Built-ins

September 2, 2026 · View on GitHub

Every built-in tool, sub-agent, input, output, user command, framework command, LLM provider, and LLM preset shipped with KohakuTerrarium.

For the shape of tools vs sub-agents, read concepts/modules/tool and concepts/modules/sub-agent. For task-oriented help, see guides/creatures and guides/custom-modules.

Tools

Built-in tool classes live in src/kohakuterrarium/builtins/tools/. Register them in a creature config under tools: by bare name.

bash: Run a shell command. Uses bash by default on every platform (Git Bash on Windows). Pass type="..." to select another supported shell: bash, zsh, sh, fish, pwsh, or powershell. Captures stdout and stderr, truncated to a cap. Direct execution.

Args: command (str), working_dir (str, optional), timeout (float, optional).

python: Run a Python subprocess. Respects working_dir and timeout. Direct.

Args: code (str), working_dir, timeout.

File operations

read: Read text, image, or PDF content. Records read-state per file. Images are returned as base64 data URLs. PDF support requires pymupdf. Direct.

Args: path (str), offset (int, optional), limit (int, optional).

write: Create or overwrite a file. Creates parent directories. Blocks overwrites unless the file was read first (unless new). Direct.

Args: path, content, new (bool, optional).

edit: Auto-detects unified-diff (@@) or search/replace form. Refuses binary files. Direct.

Args: path, old_text/new_text or diff, replace_all (bool).

multi_edit: Apply an ordered list of edits to one file. Atomic per file. Modes: strict (every edit must apply), best_effort (skip failures), default (partial apply with report). Direct.

Args: path, edits: list[{old, new}], mode.

glob: mtime-sorted glob. Respects .gitignore. Early-terminates. Direct.

Args: pattern, root (optional), limit (optional).

grep: Regex search across files. Supports ignore_case. Skips binaries. Direct.

Args: pattern, path (optional), ignore_case (bool), max_matches.

tree: Directory listing with YAML-frontmatter summaries for markdown files. Direct.

Args: path, depth.

notebook_read: Read Jupyter notebooks (.ipynb) as compact markdown summaries, selected cells, or raw JSON. Direct.

Args: path, cell, include_outputs, mode.

notebook_edit: Edit Jupyter notebooks by replacing, inserting, deleting, or appending cells while preserving notebook metadata. Direct.

Args: path, cell, source, cell_type, action.

Structured data

json_read: Read a JSON document by dot-path. Direct.

Args: path, query (dot-path).

json_write: Assign a value at a dot-path. Creates nested objects as needed. Direct.

Args: path, query, value.

Web

web_fetch: Fetch a URL as markdown. Tries crawl4ai → trafilatura → Jina proxy → httpx + html2text. 100k-char cap, 30s timeout. Direct.

Args: url.

web_search: Web search with a DuckDuckGo default plus optional Codex subscription and DeepSeek Responses backends. After kt login codex, select backend: codex; this works independently of the creature's active LLM. DeepSeek requires kt config key set deepseek and backend: deepseek. Either explicit provider backend can use fallback: duckduckgo for transient failures. Direct.

Args: query, max_results (int), region (str).

Provider-native media

image_gen: Generate or edit an image through the provider's own native image backend. Currently auto-injected for Codex-backed creatures unless opted out with disable_provider_tools: [image_gen]. The executor never runs it; the provider returns structured image content and the session store persists the generated file into the session artifacts dir.

Args: prompt plus provider-specific knobs when explicitly wired: output_format, size, quality, action, background.

Interactive and memory

ask_user: Prompt the user over stdin (CLI or TUI only). Stateful.

Args: question.

think: No-op; preserves reasoning as a tool event for the event log. Direct.

Args: thought.

scratchpad: Session-scoped KV store. Shared across agents in a session.

Args: action (get | set | delete | list), key, value.

search_memory: FTS / semantic / auto search over the session's indexed events. Per-agent filter.

Args: query, mode (auto/fts/semantic/hybrid), k, agent.

Communication

send_message: Emit a message to a channel. Resolves creature- local channels first, then the environment's shared channels. Direct.

Args: channel, content, sender (optional).

Introspection

info: Load on-demand documentation for any tool or sub-agent. Delegates to skill manifests under src/kohakuterrarium/builtin_skills/ and per-agent overrides. Direct.

Args: target (tool or sub-agent name).

stop_task: Cancel a running background task or trigger by id. Direct.

