Kosmokrator

██╗  ██╗ ██████╗ ███████╗███╗   ███╗ ██████╗ ██╗  ██╗██████╗  █████╗ ████████╗ ██████╗ ██████╗
██║ ██╔╝██╔═══██╗██╔════╝████╗ ████║██╔═══██╗██║ ██╔╝██╔══██╗██╔══██╗╚══██╔══╝██╔═══██╗██╔══██╗
█████╔╝ ██║   ██║███████╗██╔████╔██║██║   ██║█████╔╝ ██████╔╝███████║   ██║   ██║   ██║██████╔╝
██╔═██╗ ██║   ██║╚════██║██║╚██╔╝██║██║   ██║██╔═██╗ ██╔══██╗██╔══██║   ██║   ██║   ██║██╔══██╗
██║  ██╗╚██████╔╝███████║██║ ╚═╝ ██║╚██████╔╝██║  ██╗██║  ██║██║  ██║   ██║   ╚██████╔╝██║  ██║
╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝     ╚═╝ ╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝   ╚═╝    ╚═════╝ ╚═╝  ╚═╝

AI coding agent for the terminal, headless automation, editors, and apps

Documentation website: kosmokrator.dev

---

KosmoKrator is a mythology-themed AI coding agent with one reusable runtime and several entry points. Use it interactively in the terminal, run it headlessly from scripts and CI, call integrations and MCP servers directly, embed it in PHP through the SDK, or wrap it from non-PHP apps through ACP.

The agent can read, write, and edit files, search your codebase, execute shell commands, run Lua workflows, call OpenCompany integrations, call MCP servers, and spawn parallel subagents. Permission modes, blocked paths, trust checks, sessions, settings, and credentials are shared across the terminal, headless CLI, SDK, and ACP surfaces.

Built with PHP 8.4+, Symfony Console, Symfony TUI, and async streaming via Amp/Revolt. The offline provider catalog exposes 90+ provider configurations and 3,500+ catalogued models through Prism, provider-native clients, OpenAI-compatible APIs, OAuth, and custom providers; live provider refresh can cache newer model inventories when credentials are available.

Table of Contents

• Installation
• CLI Usage
• Headless Agent CLI
• Headless Configuration
• ACP Server
• Agent SDK
• Integration CLI
• MCP CLI
• Agent Modes
• Slash Commands
• Tools
• Permission System
• Subagent System
• Context Management
• Providers
• Settings
• Configuration Files
• Sessions & Persistence
• Completion Sounds
• Rendering
• Architecture
• Development
• Changelog & Releases
• License

Installation

Quick install (recommended)

Auto-detects your OS and architecture:

curl -fsSL https://raw.githubusercontent.com/OpenCompanyApp/kosmokrator/main/install.sh | bash

Manual download

Pick the binary for your platform — no PHP required:

# macOS (Apple Silicon)
sudo curl -fSL https://github.com/OpenCompanyApp/kosmokrator/releases/latest/download/kosmokrator-macos-aarch64 \
  -o /usr/local/bin/kosmo && sudo chmod +x /usr/local/bin/kosmo

# macOS (Intel)
sudo curl -fSL https://github.com/OpenCompanyApp/kosmokrator/releases/latest/download/kosmokrator-macos-x86_64 \
  -o /usr/local/bin/kosmo && sudo chmod +x /usr/local/bin/kosmo

# Linux (x86_64)
sudo curl -fSL https://github.com/OpenCompanyApp/kosmokrator/releases/latest/download/kosmokrator-linux-x86_64 \
  -o /usr/local/bin/kosmo && sudo chmod +x /usr/local/bin/kosmo

# Linux (ARM)
sudo curl -fSL https://github.com/OpenCompanyApp/kosmokrator/releases/latest/download/kosmokrator-linux-aarch64 \
  -o /usr/local/bin/kosmo && sudo chmod +x /usr/local/bin/kosmo

The installer creates both commands. If you install manually and want the legacy command to keep working too, add:

sudo ln -sfn kosmo /usr/local/bin/kosmokrator

PHAR (requires PHP 8.4+)

If you already have PHP installed, the PHAR is smaller (~5MB vs ~25MB):

sudo curl -fSL https://github.com/OpenCompanyApp/kosmokrator/releases/latest/download/kosmokrator.phar \
  -o /usr/local/bin/kosmo && sudo chmod +x /usr/local/bin/kosmo

Add sudo ln -sfn kosmo /usr/local/bin/kosmokrator if you also want the legacy command name.

