compatible (LM Studio / LiteLLM-style) option, set the URL to http://localhost:7787/v1, click Test Connection. Key not mandated for local providers (dummy works if a field exists). Note: local models don't support MCP tool calls in AI Assistant, and default context window is 64k (adjustable). Verified: https://www.jetbrains.com/help/ai-assistant/use-custom-models.html

Open Interpreter

CLI flag --api_base (or interpreter.llm.api_base in Python), via LiteLLM.

interpreter --api_base http://localhost:7787/v1 --api_key dummy --model gpt-4o

OpenAI Chat Completions path; include /v1 yourself. LiteLLM requires a non-empty --api_key placeholder. (Flags belong to the 0.x CLI; the newer rewrite may differ.) Verified: https://docs.openinterpreter.com/language-models/local-models/custom-endpoint

llm (Simon Willison's CLI)

api_base in an extra-openai-models.yaml entry (in the dir from dirname "$(llm logs path)").

# extra-openai-models.yaml
- model_id: tokdiet-gpt4o
  model_name: gpt-4o
  api_base: "http://localhost:7787/v1"
# then: llm -m tokdiet-gpt4o 'hello'

When api_base is set, the OpenAI key is not sent by default → no dummy key needed unless your endpoint demands one. (For Anthropic you'd use the separate llm-anthropic plugin + ANTHROPIC_BASE_URL.) Verified: https://llm.datasette.io/en/stable/other-models.html

Crush (Charm)

providers.<id>.base_url in crush.json. Handles both protocols via type.

// ~/.config/crush/crush.json
{ "$schema": "https://charm.land/crush.json",
  "providers": { "tokdiet": {
    "type": "openai-compat", "base_url": "http://localhost:7787/v1",
    "api_key": "dummy", "models": [ { "id": "gpt-4o", "name": "gpt-4o" } ] } } }
// Anthropic-format: "type":"anthropic","base_url":"http://localhost:7787"

Use openai-compat (not openai) for third-party/proxied OpenAI traffic. base_url supports $VAR expansion. Non-empty api_key required by schema. Verified: https://github.com/charmbracelet/crush

pi (earendil-works/pi — the BYOK coding-agent CLI)

Not Inflection's consumer "Pi". This is @earendil-works/pi-coding-agent (formerly badlogic/pi-mono).

Base URL is set per-provider in ~/.pi/agent/models.json (no global env var). You can override a built-in provider by re-declaring its name.

// ~/.pi/agent/models.json — Anthropic-format route
{ "providers": { "anthropic": {
    "baseUrl": "http://localhost:7787", "api": "anthropic-messages", "apiKey": "$ANTHROPIC_API_KEY" } } }
// OpenAI route: "baseUrl":"http://localhost:7787/v1", "api":"openai-completions"

openai-completions needs /v1; anthropic-messages uses the bare host. apiKey is required (any value for keyless local endpoints). Verified: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/models.md

oh-my-pi / omp (can1357/oh-my-pi)

A separate "pi" coding agent. ~/.omp/agent/models.yml, providers.<id>.baseUrl. Lets you override the built-in openai/anthropic provider in place.

# ~/.omp/agent/models.yml
providers:
  openai:
    baseUrl: http://localhost:7787/v1
    apiKey: dummy        # resolved as an env-var NAME first, else a literal token
# Anthropic: providers.anthropic.baseUrl: http://localhost:7787 , api: anthropic-messages

openai-completions baseUrl needs /v1; anthropic-messages uses the bare host. For truly keyless endpoints set auth: none. Verified: https://github.com/can1357/oh-my-pi/blob/main/docs/models.md

Hermes Agent (NousResearch/hermes-agent)

~/.hermes/config.yaml model: { provider: custom, base_url, api_key }, or run hermes model → "Custom endpoint". The custom path is OpenAI-compatible.

# ~/.hermes/config.yaml
model:
  provider: custom
  base_url: "http://localhost:7787/v1"
  api_key: "dummy"

Anthropic-format goes through Hermes's named Anthropic provider (target http://localhost:7787), not the generic custom path. Verified: https://github.com/NousResearch/hermes-agent

⚠️ Partial / read the caveat

Cursor

Settings → Models → OpenAI API Key → enable "Override OpenAI Base URL" → http://localhost:7787/v1 → paste a dummy key → Verify. Big limits (from Cursor's own docs): (1) requests still route through Cursor's servers for prompt building; (2) custom keys only work with chat models — Composer/agent/inline-edit/apply/tab-completion stay on Cursor's backend and won't hit tokdiet. So you only meter/save on the Chat-Plan panel, a small slice of usage. Verified: https://cursor.com/help/models-and-usage/api-keys

Codex CLI (OpenAI)

~/.codex/config.toml (user-level only; project config is ignored for these keys).