Args: job_id (job id from any tool call; or the trigger id returned by add_timer/watch_channel/add_schedule).

Setup-able triggers (exposed as tools via `type: trigger`)

Each universal trigger class is wrapped as its own tool via modules/trigger/callable.py:CallableTriggerTool. A creature opts in by listing the trigger's setup_tool_name under tools: with type: trigger. The tool's description is prefixed with **Trigger**: so the LLM knows calling it installs a long-lived side-effect. All three return immediately with the installed trigger id; the trigger itself runs in the background.

add_timer (wraps TimerTrigger): Install a periodic timer.

Args: interval (seconds, required), prompt (required), immediate (bool, default false).

watch_channel (wraps ChannelTrigger): Listen on a named channel.

Args: channel_name (required), prompt (optional, supports {content}), filter_sender (optional).
The agent's own name is auto-set as ignore_sender to prevent self-triggering.

add_schedule (wraps SchedulerTrigger): Clock-aligned schedule.

Args: prompt (required); exactly one of every_minutes, daily_at (HH:MM), hourly_at (0-59).

Group tools

These mutate or inspect the graph the caller belongs to. They split into two tiers:

Basic: registered on every creature in a graph. The tool body gates per-call (e.g., group_send enforces the privileged-recipient rule for non-privileged callers):

send_channel: Broadcast a message to a channel the caller is wired to send on. Args: channel, content.
group_send: Direct point-to-point send to another creature in the caller's graph (bypasses channels). Non-privileged callers can only target privileged recipients (workers report to supervisors, not peer-workers). Args: target_creature_id, content.

Privileged: registered ONLY on privileged nodes:

group_add_node: Spawn a creature into the caller's graph. Args: config_path (file or @pkg/ref), optional name, llm, pwd. Returns the new creature's id.
group_remove_node: Stop and remove a creature from the graph. May trigger an auto-split if it was a bridge. Args: creature_id.
group_start_node / group_stop_node: Start or stop a creature without removing it. Args: creature_id.
group_channel: CRUD on channels and per-creature wiring. Args: action ∈ create | delete | wire | unwire, channel, optional creature_id and direction. Cross-graph wires route through Terrarium.connect and may auto-merge.
group_wire: CRUD on output-wiring edges. Args: action ∈ add | remove, plus an edge spec (from, to, optional with_content, prompt).
group_status: Snapshot the caller's graph: creatures (with status, privilege, parent), channels (with history pointers, connected creatures), output wires, and the available catalog to spawn from. Args: none.

Sub-agents

Shipped sub-agent configs under src/kohakuterrarium/builtins/subagents/. Reference them in a creature config under subagents: by name.

All builtin sub-agents load default_plugins: ["auto-compact"] and the unified budget plugin with minimal runtime options: turn soft/hard 40/60, tool-call soft/hard 75/100, and no walltime budget.

Name	Tools	Purpose
`worker`	`read`, `write`, `bash`, `glob`, `grep`, `edit`, `multi_edit`	Fix bugs, refactor, run validations.
`coordinator`	`send_message`, `scratchpad`	Decompose → dispatch → aggregate.
`explore`	`glob`, `grep`, `read`, `tree`, `bash`	Read-only exploration.
`plan`	`explore` tools + `think`	Read-only planning.
`research`	`web_search`, `web_fetch`, `read`, `write`, `think`, `scratchpad`	External research.
`critic`	`read`, `glob`, `grep`, `tree`, `bash`	Code review.
`response`	`read`	User-facing copy generator. Typically `output_to: external`.
`memory_read`	`tree`, `read`, `grep` over the memory folder	Recall from agent memory.
`memory_write`	`tree`, `read`, `write`	Persist findings into memory.
`summarize`	(no tools)	Condense conversation for handoff or reset.

Inputs

Shipped input modules under src/kohakuterrarium/builtins/inputs/.

cli: Stdin prompt. Options: prompt, exit_commands.

cli_nonblocking: Same surface as cli but returns control to the event loop between keystrokes (useful when triggers fire during input).

none: No input. For trigger-only agents.

Audio/ASR implementations are not built-ins. The conversational example ships opt-in ASRModule/Whisper custom input files under examples/agent-apps/conversational/custom/; load them with type: custom.

Two further input types are resolved dynamically:

tui: mounted by the Textual app when running under TUI mode.
custom / package: loaded via module + class fields.