From source

git clone https://github.com/OpenCompanyApp/kosmokrator.git
cd kosmokrator
composer install
bin/kosmo setup

Requires PHP 8.4+ with pcntl, posix, and mbstring extensions, and Composer 2.x.

Getting started

kosmo setup    # First run — select provider and enter API key
kosmo          # Start the agent

TUI mode requires a modern terminal (iTerm2, Kitty, Ghostty, Windows Terminal, etc.). Falls back to ANSI mode automatically on simpler terminals.

CLI Usage

kosmo                         # Auto-detect renderer: TUI when available, ANSI fallback
kosmo --renderer=tui          # Force TUI mode
kosmo --renderer=ansi         # Force ANSI mode
kosmo --no-animation          # Skip the animated intro
kosmo --resume                # Resume the last session for the current project
kosmo --session <id>          # Resume a specific session by ID or prefix

The installed binary and bin/kosmo from a source checkout expose the same command surface.

Headless Agent CLI

Passing a prompt enables non-interactive execution. Use -p/--print when you want KosmoKrator to complete the task and exit:

kosmo -p "Explain the routing in this repository"
kosmo -p "Run the test suite and fix the failing tests" --mode edit --permission-mode guardian
kosmo -p "Review this branch" --mode plan --output-format json
kosmo -p "Summarize the architecture" --mode ask --output-format stream-json --no-session

Useful headless options:

Option	Purpose
-p, --print	Run one prompt and exit
-o, --output-format	text, json, or stream-json
-m, --model	Override the configured default model
--mode	edit, plan, or ask
--permission-mode	guardian, argus, or prometheus
--yolo	Alias for --permission-mode prometheus
--max-turns	Cap agentic turns
--timeout	Cap wall-clock runtime in seconds
--continue	Continue the most recent session
--session	Resume a specific persisted session
--system-prompt	Replace the system prompt
--append-system-prompt	Append extra system instructions
--no-session	Disable session persistence for this run

Headless mode uses HeadlessRenderer and the same AgentSessionBuilder/AgentLoop path as the interactive terminal. Tool permissions still apply. If a headless run cannot ask an approval question, choose a permission mode that matches the automation context.

Headless Configuration

Everything needed to run KosmoKrator without the interactive setup can be configured from the CLI:

kosmo setup                         # Interactive provider/model wizard
printf %s "$OPENAI_API_KEY" | kosmo setup --provider openai \
  --model gpt-5.4-mini --api-key-stdin --global --json

kosmo config show --json            # Inspect resolved configuration
kosmo config paths --json           # Show global/project config paths
kosmo settings:list --json          # Discover configurable settings
kosmo settings:set agent.mode plan --global --json
kosmo settings:options agent.default_model --provider openai --json
kosmo settings:doctor --json        # Diagnose headless readiness

kosmo providers:list --json
kosmo providers:models openai --live --json
kosmo providers:refresh-models openai --json
kosmo providers:doctor openai --model gpt-5.4-mini --json
printf %s "$OPENAI_API_KEY" | \
  kosmo providers:configure openai --api-key-stdin --model gpt-5.4-mini --global --json
kosmo providers:configure openai --model future-model-id --allow-unlisted-model --global --json
kosmo providers:configure codex --device --global --json
kosmo providers:custom:upsert local_ai --url http://127.0.0.1:11434/v1 --model qwen3:32b --json

printf %s "$OPENAI_API_KEY" | \
  kosmo secrets:set provider.openai.api_key --stdin --json

printf %s "$TELEGRAM_BOT_TOKEN" | \
  kosmo gateway:telegram:configure --token-stdin --enabled on --json
kosmo gateway:telegram:status --json
kosmo codex:login                   # ChatGPT OAuth for the Codex provider

The settings:*, providers:*, secrets:*, gateway:telegram:*, integrations:*, mcp:*, and web:* commands are designed for agents and scripts: they provide discovery helpers, stable JSON output, stdin secret entry, and non-zero exit codes on failures.

