Use tokenflow in your coding agent

Your key

Export the API key we sent with your invitation in your shell profile. Both setups below read it from this variable.

export TOKENFLOW_API_KEY="tfk_…"

Claude Code

Claude Code speaks the Anthropic Messages API, which tokenflow serves natively. Add a shell function so your normal claude keeps talking to Anthropic and claude-tf talks to tokenflow.

1. Add a launcher

claude-tf() {
  ANTHROPIC_BASE_URL=https://api.tokenflow.dev \
  ANTHROPIC_AUTH_TOKEN="$TOKENFLOW_API_KEY" \
  ANTHROPIC_MODEL=qwen3.8-27b \
  ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3.8-27b \
  ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3.8-27b \
  ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3.8-27b \
  CLAUDE_CODE_SUBAGENT_MODEL=qwen3.8-27b \
  CLAUDE_CODE_MAX_CONTEXT_TOKENS=131072 \
  CLAUDE_CODE_AUTO_COMPACT_WINDOW=120000 \
  CLAUDE_CODE_DISABLE_THINKING=1 \
  CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
  claude "$@"
}

2. Cap the effort level

Add this to ~/.claude/settings.json. It only applies when the model is qwen3.8-27b, so your Claude sessions are unaffected.

{
  "modelSettings": {
    "qwen3.8-27b": { "effortLevel": "medium", "maxEffortLevel": "medium" }
  }
}

3. Run it

claude-tf

Why these settings

• Every model tier maps to one model. Claude Code sends background requests (titles, summaries, subagents) to its Haiku, Sonnet and Opus tiers. Pointing all of them at qwen3.8-27b keeps every request on tokenflow.
• Context and compaction. The model has a 131,072-token context. Compacting at 120,000 leaves room for the reply, so long sessions don’t fail at the limit.
• Effort capped at medium. Qwen3.8 accepts low, medium and xhigh, not high. Claude Code’s helper requests ignore --effort and default to high for models it doesn’t know; only maxEffortLevel clamps them. Passing your own --settings file replaces this cap.
• Thinking off. Faster turns for agent work. This is the only thinking setting we test.
• Claude Code may log unrecognized_model for qwen3.8-27b. It’s informational.

Codex

Codex uses the OpenAI Responses API, which tokenflow also serves. Put tokenflow in its own profile file so your default Codex setup stays as it is.

1. Create a profile

model = "qwen3.8-27b"
model_provider = "tokenflow"
model_reasoning_effort = "none"
model_context_window = 131072
model_auto_compact_token_limit = 95000

[model_providers.tokenflow]
name = "tokenflow"
base_url = "https://api.tokenflow.dev/v1"
wire_api = "responses"
env_key = "TOKENFLOW_API_KEY"

2. Run it

codex --profile tokenflow

# non-interactive, e.g. in a script or CI job
codex exec --profile tokenflow "fix the failing test" </dev/null

Why these settings

• Profile file, not a table. Codex 0.160 reads profiles from ~/.codex/<name>.config.toml and rejects the older [profiles.<name>] table in config.toml.
• Context window set explicitly. Codex has no metadata for qwen3.8-27b and would otherwise assume a much larger window. Compacting at 95,000 keeps sessions well inside 131,072.
• Reasoning effort none. Faster turns. Set "medium" to turn reasoning on.
• </dev/null in scripts. Without a terminal, codex exec waits on standard input; closing it lets the run start.
• The warning Model metadata for qwen3.8-27b not found is expected; the settings above supply what it needs.

Other tools

Any tool that lets you set an OpenAI- or Anthropic-compatible endpoint can use tokenflow. These are the three values to enter:

We’ve only listed the agents we test end to end. If you run tokenflow in another one, tell us at support@tokenflow.dev and we’ll add it.