Outputs

Shipped output modules under src/kohakuterrarium/builtins/outputs/.

stdout: Print to stdout. Options: prefix, suffix, stream_suffix, flush_on_stream.

stdout_prefixed: stdout with a per-line prefix, useful for tagging side outputs.

console_tts: Console-only TTS shim that prints the synthesized text character-by-character with a configurable char_delay. Intended for demos and testing; there is no audio backend.

dummy_tts: Silent TTS that fires the usual TTS lifecycle events without any output. Useful in tests.

Additional routed types:

tui: renders into the Textual TUI widget tree.
custom / package: loaded via module + class.

There is no plain tts registry key. Real TTS backends (Fish, Edge, OpenAI, etc.) are shipped as custom/package outputs that subclass TTSModule.

User commands

Slash commands available inside input modules. Under src/kohakuterrarium/builtins/user_commands/.

Command	Aliases	Purpose
`/help`	`/h`, `/?`	List commands.
`/status`	`/info`	Model, message count, tools, jobs, compact state.
`/clear`		Clear conversation (the session log retains history).
`/model [name]`	`/llm`	Show current model or switch profile. Accepts canonical `provider/name[@variations]`.
`/compact`		Manual context compaction.
`/regen`	`/regenerate`, `/retry`	Re-run the last assistant turn as a sibling branch.
`/edit <message_index> <new content>`	(none)	Edit a past user message and re-run from that point as a new branch.
`/branch [<turn> <branch_id>\|latest]`	`/br`	List or switch the live branch for regen/edit alternatives.
`/fork [event_id] [--name name]`	(none)	Copy the current session into a new `.kohakutr` file for alternate exploration.
`/plugin [list\|enable\|disable\|toggle] [name]`	`/plugins`	Inspect or toggle plugins.
`/skill [list\|enable\|disable\|toggle\|show] [name]`	`/skills`	Inspect or toggle procedural skills.
`/<skill-name> [args]`	(none)	User-invoke path for an enabled procedural skill when no built-in slash command shadows that name.
`/exit`	`/quit`, `/q`	Graceful exit. On web, a force flag may be required.

Framework commands

Inline directives the LLM can emit instead of a tool call. They talk to the framework directly (no tool round-trip). Defined under src/kohakuterrarium/commands/.

Framework commands use the same syntax family as tool calls: they follow the creature's configured tool_format (bracket / XML / native). The default bracket form with bare-identifier placeholders:

[/info]tool_or_subagent[info/]: Load a tool's, sub-agent's, or procedural skill's documentation on demand.
[/read_job]job_id[read_job/]: Read output from a background job. Supports --lines N and --offset M in the body.
[/jobs][jobs/]: List running jobs with IDs.
[/wait]job_id[wait/]: Block the current turn until a background job finishes.
[/skill]skill_name [args][skill/]: Return a procedural skill body to the model for explicit invocation.

Command names share a namespace with tool names; the command for reading job output is called read_job to avoid colliding with the read file-reader tool. Defined under src/kohakuterrarium/commands/.

LLM providers

Built-in provider types (backends):

Provider	Backend type	Transport	Notes
`codex`	`codex`	OpenAI Responses API	Routed via `CodexOAuthProvider`. No `base_url` → ChatGPT-subscription OAuth (`kt login codex`); with a `base_url` → API-key auth against a custom OpenAI-Responses endpoint (no OAuth). Ships provider-native tools such as `image_gen`.
`openai`	`openai`	OpenAI `/chat/completions`	API-key auth (`OPENAI_API_KEY`).
`openrouter`	`openai`	OpenAI-compat against OpenRouter	API-key auth (`OPENROUTER_API_KEY`); unified `reasoning` param.
`anthropic`	`anthropic`	Anthropic-compatible Messages API	API-key auth (`ANTHROPIC_API_KEY`). Uses the official `anthropic` SDK. Claude-specific knobs go through `extra_body` (`thinking.`, `output_config.`). Prompt-caching markers are auto-applied unless disabled.
`gemini`	`openai`	Google's OpenAI-compat endpoint	API-key auth (`GEMINI_API_KEY`).
`mimo`	`openai`	Xiaomi MiMo	`kt login mimo`.
`kimi-code`	`anthropic`	Kimi Code Anthropic-compatible endpoint	API-key auth (`KIMI_CODE_API_KEY`).
`glm-coding`	`anthropic`	GLM Coding Plan Anthropic-compatible endpoint	API-key auth (`GLM_CODING_API_KEY`) with Bearer-token wiring.

Canonical backend types are openai, anthropic, and codex. Legacy codex-oauth backend type values are silently migrated on read (see configuration reference).