ACP Server

KosmoKrator can run as an ACP (Agent Client Protocol) server for editors and IDEs that speak ACP:

kosmo acp
kosmo acp --cwd /path/to/project --mode edit --permission-mode guardian
kosmo acp --yolo

ACP mode uses newline-delimited JSON-RPC over stdio. Stdout is reserved for protocol frames; diagnostics go to stderr. It reuses normal KosmoKrator provider credentials, sessions, permissions, Lua, integrations, MCP, memory, tasks, and subagents.

Supported ACP surface: initialize, new/load/resume/list/close sessions, prompt turns, cancellation, streamed assistant chunks, tool call updates, permission requests, session/set_config_option, and legacy session/set_model for clients that still use it. ACP mcpServers are added as runtime-only session MCP overlays and exposed through app.mcp.* without writing project .mcp.json.

KosmoKrator also advertises kosmoCapabilities and emits namespaced kosmo/* notifications for wrappers that need the terminal-grade model: phase changes, text/thinking deltas, tool lifecycle, permission lifecycle, usage, runtime changes, subagent spawn/tree/dashboard/completion, integration events, MCP events, and errors. Direct runtime JSON-RPC methods are available for non-PHP apps:

kosmo/runtime/set
kosmo/settings/set
kosmo/providers/configure
kosmo/integrations/list
kosmo/integrations/describe
kosmo/integrations/configure
kosmo/integrations/call
kosmo/mcp/list_servers
kosmo/mcp/list_tools
kosmo/mcp/schema
kosmo/mcp/add_stdio_server
kosmo/mcp/set_secret
kosmo/mcp/call_tool
kosmo/lua/execute

Those methods require a session ID so project root, permissions, session MCP overlays, credentials, and Lua namespaces match the active agent.

Example editor configuration:

{
  "agent_servers": {
    "kosmo": {
      "command": "kosmo",
      "args": ["acp"]
    }
  }
}

Agent SDK

KosmoKrator can also be embedded as a PHP library. The public SDK lives under Kosmokrator\Sdk\* and uses the same runtime path as kosmo -p:

use Kosmokrator\Sdk\AgentBuilder;

$result = AgentBuilder::create()
    ->forProject('/path/to/project')
    ->withMode('edit')
    ->withPermissionMode('guardian')
    ->withMaxTurns(20)
    ->build()
    ->collect('Fix the failing tests');

echo $result->text;

The SDK supports the headless CLI feature set: model/mode/permission overrides, sessions and resume, system prompt overrides, max turns, timeout, Lua code mode, integrations, MCP runtime overlays, subagents, callbacks, and programmatic provider/integration/MCP configuration helpers.

Integration CLI

KosmoKrator can run as a unified headless CLI for OpenCompany integrations. Use direct commands for single calls and Lua for multi-step workflows. The currently bundled command shortcuts include ClickUp, CoinGecko, Plane, and Plausible; the runtime is catalog-driven, so additional integration packages can expose the same command shape.

kosmo integrations:list --json
kosmo integrations:status --json
kosmo integrations:doctor clickup --json
kosmo integrations:fields clickup --json
kosmo integrations:configure clickup --set api_key="$CLICKUP_API_KEY" --enable --read=allow --write=ask --json
kosmo integrations:configure plane --account work --set api_key="$PLANE_API_KEY" --enable --json
kosmo integrations:search "create clickup task" --json
kosmo integrations:docs clickup
kosmo integrations:docs clickup.create_task
kosmo integrations:schema clickup.create_task

kosmo integrations:call clickup.create_task --list-id=123 --name="Ship it" --dry-run
kosmo integrations:call clickup.create_task --list-id=123 --name="Ship it" --force
kosmo integrations:clickup create_task --list-id=123 --name="Ship it"

kosmo integrations:lua --eval 'print(docs.read("clickup.create_task"))'
kosmo integrations:lua workflow.lua --force --json

Discovery commands expose credentials, activation, CLI compatibility, auth strategy, required runtime binaries, operations, and input schemas. Integrations that cannot be configured fully from the CLI are still listed because future proxy/web flows can support them.

Credentials support account aliases through --account, so one provider can have multiple independently configured identities. Headless calls still follow each provider function's read/write permission policy; --force bypasses only that integration permission policy for one trusted automation call.

