Deployment config (mira.yaml)

mira.yaml is where you configure Mira. One file, three sections (llm, filter, review), and every key is optional. Whatever you set here applies to every PR Mira reviews, unless a repo overrides it in .mira.yaml or the dashboard.

This page is the canonical reference for those three sections: every key, default, and what it does.

Loading

# Local CLI
mira serve --config /etc/mira/mira.yaml

# Docker — mount the file and point the env var at it
docker run -p 8000:8000 \
  -v /etc/mira/mira.yaml:/app/mira.yaml:ro \
  -e MIRA_CONFIG=/app/mira.yaml \
  -e MIRA_GITHUB_APP_ID=123456 \
  -e MIRA_GITHUB_PRIVATE_KEY="$(cat private-key.pem)" \
  -e MIRA_WEBHOOK_SECRET=… \
  -e OPENROUTER_API_KEY=sk-or-… \
  ghcr.io/miracodeai/mira:latest

Full example

A complete mira.yaml showing every key at its default. Copy this, delete anything you're happy with the default for, and you have your config.

# mira.yaml — deployment-wide defaults. Every key is optional.
llm:
  provider: "openai" # openai | bedrock | codex-cli
  model: "anthropic/claude-sonnet-4-6"
  fallback_model: "anthropic/claude-haiku-4-5"
  indexing_model: "anthropic/claude-haiku-4-5"
  review_model: "anthropic/claude-sonnet-4-6"
  security_model: "anthropic/claude-sonnet-4-6"
  review_reasoning_effort: "off" # off | low | medium | high | xhigh | max
  temperature: 0.2
  max_tokens: 4096
  max_context_tokens: 120000
  # Point at any OpenAI-compatible endpoint — vLLM, Ollama, LiteLLM proxy,
  # LocalAI, llama.cpp, Together, Fireworks, Groq, etc.
  # Defaults to OpenRouter.
  api_style: "chat" # chat | responses
  base_url: "https://openrouter.ai/api/v1"
  api_key_env: "OPENROUTER_API_KEY"
  region: "us-east-1"
  aws_profile: null
  max_retries: 3
  request_timeout: 120
  retry_min_wait: 2
  retry_max_wait: 30
  codex_command: "codex"
  codex_home: null
  codex_sandbox: "read-only"
  codex_timeout_seconds: 900

filter:
  confidence_threshold: 0.7
  category_confidence_thresholds: {}
  max_comments: 5
  min_severity: "nitpick"
  exclude_deleted: true
  max_files: 50
  exclude_patterns:
    - "*.lock"
    - "package-lock.json"
    - "*.min.js"
    - "*.svg"
    - "*.png"
  # Only auto-review PRs from these authors (empty = review all).
  allowed_authors: []
  # Never auto-review PRs from these authors; wins over allowed_authors.
  # "dependabot" also matches "dependabot[bot]".
  blocked_authors: []

review:
  context_lines: 3
  max_diff_size: 250000
  max_file_size: 50000
  max_chunks_per_review: 5
  include_summary: true
  focus_only_on_problems: false
  walkthrough: true
  walkthrough_sequence_diagram: true
  pr_summary: "disable" # disable | append | replace
  code_context: true
  context_token_budget: 8000
  max_concurrent_chunks: 5
  self_critique: true
  security_pass: true
  security_agentic: true
  osv_scan: true
  secrets_scan: true
  agentic_tools: true
  jit_java_go: true
  blast_radius: true
  ensemble_runs: 1
  ensemble_temperature: 0.3
  auto_resolve_conversations: true
  review_on_synchronize: true
  dependency_overlap: true
  overlap:
    enabled: true
    max_candidates: 20
    confidence_floor: 0.6
    title_similarity_threshold: 0.4

index:
  max_file_size: 1048576

llm

Controls which model is used and how it's invoked.

Custom endpoints

Any compatible Chat Completions API works. Endpoints that implement the OpenAI Responses API can instead set llm.api_style: responses. Point llm.base_url at the right URL and set llm.api_key_env to the env var holding the key. Mira looks up a provider profile by base_url to apply any per-provider quirks (attribution headers, whether the model id keeps its vendor/ prefix, reasoning remapping); an endpoint with no profile gets the portable default — bare model name, no extra headers — so most providers need nothing more than base_url + api_key_env.

Ollama (local, no auth):

llm:
  base_url: "http://localhost:11434/v1"
  api_key_env: ""
  model: "llama3.1:70b"

vLLM (your own GPU server):

llm:
  base_url: "http://vllm.internal:8000/v1"
  api_key_env: "VLLM_API_KEY"
  model: "meta-llama/Meta-Llama-3.1-70B-Instruct"

LiteLLM proxy (route to many providers under one URL):

llm:
  base_url: "http://litellm:4000/v1"
  api_key_env: "LITELLM_MASTER_KEY"
  model: "claude-sonnet-4-6"

Together / Fireworks / Groq / Cerebras / Anyscale / LocalAI use the same shape; set base_url to the provider's /v1 endpoint and api_key_env to the env var holding your key.

Adding a provider profile

Profiles only matter when a provider needs a quirk the default doesn't cover (ranking headers, keeping the vendor/model prefix, remapping a reasoning level). The bundled set lives in src/mira/llm/providers.json; to add or override one at runtime without forking, point MIRA_PROVIDERS_JSON_PATH at your own file — the same override idiom as MIRA_MODELS_JSON_PATH for the model registry. A profile is matched to a request by base_url:

{
  "acme": {
    "base_url": "https://llm.acme.ai/v1",
    "api_key_env": "ACME_API_KEY",
    "model_prefix": "keep",
    "extra_headers": { "X-Acme-Project": "mira" },
    "reasoning_effort_map": { "max": "xhigh" }
  }
}

Reviews require the model to support tool/function calling. Most hosted providers support it; for self-hosted models pick one with structured-output support (Llama 3.1, Mixtral 8x22B, Qwen 2.5, DeepSeek V3, etc.).

filter

Controls which comments get posted. Mira filters aggressively by default; the goal is "every comment is one you'd want to land."

review

Controls how Mira reviews: diff sizing, walkthrough behaviour, parallelism.

review.overlap

Cross-PR overlap detection — while reviewing a PR, Mira compares it against the repo's other open PRs and flags ones stepping on it in the walkthrough: same files (merge-conflict risk) or same goal (duplicate effort). A cheap deterministic pre-filter runs first, so only genuinely-overlapping candidates cost an LLM call; stacked branches and draft PRs are excluded.

index

Controls the codebase indexing pass that powers cross-file review context. Indexing also honours filter.exclude_patterns (see above).

How merging works

For each PR, Mira merges layers in this order (later wins):

1. mira.yaml, this file.
2. .mira.yaml in the repo, or the repo's dashboard settings.
3. Env-var fallbacks (MIRA_MODEL, DATABASE_URL) if nothing above set the value.

Merging is per-key. A repo that sets filter.confidence_threshold: 0.5 inherits everything else from mira.yaml. Most repos don't need a .mira.yaml at all.

Validation

mira serve --config … validates the YAML at startup. A typo or a wrong type (e.g. confidence_threshold: "high") fails the boot rather than silently no-op'ing on the first PR. The error names the offending field.

Where to next