model = "gpt-5.4"
model_provider = "tokdiet"
[model_providers.tokdiet]
name = "tokdiet proxy"
base_url = "http://localhost:7787/v1"
wire_api = "responses"          # the only accepted value now
env_key = "OPENAI_API_KEY"      # export OPENAI_API_KEY=sk-dummy

Critical: as of Feb 2026 Codex speaks only the OpenAI Responses API (/v1/responses); wire_api="chat" was removed. Codex is a clean tokdiet target only if tokdiet passes the Responses surface through — tokdiet's compaction targets Chat Completions, so Codex traffic may flow through unmetered/uncompacted, or you'd need a translating gateway (LiteLLM/OpenRouter) in between. Treat as experimental. Verified: https://developers.openai.com/codex/config-reference

gptme

OpenAI path only: OPENAI_BASE_URL env (or [env] in ~/.config/gptme/config.toml), model prefixed local/ or openai/.

OPENAI_BASE_URL="http://localhost:7787/v1" OPENAI_API_KEY=dummy gptme 'hello' -m local/gpt-4o

No documented ANTHROPIC_BASE_URL knob — route Anthropic-format through tokdiet is unconfirmed. /v1 required (not auto-appended). Verified: https://gptme.org/docs/providers.html

openclaw

openclaw is a real project (the "lobster" personal AI assistant; openclaw-code-agent adds coding via Claude Code / Codex / OpenCode backends) — but it has no first-class base-URL config of its own. It delegates to whichever backend you pick, so support reduces to that backend's knob:

export ANTHROPIC_BASE_URL=http://localhost:7787      # if backend = Claude Code
export OPENAI_BASE_URL=http://localhost:7787/v1      # if backend = Codex

Could not confirm an openclaw-specific base-URL key from docs; the vars above are the backends' own. Vendor-OAuth backends = metering only. Source (unconfirmed): https://github.com/openclaw/openclaw

❌ Not a fit (today)

Windsurf (Codeium / Devin Desktop)

BYOK lets you paste an Anthropic key but exposes no editable base URL, so you can't point native Windsurf at a loopback proxy. Non-BYOK usage is vendor-cloud metered (locked host) = $0 saved. Third-party "OpenAI-compatible URL" claims aren't in Windsurf's docs. Verified: https://windsurf.com/subscription/provider-api-keys

Inflection Pi (pi.ai / heypi.com)

The consumer chatbot, not a coding agent. No public API with an overridable base URL → not a tokdiet target. (If someone says "pi" as a coding agent, they mean earendil-works/pi or oh-my-pi, both ✅ above.) Verified: https://pi.ai/

The 10-second test for anything not listed

Search the tool's docs/config for any of: baseURL · base_url · apiBase · api_base · OPENAI_BASE_URL · OPENAI_API_BASE · ANTHROPIC_BASE_URL · OPENAI_HOST · endpoint · custom provider

Found one → set it to http://localhost:7787/v1 (OpenAI) or http://localhost:7787 (Anthropic), add a dummy key, run the tool, and watch the dashboard at http://localhost:7878. Traffic shows up = supported.

Notes were verified against each tool's official docs/repo on 2026-06-18. Tools move fast — if a flag changed, please open an issue or PR.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Works with your agent

✅ Confirmed

opencode (sst/opencode)

Aider

Continue.dev (IDE + `cn` CLI)

Cline (VS Code)

Roo Code

Kilo Code

Goose (Block / codename goose)

Zed (editor AI)

JetBrains AI Assistant

Open Interpreter

llm (Simon Willison's CLI)

Crush (Charm)

pi (earendil-works/pi — the BYOK coding-agent CLI)

oh-my-pi / omp (can1357/oh-my-pi)

Hermes Agent (NousResearch/hermes-agent)

⚠️ Partial / read the caveat

Cursor

Codex CLI (OpenAI)

gptme

openclaw

❌ Not a fit (today)

Windsurf (Codeium / Devin Desktop)

Inflection Pi (pi.ai / heypi.com)

The 10-second test for anything not listed

FilesExpand file tree

integrations.md

Latest commit

History

integrations.md

File metadata and controls

Works with your agent

✅ Confirmed

opencode (sst/opencode)

Aider

Continue.dev (IDE + cn CLI)

Cline (VS Code)

Roo Code

Kilo Code

Goose (Block / codename goose)

Zed (editor AI)

JetBrains AI Assistant

Open Interpreter

llm (Simon Willison's CLI)

Crush (Charm)

pi (earendil-works/pi — the BYOK coding-agent CLI)

oh-my-pi / omp (can1357/oh-my-pi)

Hermes Agent (NousResearch/hermes-agent)

⚠️ Partial / read the caveat

Cursor

Codex CLI (OpenAI)

gptme

openclaw

❌ Not a fit (today)

Windsurf (Codeium / Devin Desktop)

Inflection Pi (pi.ai / heypi.com)

The 10-second test for anything not listed

Continue.dev (IDE + `cn` CLI)