MCP CLI

KosmoKrator can call MCP servers headlessly and from Lua without registering them as native model tools. It reads the common project .mcp.json mcpServers shape, plus VS Code/Cursor servers files and ~/.kosmo/mcp.json.

Runtime execution is stdio-first. Remote HTTP servers can be used through a stdio bridge such as mcp-remote until native Streamable HTTP execution is added.

kosmo mcp:list --json
kosmo mcp:add github --project --type=stdio \
  --command=github-mcp-server --env GITHUB_TOKEN --read=allow --write=ask --json
kosmo mcp:trust github --project --json
kosmo mcp:tools github --json
kosmo mcp:schema github.search_repositories --json

kosmo mcp:call github.search_repositories --query="kosmokrator" --json
kosmo mcp:github search_repositories --query="kosmokrator" --json
kosmo mcp:resources github --json
kosmo mcp:prompts github --json
kosmo mcp:lua workflow.lua --json

MCP project servers must be trusted before normal discovery/execution because they can start local commands. MCP read/write policy is configured separately under mcp.*; --force bypasses MCP trust and MCP read/write policy for one trusted automation call. Agent-side execute_lua, integrations:lua, and mcp:lua expose configured servers as app.mcp.*.

Agent Modes

Three agent modes control the tool families available to the model. They are separate from permission modes: edit can still ask for approval in Guardian/Argus, and ask can still run safe read/search commands. Switch modes interactively with /edit, /plan, and /ask, or headlessly with --mode.

Mode	Tools Available	Use Case
Edit (default)	Full coding tools, shell sessions, Lua execution, subagents, tasks, memories, integrations, MCP through Lua	Implementing features and fixing bugs
Plan	Read/search/bash/shell/subagent/task/ask/session/Lua docs tools; no file mutation or memory writes	Explore code and produce an implementation plan
Ask	Read/search/bash/shell/task/ask/session/Lua docs tools; no file mutation, subagents, or Lua execution	Answer questions with file context

Slash Commands

Type these at the prompt during a session.

Modes

Command	Description
/edit	Switch to Edit mode (full tool access)
/plan	Switch to Plan mode (read-only exploration)
/ask	Switch to Ask mode (Q&A with file context)

Permissions

Command	Description
/guardian	Heuristic auto-approve for safe commands (default)
/argus	Ask permission for every tool call
/prometheus	Auto-approve governed tool prompts until next prompt; hard denies still apply

Session Management

Command	Description
/new	Clear conversation and start a new session
/sessions	List recent sessions
/resume	Resume a previous session
/rename [name]	Rename the current session

Context & Memory

Command	Description
/compact	Force context compaction now
/memories	List stored memories
/forget <id>	Delete a memory by ID

Utilities

Command	Description
/settings	Open the settings workspace
/agents	Show the live subagent swarm dashboard
kosmo update	Check for updates and update based on install method
/feedback <text>	Submit feedback or a bug report as a GitHub issue (requires gh CLI)
/tasks-clear	Remove all tasks
/clear	Clear the terminal screen
/theogony	Replay the mythological intro animation
/quit	Exit KosmoKrator

Tools

KosmoKrator provides the LLM with a set of tools for interacting with your codebase and environment.

File Operations

Tool	Description
file_read	Read file contents with line numbers. Supports offset/limit for partial reads. Caches unchanged content to avoid redundant context.
file_write	Write entire files (new or complete overwrite). Creates missing parent directories.
file_edit	Targeted find-and-replace edits within a file.
apply_patch	Apply unified diff patches. Multi-file and multi-hunk support.

Search

Tool	Description
grep	Regex-powered code search (ripgrep-style). Supports file type/glob filters, content/file/count output modes, multiline matching.
glob	Fast file pattern matching (**/*.ts, src/**/*.php). Returns results sorted by modification time.

Shell

Tool	Description
bash	Execute a shell command. Streams output in real time. Configurable timeout (default 120s).
shell_start	Start a persistent interactive shell session. Returns a session ID.
shell_write	Send input to a running shell session.
shell_read	Read output from a shell session.
shell_kill	Terminate a shell session.

Coordination

Tool	Description
subagent	Spawn a parallel child agent with its own context window. Supports agent types, dependency chains, sequential groups, and concurrency control.

Interaction

Tool	Description
ask_user	Ask the user a question and wait for a response.
ask_choice	Present a choice to the user with optional visual mockups.

