codecraft/config.example.toml

# =============================================================================
#  CodeCraft – configuration
# =============================================================================
#
#  This file is the single source of truth for the server. Everything here can
#  also be edited in the web UI (Settings). The UI edits this file in place and
#  keeps your comments and formatting intact.
#
#  Location: $CODECRAFT_CONFIG, otherwise <data_dir>/config.toml (Docker: /data/config.toml)
#
#  Secrets: any string value may reference environment variables with ${NAME},
#  e.g. api_key = "${ANTHROPIC_API_KEY}". Prefer this over literal keys – keys
#  entered in the UI are stored here in plain text.
# =============================================================================

[server]
# Address the HTTP server listens on.
bind = "0.0.0.0:8787"

# Root directory the agents may work in. Sessions pick a project folder below it.
# Docker: mount your projects here (see docker-compose.yml).
workspace = "/workspace"

# Where sessions (SQLite) and runtime state are stored.
data_dir = "/data"

# Default model for new sessions, as "<provider>/<model>".
default_model = "anthropic/claude-sonnet-5-5"

# Cheap/fast model for titles and history compaction. Falls back to default_model.
small_model = "anthropic/claude-haiku-5-5"

# Compact the conversation automatically when the context is this full (0.0–1.0).
auto_compact_threshold = 0.85

# Max. subagents running at the same time (all sessions). Additionally, each
# subagent can be limited with max_parallel, and each provider/model with
# max_concurrency (see below).
max_parallel_subagents = 4

# MCP servers with more tools than this are loaded on demand (see [mcp.*] load).
mcp_direct_limit = 15

# Currency label for the cost statistics (prices below are in this currency).
currency = "USD"


# -----------------------------------------------------------------------------
#  Providers
# -----------------------------------------------------------------------------
#  kind = "anthropic"  → Anthropic Messages API (and compatible endpoints)
#  kind = "openai"     → OpenAI-compatible /chat/completions (Gemini, opencode Go,
#                        fal, OpenRouter, Ollama, vLLM, llama.cpp, LM Studio, ...)
#
#  models: optional list. If empty, the model list is fetched from <base_url>/models.
#  Per-model options: context (tokens), max_output (tokens), reasoning (bool),
#  vision (bool, can read images), effort ("low" | "medium" | "high" | "xhigh" | "max").
#
#  Prices (per 1M tokens, used for the statistics):
#    input_price, output_price, cache_read_price, cache_write_price
#  Cache prices default to input_price when omitted. Costs are stored per request,
#  so changing a price only affects future requests.
#
#  Concurrency: max_concurrency limits simultaneous requests – on the provider
#  (= the host) and/or per model. Requests beyond the limit wait in a queue,
#  e.g. for self-hosted models that can only process one request at a time.
#
#  options: extra request body fields sent as-is, e.g.
#    options = { temperature = 0.5, reasoning_budget_tokens = 5000 }
#
#  Authentication: by default "Authorization: Bearer <api_key>" (openai) or
#  "x-api-key: <api_key>" (anthropic). Override with auth_header / auth_prefix.
#  Extra static headers: headers = { "X-Foo" = "bar" }
# -----------------------------------------------------------------------------

[providers.anthropic]
kind = "anthropic"
base_url = "https://api.anthropic.com/v1"
api_key = "${ANTHROPIC_API_KEY}"
# Current Claude models think adaptively; "effort" controls depth and cost.
# Prices: Anthropic list prices in USD per 1M tokens (Nov 2026) – adjust if needed.
models = [
  { id = "claude-opus-5-5", context = 1000000, max_output = 64000, reasoning = true, vision = true, effort = "high",
    input_price = 4.0, output_price = 20.0, cache_read_price = 0.20, cache_write_price = 5.0 },
  { id = "claude-sonnet-5-5", context = 1000000, max_output = 64000, reasoning = true, vision = true, effort = "medium",
    input_price = 2.0, output_price = 10.0, cache_read_price = 0.20, cache_write_price = 2.5 },
  { id = "claude-haiku-5-5", context = 1000000, max_output = 32000, reasoning = true, vision = true, effort = "low",
    input_price = 0.10, output_price = 0.50, cache_read_price = 0.01, cache_write_price = 0.125 },
]

# Google Gemini through its OpenAI-compatible endpoint.
[providers.gemini]
kind = "openai"
base_url = "https://generativelanguage.googleapis.com/v1beta/openai"
api_key = "${GEMINI_API_KEY}"
models = []          # discovered automatically
enabled = false      # set to true once GEMINI_API_KEY is available