LLM presets

Shipped in src/kohakuterrarium/llm/presets.py. Use them as llm: or --llm values.

Naming convention (post-2026-04 refactor):

Direct / native-API variants are the primary name (claude-opus-4.7, gemini-3.1-pro, mimo-v2-pro).
OpenRouter-routed variants use the -or suffix (claude-opus-4.7-or).
OpenAI is an exception: gpt-5.4 stays bound to the Codex OAuth provider; the direct OpenAI API variant uses -api, OpenRouter uses -or.
Legacy names (claude-opus-4.6-direct, or-gpt-5.4, gemini-3.1-pro-direct, mimo-v2-pro-direct, …) survive as aliases so existing configs keep working.

OpenAI via Codex OAuth

gpt-5.3-codex-spark (ChatGPT Pro research preview; text-only)
gpt-5.5
gpt-5.4 (aliases: gpt5, gpt54)
gpt-5.3-codex (gpt53)
gpt-5.1
gpt-4o-codex (aliases: gpt4o, gpt-4o)
gpt-4o-mini-codex (alias: gpt-4o-mini)

OpenAI Direct API (`-api` suffix)

gpt-5.4-api (legacy alias: gpt-5.4-direct)
gpt-5.4-mini-api (gpt-5.4-mini-direct)
gpt-5.4-nano-api (gpt-5.4-nano-direct)
gpt-5.3-codex-api (gpt-5.3-codex-direct)
gpt-5.1-api (gpt-5.1-direct)
gpt-4o-api (gpt-4o-direct)
gpt-4o-mini-api (gpt-4o-mini-direct)

OpenAI via OpenRouter (`-or` suffix)

gpt-5.4-or (legacy alias: or-gpt-5.4)
gpt-5.4-mini-or (or-gpt-5.4-mini)
gpt-5.4-nano-or (or-gpt-5.4-nano)
gpt-5.3-codex-or (or-gpt-5.3-codex)
gpt-5.1-or (or-gpt-5.1)
gpt-4o-or (or-gpt-4o)
gpt-4o-mini-or (or-gpt-4o-mini)

Anthropic Claude Direct (no suffix: primary)

Routed through the native Anthropic-compatible Messages API. Effort via extra_body.output_config.effort.

claude-opus-4.7 (aliases: claude-opus, opus)
claude-opus-4.6 (legacy alias: claude-opus-4.6-direct)
claude-sonnet-4.6 (aliases: claude, claude-sonnet, sonnet; legacy: claude-sonnet-4.6-direct)
claude-haiku-4.5 (aliases: claude-haiku, haiku; legacy: claude-haiku-4.5-direct)

Anthropic Claude via OpenRouter (`-or` suffix)

claude-opus-4.7-or
claude-opus-4.6-or
claude-sonnet-4.6-or
claude-sonnet-4.5-or
claude-haiku-4.5-or
claude-sonnet-4-or (legacy alias: claude-sonnet-4)
claude-opus-4-or (legacy alias: claude-opus-4)

Google Gemini Direct (OpenAI-compat)

gemini-3.1-pro (aliases: gemini, gemini-pro; legacy: gemini-3.1-pro-direct)
gemini-3-flash (gemini-flash; legacy: gemini-3-flash-direct)
gemini-3.1-flash-lite (gemini-lite; legacy: gemini-3.1-flash-lite-direct)

Google Gemini via OpenRouter (`-or` suffix)

gemini-3.1-pro-or
gemini-3-flash-or
gemini-3.1-flash-lite-or
nano-banana (image-generation model, OpenRouter)

Google Gemma (OpenRouter)

gemma-4-31b (aliases: gemma, gemma-4)
gemma-4-26b

Qwen (OpenRouter)

qwen3.5-plus (qwen)
qwen3.5-flash
qwen3.5-397b
qwen3.5-27b
qwen3-coder (qwen-coder)
qwen3-coder-plus

Moonshot Kimi (OpenRouter)

kimi-k2.5 (kimi)
kimi-k2-thinking

Moonshot Kimi Code Direct (Anthropic-compatible)

kimi-for-coding (kimi-code)

MiniMax (OpenRouter)

minimax-m2.7 (minimax)
minimax-m2.5

Xiaomi MiMo Direct (no suffix: primary)

mimo-v2-pro (mimo; legacy alias: mimo-v2-pro-direct)
mimo-v2-flash (legacy alias: mimo-v2-flash-direct)