Lua

Tool	Description
lua_list_docs	List Lua and integration documentation pages available to the agent.
lua_search_docs	Search Lua and integration docs by topic or function name.
lua_read_doc	Read a specific Lua docs page before writing a script.
execute_lua	Run Lua with app.integrations.*, app.mcp.*, and approved native helper access.

Web

Tool	Description
web_search	Search the web through an optional configured web provider.
web_fetch_external	Fetch a URL through an optional configured external web provider.
web_extract	Extract a URL through an optional configured external web provider.
web_crawl	Crawl a site through an optional configured web provider.

Memory & Tasks

Tool	Description
memory_search	Search saved memories by type and text.
memory_save	Create or update a persistent memory (project, user, or decision).
session_search	Browse recent sessions or search prior session history for this project.
session_read	Read a prior session transcript by ID or unique prefix.
task_create	Create tasks with status tracking. Supports batch creation.
task_update	Update task status, description, or dependencies.
task_get	Retrieve a task by ID with full details.
task_list	List all tasks with status and blocked-by info.

Permission System

Three permission modes control tool approval. Governed native tools (file_write, file_edit, apply_patch, bash, shell_start, shell_write, execute_lua) go through permission checks. Integration and MCP calls also have read/write policies in headless CLI, Lua, SDK, and ACP flows.

Mode	Symbol	Behavior
Guardian (default)	◈	Auto-approve known-safe operations. Ask for writes and unknown commands.
Argus	◉	Ask permission for every governed tool call.
Prometheus	⚡	Auto-approve all governed calls. Deny rules still enforced.

Guardian Heuristics

Guardian auto-approves tool calls that match its safety rules:

• file_read, glob, grep — always auto-approved
• file_write, file_edit — auto-approved when the target is inside the project root
• bash — auto-approved when the command matches a known safe pattern AND contains no shell operators (;, &&, ||, |, redirects, substitutions, newlines)
• web_search, web_fetch_external, web_extract, web_crawl — run only when the relevant web provider setting is enabled

Safe command patterns include: git, ls, pwd, cat, head, tail, wc, find, which, echo, diff, php vendor/bin/phpunit, php vendor/bin/pint, composer, npm, npx, node, python, cargo, go, make.

Always Enforced

Regardless of permission mode:

• Blocked paths are always denied: *.env, .git/*, *.pem, *id_rsa*, *id_ed25519*, *.key
• Blocked bash patterns are always denied
• MCP project trust is required before normal MCP discovery/execution
• Session grants persist for the duration of the session — approve once, the tool is approved until you close

See docs/architecture/permission-modes.md for the full evaluation order and interaction with agent modes.

Subagent System

KosmoKrator can spawn parallel child agents, each with their own context window, to handle complex multi-part tasks.

Agent Types

Type	Can Write	Can Spawn	Purpose
General	Yes	General, Explore, Plan	Full coding access
Explore	No	Explore	Read-only research and search
Plan	No	Explore	Read-only planning and analysis

Permissions only narrow downward — a General agent can spawn any type, but an Explore agent can only spawn other Explore agents.

Features

• Dependency chains — agents can depend on other agents and wait for their results
• Sequential groups — run a batch of agents one at a time in order
• Parallel execution — up to 10 concurrent agents (configurable)
• Automatic retries — failed agents retry with exponential backoff (configurable)
• Stuck detection — headless agents are monitored for repetitive tool call patterns. Three-stage escalation: nudge, final notice, force return.
• Per-depth model overrides — run cheaper/faster models at subagent depths (configured via settings)
• Live dashboard — view progress, resource usage, and failures with /agents

See AGENTS.md for full documentation.

