mini-SWE-agent

mini-SWE-agent has no base-URL environment variable — the endpoint goes under model.model_kwargs.api_base in a YAML config, which mini hands straight to litellm.completion.

# mini has no base-URL environment variable and no settings UI. The endpoint
# goes in an agent config file, under model.model_kwargs — which mini's docs
# describe as "directly passed to litellm.completion".
model:
  model_name: "openai/claude-sonnet-5"
  model_kwargs:
    custom_llm_provider: "openai"
    api_base: "https://api.kunavo.com/v1"   # keep the /v1
  litellm_model_registry: "kunavo-registry.json"   # see "Cost tracking" below

# The key does not live in this file. With custom_llm_provider: "openai",
# litellm reads OPENAI_API_KEY, and mini documents two ways to set it:
#
#   export OPENAI_API_KEY=sk-kn-...          # environment, wins over .env
#   mini-extra config set OPENAI_API_KEY sk-kn-...   # mini's own .env
#
# Then run it:  mini -c kunavo.yaml
# Or make it the default:  mini-extra config set MSWEA_MINI_CONFIG_PATH kunavo.yaml

Step by step

1. Create a key at /app/keys and copy it — it is shown once.
2. Install and run once so the paths exist: pip install mini-swe-agent, then mini. The first run prints where its .env and agent config live, and offers mini-extra config setup.
3. Put the key where litellm will look for it: export OPENAI_API_KEY=sk-kn-..., or persist it with mini-extra config set OPENAI_API_KEY sk-kn-.... mini notes that “Environment variables take precedence over variables set in the .env file”, which is the usual reason a key you just changed appears not to have changed.
4. Save the YAML above as kunavo.yaml beside your other agent configs, and add the price registry from the section below — without it the run stops on a cost-calculation error rather than on a bad answer.
5. Start it with mini -c kunavo.yaml, or mini -c kunavo.yaml -m openai/claude-haiku-4-5 to override the id for one run. mini opens in confirm mode, where you approve each command — a good default for a first run against a new endpoint.
6. Give it a task that actually runs a command, not a greeting. mini's v2 default is native tool calling and its shipped prompt insists that “Every response needs to use the 'bash' tool at least once to execute commands” — so one real tool round-trip is what tells you the pairing works. If tool calls come back empty or malformed, mini still ships the older text-parsing path: mini -c mini_textbased.yaml, or model_class: litellm_textbased in your own file.

Checked against mini-SWE-agent's local models guide on September 21, 2026. Third-party settings move; if a field name here no longer matches what you see, that page is the authority, not this one.

Verify before you debug the client

One request settles whether a failure is the endpoint, the key, or the configuration file. If this returns JSON, the same base URL and key work in mini-SWE-agent.

# Settles whether a failure is the endpoint, the key, or the client.
curl -sS https://api.kunavo.com/v1/models \
  -H "Authorization: Bearer sk-kn-..."

Which model id to put in the field

Every text model is reachable as a model id — the live list is GET /v1/models, and the catalog with prices is on the models page. Rates are USD per 1M tokens, input / output.

Cost tracking, which is not optional here

mini's shipped mini.yaml carries cost_limit: 3. — a per-run ceiling in dollars — and that ceiling is enforced by litellm's cost calculator, which prices a run by looking the model id up in its registry. Kunavo's ids are not in that registry, so the first thing most people see is not a bad answer but an error: mini's own troubleshooting page shows it as Exception: This model isn't mapped yet. model=…, custom_llm_provider=….

There are two ways out and they are not equivalent. The global switch MSWEA_COST_TRACKING="ignore_errors" (or cost_tracking: "ignore_errors" in the file) removes the guard rather than repairing it, and mini labels it “CAREFUL: This can lead to unmanaged spending!” The other way is to tell litellm the rates, which is what litellm_model_registry in the setup block points at. The rates below are this site's live catalog rates, converted to litellm's per-token format:

{
  "claude-sonnet-5": {
    "input_cost_per_token": 0.000002,
    "output_cost_per_token": 0.00001,
    "litellm_provider": "openai",
    "mode": "chat"
  },
  "claude-opus-5": {
    "input_cost_per_token": 0.000002,
    "output_cost_per_token": 0.00001,
    "litellm_provider": "openai",
    "mode": "chat"
  },
  "claude-haiku-4-5": {
    "input_cost_per_token": 0.0000004,
    "output_cost_per_token": 0.000002,
    "litellm_provider": "openai",
    "mode": "chat"
  }
}

1. Model names are matched exactly and case-sensitively, and mini's example keys the entry on the name without its provider prefix — so claude-sonnet-5 here, even though the config says openai/claude-sonnet-5.
2. litellm_provider has to agree with the prefix and with custom_llm_provider. mini's warning is explicit: “If you use the custom_llm_provider or have a provider prefixed to the model name (e.g., openai/…), then this must also match litellm_provider in the config!”
3. The path can also come from LITELLM_MODEL_REGISTRY_PATH instead of the config key — useful for the batch runners, e.g. LITELLM_MODEL_REGISTRY_PATH=kunavo-registry.json mini-extra swebench …
4. These rates are a budget input, not a bill. What you are actually charged is what your Kunavo balance records; re-copy them if the catalog moves, or read them off GET /v1/models.