Choose models and providers

A Mecatl session runs with a provider and a base model selected by the server and, optionally, by the client. The server returns the effective selection and input capabilities when the session is created.

Choose the path that matches how you use Mecatl:

• use mecatui for interactive selection;
• use the CLI to configure a server or one-shot run; or
• use the API when your client creates sessions directly.

For the rest of the terminal workflow, see Use mecatui.

Mecatui journey

When the connected server advertises model selection, type /models in mecatui. Filter the server's inventory and use the keyboard or a primary click on a visible model row to move the selection. The mouse wheel scrolls the viewport while the selection remains pinned. Press Enter to switch to the selected model. The picker warns that the choice creates a peer session and carries over the visible conversation. Replaying a long history may be costly. The existing session's provider and base model do not change.

A switch across providers keeps the visible conversation but drops provider- private replay state, such as reasoning state that the new provider cannot understand. The new session's provider, model, and capabilities are reported by the server.

When broker OAuth is enabled, this peer is also a new broker session. Protected MCP enrollment is session-scoped, so switching models may require enrolling the protected backends again; authorization is not silently copied from the old session.

Type /effort to choose a reasoning-effort tier. mecatui applies a changed tier by creating a peer session with the same provider, model, and conversation. The picker is available only when the connected server advertises the relevant capability.

mecatui model choices are server-backed. In embedded mode, the local server's configuration and credentials determine the inventory. In connect mode, the remote server determines it; local embedded-server flags and credentials do not apply.

See Use mecatui for the command-line startup, connection, and keybinding details.

CLI journey

Set up a local provider

Run mecatui providers setup [PROVIDER] to configure the embedded server. Without a name, choose from the listed providers. It does not start a server or open the TUI. mecatui connect ADDRESS always uses the remote server's configuration.

For API-key providers, setup can reuse an effective credential without copying an environment value to disk, or accept a replacement through hidden terminal input. It explains where to obtain provider-specific keys; API use may incur charges. Keys are never command arguments, and empty, invalid, or control-character values are rejected before saving. Saving requires confirmation.

Environment credentials or an operator-selected owner-only API-key file supply API keys. The file is plaintext, so its permissions are not encryption: same-UID processes, including permitted agent Shell commands, can read it. Do not put a secret in flags, prompts, settings, or logs. Use an environment variable or the operator-managed credential file when interactive entry is unsuitable. See Run mecated standalone for credential-file and daemon configuration details.

After setup, you may set the embedded default with mecatui providers set-default PROVIDER [MODEL]; declining leaves the current default unchanged. mecatui providers and mecatui providers status [PROVIDER] report local configuration without revealing credentials. Presence does not prove model access, billing, or account health.

openai-codex is distinct from the public openai API-key provider. Setup can reuse a locally usable manual Codex subscription token for default selection but never requests, writes, refreshes, imports, or removes that token. See the manual subscription-token deployment guidance.

Use providers add PROVIDER to define a custom provider, login PROVIDER to manage its locally owned credential, and logout or remove to remove it. Custom OIDC login can use --no-browser on a headless host. ToolHive credentials and lifecycle are external: use thv llm tooling, not these provider commands.

Endpoint overrides

The built-in provider endpoint flags (--openai-base-url, --openrouter-base-url, --anthropic-base-url, and --opencode-base-url) are non-secret configuration. They override the matching operator provider_overrides setting; settings override the built-in endpoint. Custom provider URLs remain defined only by their provider definition. OpenAI and Anthropic keep their SDK endpoint when neither source supplies an override.

Configure a server default

Use deployment flags when every session on a server should start from the same provider and model:

mecated serve \
  --default-provider openai \
  --default-model gpt-5.6-terra

gpt-5.6-terra is an example model ID. Model IDs are provider-specific, so the same example is valid only when that server's OpenAI provider can use it. The server's built-in OpenAI default remains available when no explicit model is configured.

For a zero-selector session, server-side resolution is separate for provider and model:

• provider: --default-provider, otherwise the automatic available-provider preference;
• model: --model, then --default-model, then the selected provider's built-in default.

--model is a higher-priority deployment override. --default-model is the validated default for the configured default provider. An invalid deployment default fails startup rather than silently selecting a different provider or model.

mecatui accepts these flags for its embedded server. They do not reconfigure a server used through mecatui connect. mecak8s exposes the corresponding server configuration. See the operator provider and model reference for credential sources and deployment options.

