basemode

Make any LLM do raw text continuation.

basemode coerces chat-tuned models into clean next-token continuation mode (instead of assistant-style replies), with strategy selection handled per model/provider.

The problem

Ask a chat-tuned model to continue a piece of text and it answers you instead:

> The ship rounded the headland and

I'd be happy to help continue this passage! Here's one possibility:

"The ship rounded the headland and the harbour opened before them..."

Would you like me to continue in a different style?

You wanted prose. You got a customer-service interaction wrapped around it.

basemode gets you the prose:

$ basemode "The ship rounded the headland and"
The ship rounded the headland and the wind dropped all at once, as though
the cliffs had swallowed it. Sails that had been drum-tight for three days
went slack, and in the sudden quiet the crew could hear water moving along
the hull.

No preamble, no offer to help, no quotation marks. Just the continuation.

How

Base models do this natively; chat-tuned models have it trained out of them. basemode puts it back by wrapping the prefix in whatever prompt shape a given model will actually continue — a prefilled assistant turn, a system instruction, a few-shot frame, an FIM template — because no single trick works everywhere. Anthropic models took a prefill until they didn't; some providers reject the parameter outright.

So the strategy is per-model data, not a guess. basemode bench scores each strategy against a model on real continuations and pins the winner; the results ship in the verified-model table below.

$ basemode info claude-opus-5

┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Field                  ┃ Value                                               ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ Resolved               │ anthropic/claude-opus-5                             │
│ Strategy               │ few_shot                                            │
│ Strategy source        │ verified models registry                            │
│ Quirks                 │ no_prefill, no_temperature                          │
│ Input price            │ $5.00/1M                                            │
│ Output price           │ $25.00/1M                                           │
│ Your rating            │ unrated                                             │
│ Observed health        │ 181 attempts, 1 failed (1%); last empty_response at │
│                        │ 2026-08-24T20:49:09.033360+00:00                    │
└────────────────────────┴─────────────────────────────────────────────────────┘

Strategy, quirks, pricing, and your own observed success rate for that model, in one place — so a bad continuation is diagnosable rather than mysterious.

Install

pip install basemode

Set provider keys via environment variables or .env (for example OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY, GEMINI_API_KEY, GROQ_API_KEY, TOGETHER_API_KEY).

Quickstart

# Single continuation (default model if configured, else fallback)
basemode "The ship rounded the headland and"

# Parallel continuations
basemode "The ship rounded the headland and" -n 3

# Inspect selected strategy and pricing metadata
basemode info claude-sonnet-4-6

# Rank coercion strategies for a model by how cleanly they continue text
basemode bench claude-sonnet-4-6

# Show only key-configured models
basemode models --available

CLI

basemode --help
basemode run --help
basemode models --help
basemode info --help
basemode strategies --help

Useful commands:

• basemode run (default): stream continuation text
• basemode models: list models (supports --verified and --json for picker UIs)
• basemode providers: list provider IDs
• basemode info: show normalized model + prompt strategy (and where that choice came from) + pricing metadata
• basemode bench: rank coercion strategies for a model, and pin the winner
• basemode default: get/set your default model
• basemode keys: manage stored API keys

Choosing a strategy

Coercion either works or it doesn't, and basemode scores which. basemode bench runs each strategy against the model over four probe prefixes — prose, technical, poetry, dialogue — and scores every result for assistant behavior: preamble, refusal, echoed prefix, chat turns, stray formatting.

basemode bench claude-opus-5 --samples   # compare
basemode bench claude-opus-5 --save      # pin the winner for this model

┏━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┓
┃ Strategy ┃ Score ┃ Flags / error        ┃ Mean s ┃
┡━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━┩
│ few_shot │ 0.93  │ quoted               │ 2.07   │
│ system   │ 0.75  │ empty                │ 2.24   │
│ prefill  │ 0.00  │ BadRequestError: ... │ 0.30   │
└──────────┴───────┴──────────────────────┴────────┘

At runtime the strategy comes from the first of: an explicit --strategy, a local pin, the prompt method verified for that model in the table below, then model-name heuristics. basemode info MODEL shows which one applied.

Server

basemode serve exposes an OpenAI-completions-compatible POST /v1/completions endpoint backed by continue_text/branch_text, so llama.cpp-style clients (e.g. Tapestry Loom) can drive basemode's model-coerced continuations as a local backend instead of needing a locally-hosted base model.

pip install 'basemode[server]'
basemode serve --port 8080

Point a client's "OpenAI-style Completions" endpoint at http://127.0.0.1:8080/v1/completions. Requests accept model, prompt, max_tokens, temperature, n (parallel branches), echo, and strategy; stream is not supported (responses are always synchronous JSON). No logprobs are returned, since basemode routes through chat-completion coercion rather than a real base model.

Python API

from basemode import continue_text, branch_text

async for token in continue_text(
    "The ship rounded the headland and",
    model="gpt-4o-mini",
    max_tokens=120,
):
    print(token, end="", flush=True)

async for idx, token in branch_text(
    "The ship rounded the headland and",
    model="gpt-4o-mini",
    n=3,
    max_tokens=80,
):
    print(idx, token, end="", flush=True)

Docs

Full docs are in docs/ and can be served with MkDocs:

make docs-serve

Then open http://localhost:8001.

Integration Health Checks

Run live provider checks (real APIs, key-aware skips):

uv run pytest -m integration tests/test_integration.py -q

This writes a machine-readable report to dist/integration/provider_health.json with per-model status, latency, token estimates, and estimated USD cost.

Verified Models

Single generated table, refreshed by CI.

Legend: ✓ = LiteLLM pricing present and release date available; ⚠ = missing/approximate field or known issue.