# opencode Go subscription (OpenAI-compatible endpoint for most models).
# Current model list: https://opencode.ai/inference/go/v1/models
[providers.opencode-go]
kind = "openai"
base_url = "https://opencode.ai/inference/go/openai/v1"
api_key = "${OPENCODE_API_KEY}"
models = []
enabled = false

# fal.ai – OpenRouter-backed LLM router. fal expects "Authorization: Key <FAL_KEY>".
# NOTE: verify base_url against the fal docs for your account.
[providers.fal]
kind = "openai"
base_url = "https://fal.run/openrouter/router/openai/v1"
api_key = "${FAL_KEY}"
auth_prefix = "Key "
models = [
  { id = "google/gemini-2.5-flash", context = 1000000, max_output = 32000, vision = true },
]
enabled = false

# Self-hosted (Ollama shown; vLLM/llama.cpp/LM Studio work the same way).
# From inside Docker, the host is reachable as host.docker.internal.
[providers.local]
kind = "openai"
base_url = "http://host.docker.internal:11434/v1"
api_key = ""
max_concurrency = 1  # one request at a time on this host
models = []
enabled = false


# -----------------------------------------------------------------------------
#  MCP servers
# -----------------------------------------------------------------------------
#  type = "stdio" → command + args (+ env), started as a child process
#  type = "http"  → Streamable HTTP endpoint (url + optional headers)
#  Tools appear to the model as  mcp__<server>__<tool>.
#  users = [...]: who may use the server. Default: admins only.
#  Use ["*"] for all users or list usernames, e.g. ["anna", "ben"].
#  load = "auto" (default) | "always" | "on_demand": servers with many tools
#  (more than server.mcp_direct_limit) are not sent with every request; the
#  model finds them with mcp_search and calls them with mcp_call. This keeps
#  requests small – important for self-hosted models.
#  MCP servers may ask you questions (elicitation); they show up in the UI.
# -----------------------------------------------------------------------------

# [mcp.trilium]
# type = "http"
# url = "https://notes.example.org/mcp"
# headers = { Authorization = "Bearer ${TRILIUM_TOKEN}" }

# [mcp.filesystem]
# type = "stdio"
# command = "npx"
# args = ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"]
# env = {}
# enabled = false


# -----------------------------------------------------------------------------
#  Permissions
# -----------------------------------------------------------------------------
#  For each tool: "allow" | "ask" | "deny". "ask" forwards the request to the UI.
#  MCP tools can be addressed as "mcp__<server>__<tool>" or "mcp__<server>__*".
#  Session permission modes in the UI override this:
#    default       → these rules
#    accept_edits  → file edits allowed without asking
#    bypass        → everything allowed (use with care)
# -----------------------------------------------------------------------------

[permissions]
read = "allow"
glob = "allow"
grep = "allow"
list = "allow"
todo_write = "allow"
ask_user = "allow"
task = "allow"
write = "ask"
edit = "ask"
bash = "ask"
bash_output = "allow"
kill_shell = "allow"
web_fetch = "ask"
web_search = "allow"
diagnostics = "allow"
browser_snapshot = "allow"
image_compare = "allow"
generate_image = "ask"   # costs money per image
"mcp__*" = "ask"

# Shell command patterns (glob, matched against the full command). First match wins.
[permissions.bash_rules]
"git status*" = "allow"
"git diff*" = "allow"
"git log*" = "allow"
"ls*" = "allow"
"pwd" = "allow"
"rm -rf /*" = "deny"


# -----------------------------------------------------------------------------
#  Agents (modes)
# -----------------------------------------------------------------------------
#  Built-in: build (full access), plan (read-only, proposes a plan),
#            craft (study → build → check → ready), and subagents
#            general + explore used by the "task" tool.
#  Define your own or override built-ins here. mode = "primary" shows up in the
#  mode switcher, mode = "subagent" is only available through the task tool.
# -----------------------------------------------------------------------------

#  Subagents can run on a different model (and host) than the chat, and in the
#  background: the main agent calls task(background = true), keeps working and
#  receives the result as a message when the subagent is done.