Xiaomi MiMo via OpenRouter (`-or` suffix)

mimo-v2-pro-or
mimo-v2-flash-or

GLM (Z.ai, OpenRouter)

glm-5
glm-5-turbo (alias: glm)

GLM Coding Plan Direct (Anthropic-compatible)

glm-5.1 (glm-code, glm-coding)
glm-5-turbo
glm-4.7
glm-4.5-air

xAI Grok (OpenRouter)

grok-4 (grok)
grok-4.20
grok-4.20-multi
grok-4-fast (grok-fast)
grok-4.1-fast
grok-code-fast (grok-code)
grok-3
grok-3-mini

Mistral (OpenRouter)

mistral-large-3 (aliases: mistral, mistral-large)
mistral-medium-3.1 (mistral-medium)
mistral-medium-3
mistral-small-4 (mistral-small)
mistral-small-3.2
magistral-medium (magistral)
magistral-small
codestral
devstral-2 (devstral)
devstral-medium
devstral-small
pixtral-large
ministral-3-14b (ministral)
ministral-3-8b

Preset token windows (max_context / max_output) are set per preset; see src/kohakuterrarium/llm/presets.py for the exact values, or kt config llm show <name>. controller.max_tokens overrides max_output; to adjust the compaction window, set compact.max_tokens.

Built-in preset merging also picks up llm_presets contributed by installed packages; see Package manifest in configuration.md.

Variation groups

A variation group lets one preset expose multiple knobs without duplicating the entry. Select with the preset@group=option shorthand in llm: / --llm, or via controller.variation_selections; see Variation selector in the configuration reference.

Presets not listed here have no variation group; their defaults are fixed.

OpenAI: Codex OAuth

Preset	Group	Options
`gpt-5.3-codex-spark`	`reasoning`	`low`, `medium`, `high`, `xhigh`
`gpt-5.5`	`reasoning`	`none`, `low`, `medium`, `high`, `xhigh`
`gpt-5.5`	`speed`	`normal`, `fast` (maps to `service_tier: priority`)
`gpt-5.4`	`reasoning`	`none`, `low`, `medium`, `high`, `xhigh`
`gpt-5.4`	`speed`	`normal`, `fast` (maps to `service_tier: priority`)
`gpt-5.3-codex`	`reasoning`	`none`, `low`, `medium`, `high`, `xhigh`
`gpt-5.1`	`reasoning`	`none`, `low`, `medium`, `high`, `xhigh`

OpenAI: Direct API (`-api` suffix)

Patches extra_body.reasoning.effort.

Preset	Group	Options
`gpt-5.4-api`, `gpt-5.4-mini-api`, `gpt-5.4-nano-api`, `gpt-5.3-codex-api`, `gpt-5.1-api`	`reasoning`	`none`, `low`, `medium`, `high`, `xhigh`

OpenAI: OpenRouter (`-or` suffix)

Patches extra_body.reasoning.effort via OpenRouter's unified param.

Preset	Group	Options
`gpt-5.4-or`, `gpt-5.4-mini-or`, `gpt-5.4-nano-or`, `gpt-5.3-codex-or`, `gpt-5.1-or`	`reasoning`	`minimal`, `low`, `medium`, `high`, `xhigh`

Anthropic: Direct

Patches extra_body.output_config.effort via the compat layer.

Preset	Group	Options
`claude-opus-4.7`	`reasoning`	`low`, `medium`, `high`, `xhigh`, `max`
`claude-opus-4.6`, `claude-sonnet-4.6`	`reasoning`	`low`, `medium`, `high`, `max`

Haiku 4.5 uses the older extended-thinking (budget_tokens) and has no variation group.

Anthropic: OpenRouter (`-or` suffix)

Patches extra_body.reasoning.effort. xhigh is only honoured by Opus 4.7.

Preset	Group	Options
`claude-opus-4.7-or`	`reasoning`	`minimal`, `low`, `medium`, `high`, `xhigh`
`claude-opus-4.6-or`, `claude-sonnet-4.6-or`, `claude-sonnet-4.5-or`, `claude-opus-4-or`, `claude-sonnet-4-or`	`reasoning`	`minimal`, `low`, `medium`, `high`
`claude-haiku-4.5-or`	`reasoning`	`off`, `low`, `medium`, `high`

Google Gemini: Direct

Patches extra_body.google.thinking_config.thinking_level.