Context Management

Long conversations are managed automatically through a multi-stage pipeline:

1. Deduplication — detects and removes redundant tool outputs (e.g., reading the same file twice)
2. Pruning — removes superseded tool results (e.g., an old file_read replaced by a file_edit of the same file). Protects recent results (configurable prune_protect threshold).
3. Compaction — LLM-based summarization of older messages into a concise working memory. Extracts durable memories during compaction. Auto-triggers when context usage crosses the auto_compact_buffer_tokens threshold.
4. Trimming — emergency fallback that drops the oldest messages when context still overflows after compaction

Token Budgets

Context management uses a budget model with configurable thresholds:

Setting	Default	Purpose
reserve_output_tokens	16,000	Headroom reserved for the assistant response
warning_buffer_tokens	24,000	Show warnings when remaining input drops below this
auto_compact_buffer_tokens	12,000	Trigger auto-compaction when remaining input drops below this
blocking_buffer_tokens	3,000	Hard stop to prevent overrunning the model context window

Providers

Built-in provider configurations are defined in config/prism.yaml. KosmoKrator supports:

Provider	ID(s)	Auth
Anthropic	anthropic	API key
OpenAI	openai	API key
Codex (ChatGPT)	codex	OAuth (browser/device login)
Zhipu AI	z, z-api	API key
Moonshot (Kimi)	kimi, kimi-coding	API key
Xiaomi MiMo	mimo, mimo-api	API key / token plan
Google Gemini	gemini	API key
DeepSeek	deepseek	API key
Groq	groq	API key
Mistral	mistral	API key
xAI (Grok)	xai	API key
OpenRouter	openrouter	API key
Perplexity	perplexity	API key
MiniMax	minimax, minimax-cn	API key
StepFun	stepfun, stepfun-plan	API key
Ollama	ollama	None (local)

Model inventories come from the bundled prism-relay registry plus optional cached live discovery. Use kosmo providers:models <provider> --live --json or kosmo providers:refresh-models <provider> --json to query provider model endpoints and cache the result in SQLite. providers:list and providers:models report the active inventory source (bundled, provider_live, etc.). For newly launched models that have not appeared in metadata yet, headless setup supports --allow-unlisted-model.

Custom Providers

You can define custom providers in the settings workspace (/settings > Provider Setup) or directly in YAML config:

relay:
  providers:
    my-provider:
      label: My Provider
      driver: openai-compatible
      url: https://my-api.example.com/v1
      auth: api_key
      default_model: my-model
      models:
        my-model:
          display_name: My Model
          context: 128000
          max_output: 8192

Settings

Open the settings workspace with /settings during a session. Settings are organized into categories:

Category	Key Settings
General	Renderer (auto/tui/ansi), theme, intro animation
Models	Default provider and model. Browse providers and select via the models browser.
Provider Setup	Per-provider credential management, custom provider definitions
Auth	OAuth/device auth state such as Codex login
Agent	Default mode (edit/plan/ask), temperature, max output tokens, max retries
Permissions	Default permission mode (guardian/argus/prometheus)
Context & Memory	Auto compact, compact threshold, token buffer thresholds, prune settings, memories toggle
Gateway	Telegram gateway enablement and token configuration
Integrations	Integration activation, credential fields, account aliases, and read/write policy
MCP	MCP server config, project trust, server secrets, and read/write policy
Web	Optional web search/fetch/crawl provider configuration
Subagents	Subagent provider/model overrides (depth 1 and depth 2+), max depth, concurrency, retries, idle watchdog
Advanced	Lower-level runtime flags and diagnostics
Audio	Completion sound, soundfont path, composition timeout, max duration, retries

Scopes

Settings can be saved at two scopes:

• Project — applies only when KosmoKrator runs in the current working directory. Stored in .kosmo.yaml.
• Global — applies everywhere. Stored in ~/.kosmo/config.yaml.

Project settings override global settings which override built-in defaults.

Per-Depth Model Overrides

You can assign different models at each agent depth:

• Main agent (depth 0) — configured via the default provider/model
• Subagents (depth 1) — optional override via subagent provider/model
• Sub-subagents (depth 2+) — optional override via sub-subagent provider/model

Each level cascades: depth 2+ falls back to depth 1, which falls back to the main agent defaults.

Configuration Files

Configuration is loaded in layers (later overrides earlier):

1. config/kosmo.yaml — bundled defaults
2. config/prism.yaml — bundled provider definitions
3. ~/.kosmo/config.yaml — user overrides
4. .kosmo/config.yaml or .kosmo.yaml — project overrides discovered from the working directory upward

Environment variables (${VAR_NAME}) are resolved in all YAML files.