Operator-defined gateways

An operator can declare a named HTTPS gateway in the user-global settings.yaml under providers: and make it the deployment default with models.default_provider. API-key gateways use the matching provider ID in the operator-local auth.yaml; credentials are never read from a project file or supplied by mecatui connect. The server snapshots these settings and credentials once while it starts, so changing either file requires a restart. When a custom provider's live model listing is unreachable, unauthorized, or empty, /models keeps its configured default model selectable and displays only a safe provider status; endpoints, credentials, and raw listing errors or response bodies are never published to clients. Built-in --*-base-url flags still take precedence over eligible built-in endpoint overrides. See the provider configuration reference for the accepted flavors and fields.

Select OpenRouter models

One OpenRouter API key registers two protocol-specific providers, the same split the ToolHive gateway uses:

Provider ID	Inference
openrouter	POST /v1/responses
openrouter-anthropic	POST /v1/messages

Both providers cache. Mecatl asks for a prompt cache on every request, using the Responses protocol's own prompt_cache_breakpoint, so a model that caches only when asked is covered on any endpoint. That includes the ToolHive gateway and any OpenAI-compatible endpoint you configure yourself.

Select Anthropic models under openrouter-anthropic when you want more than the floor. That endpoint speaks the Anthropic Messages protocol, which carries four cache breakpoints instead of one and lets you set a cache lifetime with --anthropic-cache-ttl (5m or 1h). The Responses protocol expresses neither.

This matters for cost. Anthropic caches a prompt only when the caller asks, and cache reads bill at a tenth of uncached input, so a long session on an unasked path pays the full price every turn. Mecatl therefore prefers openrouter-anthropic when the default model is an Anthropic model. An explicit --default-provider or an operator models.default_provider still wins.

openrouter-anthropic lists Anthropic models only, because OpenRouter's Anthropic endpoint does not serve other vendors' models.

In /models, a row marked no-cache is a Claude model Mecatl will not ask to cache, which is where an unasked cache costs the most. You will see it if you run with --no-prompt-cache, or if you route a Claude model through a provider speaking the Chat Completions protocol: opencode, or one you defined with api_flavor: openai-chat-completions. That protocol has no way to ask for a cache, so every turn re-pays full input. Route Claude models through a Responses or Messages provider instead.

To stop asking for a prompt cache, run with --no-prompt-cache. It turns off every Responses-side ask and the three conversation breakpoints on the Anthropic Messages providers.

It does not turn off caching completely. On a Messages provider, the breakpoint covering the system prompt is emitted whatever you set, so the provider is still asked to retain that prefix for the cache lifetime. A deployment relying on a zero-retention arrangement therefore needs --no-prompt-cache and a model route that avoids the Messages providers: anthropic, openrouter-anthropic, toolhive-anthropic, and any provider you defined with api_flavor: anthropic-messages.

Configure aliases, slots, and task routing

For a deployment with several kinds of work, use the operator-global settings.yaml to give models stable aliases and assign them to internal jobs or delegation categories:

models:
  default: gpt-5.6-terra
  aliases:
    planner: gpt-5.6-sol
    heavy: gpt-5.6-terra
    coder: gpt-5.6-luna
    quick: gemini-3.5-flash
    image: gpt-5.6-terra
  slots:
    compaction: heavy
    ask-reviewer: quick
    guardrail: coder
    plan: planner
    router: coder
    title: quick
  router:
    backend: jev
    jev:
      minimum-confidence: 0.5
      maximum-input-bytes: 16384
    default-category: medium
    categories:
      - name: large
        description: Deep reasoning, architecture, and subtle concurrency bugs.
        model: heavy
      - name: medium
        description:
          Multi-file implementation, integration, and substantial tests.
        model: coder
      - name: small
        description: Focused edits, known fixes, and quick lookups.
        model: quick
      - name: image
        description: Work requiring visual input.
        model: image

The guardrail slot also accepts a strict provider-aware object when the checker must use a different configured provider:

models:
  slots:
    guardrail:
      provider: review-provider
      model: coder

Both object fields are required. Unknown providers, missing fields, unknown keys, and an unresolvable model fail startup; Mecatl never infers a provider from an opaque model ID. The scalar form binds its selector to the deployment default provider. Project-tier objects are ignored with a warning.