# [agents.writer]
# mode = "subagent"
# description = "Writes documentation, summaries, guides and translations."
# model = "local/qwen3-4b"                    # optional, otherwise the session model
# tools = ["read", "glob", "grep"]            # optional, default: all tools
# max_parallel = 1                            # optional, instances at the same time
# prompt = '''
# You are a technical writer. Write clear, well-structured text.
# '''


# -----------------------------------------------------------------------------
#  Image generation (fal.ai)
# -----------------------------------------------------------------------------
#  Agents get the tool generate_image; the "Images" page creates images manually.
#  Images are saved in the project (output_dir) and listed in a gallery.
#  Model ids: https://fal.ai/models – price is per image (for the statistics).
#  size_param: "image_size" (FLUX-style presets, default) or "aspect_ratio" ("16:9").
# -----------------------------------------------------------------------------

[images]
enabled = false
base_url = "https://fal.run"
api_key = "${FAL_KEY}"
default_model = "fal-ai/flux/schnell"
output_dir = "generated"
models = [
  { id = "fal-ai/flux/schnell", name = "FLUX.1 schnell – fast", price = 0.003 },
  { id = "fal-ai/flux-2-pro", name = "FLUX.2 pro" },
  { id = "fal-ai/ideogram/v3", name = "Ideogram 3 – text in images", size_param = "aspect_ratio" },
  { id = "fal-ai/recraft/v4/pro/text-to-image", name = "Recraft V4 pro – design" },
]


# -----------------------------------------------------------------------------
#  Users & sign-in
# -----------------------------------------------------------------------------
#  Users are managed in the web UI (Settings → Users). On first start without
#  users, the server prints a one-time setup code to its log; use it in the UI
#  to create the first admin – or set CODECRAFT_ADMIN_USER and
#  CODECRAFT_ADMIN_PASSWORD in the environment.
#
#  Every user works in their own directory <workspace>/<username>; admins can
#  use the whole workspace. Note: file tools are confined to that directory,
#  but shell commands run in the shared container and are not sandboxed.
# -----------------------------------------------------------------------------

[auth]
# Lifetime of a login in days.
session_days = 30
# External URL of this server – required for single sign-on (OIDC redirects).
public_url = ""
# Secure cookies: "auto" (when served via https), "true" or "false".
cookie_secure = "auto"

# Single sign-on via OpenID Connect (Authentik, Authelia, Keycloak, ...).
# Redirect URI to register at the IdP: <public_url>/api/auth/oidc/callback
[auth.oidc]
enabled = false
issuer = "https://auth.example.org/application/o/codecraft/"
client_id = "codecraft"
client_secret = "${OIDC_CLIENT_SECRET}"
scopes = ["openid", "profile", "email", "groups"]
username_claim = "preferred_username"
groups_claim = "groups"
# Members of this group are admins; empty = manage roles in CodeCraft.
admin_group = ""
# Only members of this group may sign in; empty = no restriction.
allowed_group = ""
# Create users automatically on their first sign-in.
auto_create = true
# Link a sign-in to an existing local user with the same username.
# Only enable this if you trust the usernames of your IdP.
link_by_username = false
button_label = "Single sign-on"


# -----------------------------------------------------------------------------
#  Web search (web_search tool)
# -----------------------------------------------------------------------------
[search]
# "duckduckgo" (no key, may be rate-limited), "searxng" (self-hosted, set url),
# "brave" or "tavily" (api_key, e.g. "${BRAVE_API_KEY}")
provider = "duckduckgo"
url = ""
api_key = ""


# -----------------------------------------------------------------------------
#  Language servers (diagnostics after edits, tool "diagnostics")
# -----------------------------------------------------------------------------
#  Built in: rust (rust-analyzer), typescript (typescript-language-server),
#  python (pyright-langserver), go (gopls), c (clangd) – used when installed.
#  TypeScript and Python servers ship with the image. Servers run in the
#  user's sandbox and stop after 15 minutes without use.
# -----------------------------------------------------------------------------
[lsp]
enabled = true
# Add or override a server:
# [lsp.servers.zig]
# command = "zls"
# extensions = ["zig"]
# root_markers = ["build.zig"]


# -----------------------------------------------------------------------------
#  Sandbox
# -----------------------------------------------------------------------------
#  Shell commands and git of a user run in a bubblewrap sandbox: system
#  read-only, only the user's folders writable, own home, own processes, no
#  server environment. The network is shared (package managers keep working).
#  Needs the security options in docker-compose.yml. Without a working
#  sandbox, non-admin users cannot run shell commands.
[sandbox]
# "auto" = non-admin users, "always" = everybody (admins too), "off" = nobody
mode = "auto"


# -----------------------------------------------------------------------------
#  Craft mode
# -----------------------------------------------------------------------------
[craft]
# Headless Chromium binary used to study websites and check builds.
chromium = "chromium"
# Screen widths used for screenshots and comparisons.
widths = [390, 768, 1440]
# Clean room: only what a visitor sees (screenshots, visible text, layout).
# Set to false for "deep" mode (page source and scripts may be read).
clean_room = true