Preset	Group	Options
`gemini-3.1-pro`	`thinking`	`low`, `medium`, `high`
`gemini-3-flash`, `gemini-3.1-flash-lite`	`thinking`	`minimal`, `low`, `medium`, `high`

Google Gemini: OpenRouter

Preset	Group	Options
`gemini-3.1-pro-or`, `gemini-3-flash-or`, `gemini-3.1-flash-lite-or`	`reasoning`	`minimal`, `low`, `medium`, `high`

Gemma / Qwen / Kimi / MiMo / GLM: OpenRouter

Share the same OpenRouter unified reasoning group (unless noted).

Preset	Group	Options
`gemma-4-31b`, `gemma-4-26b`	`reasoning`	`minimal`, `low`, `medium`, `high`
`qwen3.5-plus`, `qwen3.5-flash`, `qwen3.5-397b`, `qwen3.5-27b`, `qwen3-coder`, `qwen3-coder-plus`	`reasoning`	`minimal`, `low`, `medium`, `high`
`kimi-k2.5`	`reasoning`	`minimal`, `low`, `medium`, `high`
`mimo-v2-pro`, `mimo-v2-flash`, `mimo-v2-pro-or`, `mimo-v2-flash-or`	`reasoning`	`minimal`, `low`, `medium`, `high`
`glm-5`, `glm-5-turbo`	`reasoning`	`minimal`, `low`, `medium`, `high`

kimi-k2-thinking has always-on thinking and no variation group.

Mistral: OpenRouter

Preset	Group	Options
`mistral-small-4`	`reasoning`	`none`, `high`

Other Mistral presets (mistral-large-3, mistral-medium-*, mistral-small-3.2, codestral, devstral-*, pixtral-large, ministral-*) are not reasoning models. magistral-medium and magistral-small have always-on reasoning and no variation group.

Grok / MiniMax: OpenRouter

Grok 4.x (grok-4, grok-4.20, grok-4.20-multi, grok-4-fast, grok-4.1-fast, grok-code-fast) has mandatory non-configurable reasoning. grok-3 / grok-3-mini are legacy non-reasoning models. minimax-m2.7 / minimax-m2.5 have mandatory reasoning. None expose a variation group.

Prompt plugins

Shipped prompt plugins (loaded by prompt/aggregator.py). Ordered by priority (lower = earlier).

Priority	Name	Emits
50	`ToolListPlugin`	Tool name + one-line description.
45	`FrameworkHintsPlugin`	Framework command examples (`info`, `read_job`, `jobs`, `wait`, native tool usage) and tool-call format examples.
40	`EnvInfoPlugin`	`cwd`, platform, date/time.
30	`ProjectInstructionsPlugin`	Loads `CLAUDE.md` and `.claude/rules.md`.

Custom prompt plugins subclass BasePlugin and register via the plugins field in a creature config. See plugin-hooks.md for lifecycle and callback hooks.

Separate from prompt plugins, tools may also contribute one-line or short paragraph guidance into the aggregated ## Tool guidance section.

Compose algebra

Operator precedence: * > | > & > >>.

Operator	Meaning
`a >> b`	Sequence (auto-flatten). `>> {key: fn}` forms a Router.
`a & b`	Product (`asyncio.gather`; broadcast input).
`a \| b`	Fallback (catch exception, try next).
`a * N`	Retry (N additional attempts).

Factories: Pure, Sequence, Product, Fallback, Retry, Router, Iterator. Wrapping helpers: agent(config_path) for persistent agents, factory(config) for ephemeral per-call agents.

Runnable methods: .map(f) (post-transform output), .contramap(f) (pre-transform input), .fails_when(pred) (raise on a predicate).

MCP surface

Built-in MCP meta-tools (exposed when mcp_servers is configured):

mcp_list: list connected servers and their tools.
mcp_call: invoke a tool on a specific server.
mcp_connect: connect to a server declared in config.
mcp_disconnect: tear down a connection.

Server tools are surfaced in the system prompt under ## Available MCP Tools. Transports: stdio (subprocess), streamable_http (modern HTTP MCP), and legacy http/sse SSE endpoints.

Python surface: MCPServerConfig, MCPClientManager in kohakuterrarium.mcp.

Extensions

A package's kohaku.yaml may contribute creatures, terrariums, tools, plugins, io, triggers, skills, commands, user_commands, prompts, framework_hints, llm_presets, and python_dependencies. kt extension list inventories them. Python modules resolve by module:class refs; configs resolve via @pkg/path. See Package manifest in configuration.md.