Use a planning model in plan mode

Set models.default to the model that implements your changes and bind plan to a model for plan-mode turns:

models:
  aliases:
    implementation: gpt-5.6-terra
    planner: gpt-5.6-sol
  default: implementation
  slots:
    plan: planner

When the session enters plan mode, its next turn uses the plan model. When you approve the plan and Mecatl continues in default or accept-edits mode, it uses the session default again. Mecatl rebuilds the session engine at the mode change, so the provider stays fixed. Configure both model IDs for the same provider.

If plan is unset, Mecatl uses the reasoning slot when it is configured; otherwise plan mode uses the session default model. See Permissions and posture for the plan-review workflow.

This example selects the Jev classifier backend. Set TYPESAFE_API_KEY in the server process environment. When Jev is active, Mecatl sends each eligible delegated task description and the configured category names and descriptions to Typesafe. Keep the default backend: llm if delegated task text must remain inside your configured LLM path.

Jev uses model jev-1.13.0 by default. minimum-confidence: 0 accepts every valid choice; a higher value from 0 through 1 makes a lower-confidence choice fall back to the inherited model. The optional base-url must use HTTPS, except for loopback HTTP development endpoints. Jev routing accepts up to 255 categories. maximum-input-bytes defaults to 16384 and accepts an integer from 1 through 65536. Mecatl measures the complete rendered request text against this limit and never permits more than 64 KiB. It does not truncate an over-limit task, instructions, or taxonomy. Router observability uses the same backend-neutral outcomes for both classifiers: classifier-error, bad-verdict, unknown-category, low-confidence, input-over-limit, capacity-timeout, cancelled, and timeout. In particular, a deadline is reported as timeout; cancelled means caller cancellation. All remain fail-soft and count toward the existing three-miss per-run breaker. default-category: medium is an advisory instruction for either backend when no category clearly fits. It does not force a fallback category. A named model on a delegation or agent definition is deterministic and takes precedence over category routing; categories are advisory classification for otherwise unpinned work.

Diagnose a delegated model decision

Start with the model line on the live delegation card. A successful decision keeps the compact routed: <category> → <model> form. A fallback names the model that actually ran and can add the rejected candidate and its confidence comparison:

model: gpt-6-astra · fallback: low-confidence
candidate: medium → gpt-5.6-terra · confidence 0.42 < threshold 0.50

The candidate is evidence about the classifier result. It is not the model that ran. The model: value remains the actual model after a fallback.

Press F6 and focus the child, Parallel branch, or team member for the complete decision. The detail identifies the configured backend and classifier, candidate, actual model, final reason, threshold, miss count, and breaker state. A zero Jev threshold appears as disabled. LLM routing has no native confidence score, so its detail shows confidence as unavailable instead of 0.00. Pin, fork, resume, and breaker skips show the actual model and why the classifier was not called.

The live view explains the current run. To inspect retained evidence after the run, start a target-bound debugger with mecatui debug <SESSION_ID> or mecatui connect <ADDRESS> debug <SESSION_ID>, then ask it to use InspectSession with the delegation view. The debugger reads the stored Subagent, Parallel, and Team start events. If an older or incompletely persisted event has no routing decision, the evidence remains absent. Do not infer that a classifier ran from the displayed model or classifier configuration.

Effective configuration tells you which backend, classifier, taxonomy, threshold, and category mappings the server can use. Runtime evidence tells you what happened for one delegation. Jev confidence is a backend-native score, not measured accuracy. Calibrate a nonzero threshold against a representative labeled workload from your own tasks. Mecatl does not provide a router evaluation command or a universal recommended threshold.

These mechanisms are independent:

• Aliases map readable names to concrete provider-specific model IDs.
• Slots select models for internal calls. compaction, ask-reviewer, guardrail, plan, and router do not replace the session model. The plan slot can use a stronger model while a plan is being written; compaction and checker slots can use cheaper models.
• title is an explicit opt-in slot for automatic session-title generation. It has no fallback at all: if the binding is absent, or if it is present but cannot be resolved for the session's fixed provider, generation is disabled and the server makes no title-provider call. This differs from other invalid slot or route targets, which may warn and fall back to the session model. With a compatible title binding, the server generates a title asynchronously from up to three early genuine prompts; it never delays or changes the chat. Its token usage is stored separately as session_title, not charged to the chat's displayed usage or run budget.
• Router categories select a model for a plain delegated Subagent, an unpinned named specialist (including mode: "read-write"), a Parallel branch, or an undefined team member from the task description. A taxonomy enables the router; with no taxonomy, delegation keeps its inherited/default model.