Other runtime state is stored separately:

Path	Purpose
~/.kosmo/data/	SQLite sessions, messages, memories, settings, and token accounting
~/.kosmo/mcp.json	Global MCP servers in portable mcpServers format
.mcp.json	Project MCP servers in the common mcpServers shape
.vscode/mcp.json, .cursor/mcp.json	Compatibility reads for editor servers MCP config
Integration credential config	Global/project activation and account aliases managed by integrations:configure

Key Configuration Options

kosmo:
  agent:
    default_provider: z
    default_model: glm-5.1
    temperature: 0.0
    max_retries: 0
    subagent_max_depth: 3
    subagent_concurrency: 10
    subagent_max_retries: 2
    subagent_idle_watchdog_seconds: 900

  ui:
    renderer: auto          # auto | tui | ansi
    intro_animated: true

  tools:
    default_permission_mode: guardian
    approval_required:
      - file_write
      - file_edit
      - apply_patch
      - bash
      - shell_start
      - shell_write
    bash:
      timeout: 120
    blocked_paths:
      - "*.env"
      - ".git/*"
      - "*.pem"
      - "*id_rsa*"
      - "*id_ed25519*"
      - "*.key"

  context:
    reserve_output_tokens: 16000
    warning_buffer_tokens: 24000
    auto_compact_buffer_tokens: 12000
    blocking_buffer_tokens: 3000
    compact_threshold: 60

  audio:
    completion_sound: true
    soundfont: ~/.kosmo/soundfonts/FluidR3_GM.sf2
    llm_timeout: 60
    max_duration: 8
    max_retries: 1

Sessions & Persistence

KosmoKrator persists state in SQLite under ~/.kosmo/data/:

• Sessions — full conversation history with token usage tracking. Auto-titled from the first user message.
• Messages — every message in each session, with role, content, and token counts.
• Memories — typed memories (project, user, decision) with retention classes (durable, working, priority), optional expiration, and pinning.
• Settings — global and project-scoped settings.

Use /sessions to list recent sessions, /resume to continue a previous session, and --resume or --session <id> from the command line.

Completion Sounds

KosmoKrator can compose and play a short musical piece after each agent response. The music reflects the outcome of the task.

• Outcome classification — the final message is analyzed to determine the mood: success (fanfare), tests passed (upbeat), tests failed (interrupted drop), failure (minor/descending), question (questioning), and more.
• Per-session instrument — each session gets a unique MIDI instrument (piano, vibraphone, guitar, violin, harp, trumpet, flute, etc.) based on a hash of the session ID.
• LLM-composed — the LLM generates a Python MIDI script with validation and safety checks. Falls back to hand-crafted scripts if generation fails.
• Non-blocking — composition and playback run in a background PHP process. The REPL is never blocked.
• Requires — Python 3 with midiutil, and fluidsynth with a SoundFont (.sf2) file.

Enable via /settings or set audio.completion_sound: true in config.

Rendering

KosmoKrator has a dual renderer architecture. RendererInterface is composed from 5 focused sub-interfaces: Core, Tool, Dialog, Conversation, and Subagent.

TUI Renderer

The default interactive renderer built on Symfony TUI:

• Full-screen layout with conversation history, status bar, and editor widget
• Multi-line input with Shift+Enter / Alt+Enter
• Slash command autocomplete with Tab
• Collapsible tool results with syntax highlighting
• Overlay dialogs for settings, session picker, plan approval, and permission prompts
• Animated breathing effect and cosmic spinners during thinking
• Live subagent progress tree
• Task bar with context usage
• Context progress bar with token counts and cost

ANSI Renderer

A pure ANSI escape code fallback for simpler terminals:

• Readline-based input
• CommonMark-to-ANSI markdown rendering via league/commonmark
• Syntax-highlighted code blocks via tempest/highlight
• Inline status updates

Both renderers share Theme for a consistent mythology-themed aesthetic — planetary tool icons (☽ ☉ ♅ ⚡︎ ⊛ ✧), cosmic spinner animations, and mythological thinking phrases.