With the example above, routing resolves as:

large  → heavy  → gpt-5.6-terra
medium → coder  → gpt-5.6-luna
small  → quick  → gemini-3.5-flash
image  → image  → gpt-5.6-terra

Resolution is fail-soft for slots and routes other than title: an invalid alias, slot, or route target warns and falls back to the session model. The title slot is the exception described above; an absent or unresolvable title binding disables generation rather than falling back or making a provider call. Explicit per-call models, model-pinned named agents, fork or resume choices, and other higher- precedence selectors are not overridden by the router. A named definition with no model: is routable; model: inherit is an explicit pin. Writable named routing keeps the specialist's direct-write scope, while explicit read-write+agent+model remains invalid. Model slots and router taxonomies are operator decisions; project model settings are ignored unless the operator explicitly allows the relevant model set on a trusted project via models.allowlist.

An allowlisted model is not scoped to a particular use: a trusted project can bind any allowlisted model to any slot, including the guardrail and ask-reviewer safety checkers, not just the session default. Do not allowlist a model you would be unwilling to see used as a safety checker.

This configuration belongs in the operator-global settings file, not a checked-in project file. See the configuration reference for the complete field schema and defaults.

Route OpenRouter models through preferred downstreams

OpenRouter can serve one model through several downstream inference providers. By default, it balances among them by price. An operator can instead set a preferred order for each model in the operator-tier settings.yaml:

openrouter:
  models:
    'anthropic/claude-sonnet-4-6':
      order: ['anthropic', 'google-vertex']
      allow_fallbacks: false
    'openai/gpt-5':
      order: ['deepinfra/turbo']

order accepts lowercase-kebab downstream slugs and disables OpenRouter's default price balancing. An absent allow_fallbacks keeps OpenRouter's default (true), so it may try other downstreams after exhausting the list. Setting it to false pins the request to the listed downstreams and can fail the turn when none are available.

This configuration is operator-tier only because it controls spend, compliance, and capabilities. Mecatl ignores a project-tier openrouter block with a warning. Invalid slugs and empty orders are also dropped with a warning.

For each OpenRouter turn, Mecatl reports the selected downstream as a provider.route event when OpenRouter supplies that metadata. The value may be absent on a cache hit. It is OpenRouter's display name, such as Google, not the configuration slug such as google-vertex, so treat it as human-readable status rather than a round-trippable identifier.

See the configuration reference for the full field schema.

Run one shot with mecatequi

mecatequi creates a new session for one prompt. It accepts model/provider defaults and reasoning-effort settings, but has no interactive model picker. Use it when the caller already knows the deployment and model configuration.

Configure reasoning effort

The accepted reasoning-effort values are:

auto, low, medium, high, xhigh, max

reasoning-effort may be set as a server default or supplied per session. The important distinction is:

• omitted effort uses the server's configured default, or the provider default when no server default exists;
• explicit auto requests the provider's default effort; and
• a valid non-empty per-session value overrides the server default.

An invalid server value is ignored with a warning and becomes unset. An invalid per-session value is ignored with a warning and falls back to the server default. The server may normalize, clamp, or drop a value according to the selected provider and known model capabilities.

The effective result is returned in resolved_model.reasoning_effort, so clients can display what the server actually applied. Provider-specific effort mapping belongs in the configuration reference, not in the selection workflow.

API journey

API clients can either omit provider/model fields and use the server defaults, or name both fields when creating a session. The fields are available through gRPC CreateSession and HTTP POST /v1/sessions.

HTTP example

The HTTP/SSE API accepts JSON. For example:

curl -sS -X POST http://127.0.0.1:8081/v1/sessions \
  -H 'Content-Type: application/json' \
  -d '{
    "workspace": "/absolute/repo",
    "provider_id": "openai",
    "model_id": "gpt-5.6-terra"
  }'

The request requires a usable workspace and an available provider. The model ID is passed to the selected provider; it does not need to appear in the server's curated inventory to be accepted. A provider that is unknown or unavailable is rejected.

Selector rules

provider_id	model_id	Result
omitted	omitted	Use the server-resolved provider and model.
set	omitted	Use that provider's own default model. The server's --default-model does not carry across to a different explicitly selected provider.
set	set	Use that provider and pass the model ID through to it. An uncatalogued model may be accepted and fail later at the provider.
omitted	set	Reject the request: a bare model ID is ambiguous.
unknown or unavailable	any	Reject the request; do not silently fall back to another provider.

A bare model_id returns HTTP 400 or gRPC InvalidArgument. The same applies to an unknown or unavailable provider. The API returns the new session ID and resolved model information after successful creation.

See Drive via gRPC / HTTP for the shared session lifecycle and the HTTP/SSE API reference for endpoint details.

Model inventory and capabilities

ListModels and mecatui's /models inventory expose public metadata, including:

• provider ID and opaque model ID;
• display name when available;
• image-input support;
• reasoning support; and
• context limit when known.

They do not expose API keys or provider-private credentials. The inventory is server-specific and depends on the providers and credentials configured at startup. Each ListModels request, including client startup and SDK requests, can refresh available providers after a ten-second per-provider cooldown. Concurrent requests share a fetch; one provider's cooldown does not prevent another from refreshing. The request waits up to ten seconds and returns models and safe provider statuses from one snapshot. There is no periodic refresh or metadata cache on disk.

The server retains each provider's last non-empty model metadata for its lifetime, even if a later listing fails, is unauthorized, or returns no models. The latest status still reports that outcome where provider status is exposed. A later non-empty listing replaces the retained list. Retained metadata can become stale; it does not guarantee current model access or context limits.

Models whose catalog includes it can call the read-only DiscoverModels tool to inspect this same resolved inventory. Start without provider_id when the provider is unknown. The first unfiltered result includes every selectable provider ID and its model count, plus the first bounded model page. You can then search across all providers or add an exact provider filter. A query is a set of case-lowered literal terms; every term must occur in the provider ID, model ID, or display name of a result. Punctuation has no special query syntax.

Each result contains the exact provider_id plus model_id selection handle and the same safe metadata as ListModels; equal model IDs under different providers remain separate. Output defaults to 20 entries and is capped at 50 entries and 32 KiB. When next_cursor is present, call the tool again with only that value as cursor. A changed inventory invalidates the cursor, so restart without it. The tool does not probe or refresh providers, accept endpoints or credentials, select or route a model, or change the current session. It remains available in no-filesystem sessions.

For a known model, the session's effective capabilities combine the model's metadata with the selected adapter's transport capabilities. For an uncatalogued model ID accepted through an explicit provider, the server can report only what the adapter itself knows. Treat the capabilities returned for the created session as authoritative.

Limitations

• Provider credentials and model availability belong to the server host. A connected mecatui cannot use credentials configured only on the TUI host.
• Model IDs are provider- and deployment-specific opaque strings.
• Listing a model does not guarantee that a later provider request will succeed.
• In mecatui, changing the provider or base model creates a peer session. The client adopts the peer's complete authoritative transcript before making it interactive, then closes the source best-effort; if creation or transcript hydration fails, the open source chat remains available. API clients must implement equivalent history carryover themselves when they create a new session.
• Provider/model selection flags configure an embedded or server deployment; they do not override a remote server reached with connect.

Next steps

• Use mecatui for the interactive model and effort pickers.
• Start and resume sessions for session creation and continuation.
• Context windows for context limits and fallback.
• Capability and deployment matrix for deployment availability.

Troubleshooting

Model context metadata is unavailable

You can send the first prompt in a new or resumed session without opening /models first. If the selected model's context window is unknown and its provider supports discovery, the server starts or joins discovery for that provider before executing the prompt. Native authenticated providers perform this listing on demand rather than at startup. Known configured, retained live, or catalog windows need no listing.

When discovery fails or returns an empty list and no positive window is known, the server rejects execution with context_window_unavailable. Your existing conversation remains available, and the rejected prompt has not been recorded as a server turn. Restore provider discovery, or ask the server operator to configure the model's verified exact context window. Retry after the ten-second cooldown; another failed attempt requires another explicit request. Cancelling your wait leaves the server's bounded discovery attempt running.

In mecatui, Retry sends the identical prepared text and attachments without rereading files or the clipboard. Back restores the editable draft, including staged pastes and images, and asks before replacing a newer draft. Cancelling that confirmation keeps both drafts. Discard submission releases the rejected payload. Recovery holds one submission in client memory under the existing size limits; accepting the prompt, changing sessions, exiting, or an unrelated terminal error releases it. There is no automatic replay or recovery after client restart.

API clients receive gRPC Unavailable with ErrorInfo domain mecatl.stacklok.com and reason context_window_unavailable, or HTTP 503. Match the structured reason, not error-message text, and retry explicitly after addressing discovery. For daemon configuration, see context discovery recovery.