Architecture

bin/kosmo
  → Kernel
  → AgentCommand | AcpCommand | integration/mcp/config commands | SDK entry points
  → AgentSessionBuilder
  → AgentLoop
      ├── ToolExecutor → tools + PermissionEvaluator
      ├── ContextManager → compaction, pruning, system prompt
      ├── StuckDetector → headless loop convergence
      ├── LLM client → AsyncLlmClient | PrismService | RetryableLlmClient
      ├── Renderer → TuiRenderer | AnsiRenderer | HeadlessRenderer | ACP/SDK renderers
      ├── ToolRegistry → coding + shell + web + Lua + coordination tools
      ├── IntegrationRuntime → OpenCompany integrations + app.integrations.*
      ├── McpRuntime → stdio MCP + app.mcp.*
      └── SubagentOrchestrator → parallel child agents

Key Directories

Directory	Purpose
src/Agent/	Agent core — AgentLoop (REPL), ToolExecutor, ContextManager, StuckDetector, SubagentOrchestrator, SubagentFactory
src/LLM/	LLM clients — AsyncLlmClient (Amp HTTP, async streaming), PrismService (Prism PHP, sync), RetryableLlmClient (decorator), ModelCatalog, ProviderCatalog
src/UI/Tui/	Symfony TUI renderer — TuiRenderer, TuiModalManager, TuiAnimationManager, SubagentDisplayManager, widgets
src/UI/Ansi/	ANSI fallback renderer — AnsiRenderer, MarkdownToAnsi, AnsiTableRenderer
src/UI/HeadlessRenderer.php	Non-interactive renderer for -p, JSON, and stream-json output
src/UI/Diff/	Unified diff rendering with word-level highlighting
src/Tool/Coding/	Tool implementations — file, bash, shell, grep, glob, subagent
src/Tool/Permission/	Permission system — PermissionEvaluator, PermissionMode, Guardian rules
src/Command/	CLI commands — AgentCommand, AcpCommand, SetupCommand, ConfigCommand, AuthCommand, UpdateCommand, settings, providers, secrets, gateway, integrations, MCP, web
src/Sdk/	Stable embeddable PHP API over the headless runtime: AgentBuilder, Agent, events, renderers, and configuration helpers
src/Acp/	Agent Client Protocol stdio server, JSON-RPC transport, session manager, and ACP renderer
src/Command/Slash/	In-session slash commands
src/Integration/	Headless OpenCompany integration runtime, catalog, credential resolution, CLI argument mapping
src/Mcp/	MCP config compatibility, stdio client, permission/trust checks, headless runtime, and Lua bridge
src/Lua/	Lua sandbox, documentation service, and native tool bridge
src/Session/	SQLite persistence — sessions, messages, memories, settings
src/Task/	Task tracking system with tool integrations
src/Audio/	Completion sound composition and playback
src/Settings/	Settings schema and multi-layer resolution

LLM Clients

KosmoKrator uses two LLM client implementations:

• AsyncLlmClient — non-blocking HTTP via Amp. Used for async-capable runtime paths, including TUI and headless surfaces that need OpenAI-compatible streaming behavior. Supports streaming, prompt caching, and tool calling.
• PrismService — synchronous client via Prism PHP SDK. Used for ANSI mode or providers with native Prism drivers when that path is selected.

Both are wrapped in RetryableLlmClient for automatic retry with exponential backoff on transient errors (429, 5xx).

Development

composer install
php vendor/bin/phpunit              # Run tests
php vendor/bin/pint                 # Code style (Laravel Pint)
php vendor/bin/phpstan analyse      # Static analysis

Building a PHAR

php vendor/bin/box compile          # Uses box.json config

Changelog & Releases

CHANGELOG.md is the source of truth for release notes. Add user-facing changes under ## [Unreleased] as part of the same PR that changes behavior, commands, providers, packaging, docs surfaces, or runtime behavior.

The website changelog is generated from the root changelog during the Astro build:

cd website
npm run sync:changelog
npm run build

Release helpers:

php bin/changelog check --base-ref=origin/main
php bin/changelog prepare 0.7.2
php bin/changelog extract v0.7.2

The release workflow uses bin/changelog extract "$GITHUB_REF_NAME" as the GitHub Release body, so tags must have a matching changelog section.

Conventions

• PSR-4 autoloading: Kosmokrator\ maps to src/
• declare(strict_types=1) everywhere
• Agent runs until the LLM signals completion — no hard tool round limit
• Extracted classes communicate via return values and closures — no circular dependencies
• Static utility classes are stateless and side-effect-free
• Markdown responses rendered with league/commonmark and tempest/highlight

Star History

Documentation

The full documentation site is available at https://kosmokrator.dev.

License

MIT