Unified Agent CLI + reusable Go agent core.

Other languages: 简体中文 | 日本語

• Why Mister Morph
• Quickstart
• Supported Models
• Daemon mode
• Console mode
• Telegram bot mode
• Slack bot mode
• Embedding
• Built-in Tools
• Skills
• Security
• Troubleshoots
• Debug
• Configuration

What makes this project worth looking at:

• 🧩 Reusable Go core: Run the agent as a CLI, or embed it as a library/subprocess in other apps.
• 🤝 Mesh Agent Exchange Protocol (MAEP): You and your amigos run multiple agents and want them to message each other: use the MAEP, a p2p protocol with trust-state and audit trails. (see docs/maep.md, WIP).
• 🔒 Serious secure defaults: Profile-based credential injection, Guard redaction, outbound policy controls, and async approvals with audit trails (see docs/security.md).
• 🧰 Practical Skills system: Discover + inject SKILL.md from file_state_dir/skills, with simple on/off control (see docs/skills.md).
• 📚 Beginner-friendly: Built as a learning-first agent project, with detailed design docs in docs/ and practical debugging tools like --inspect-prompt and --inspect-request.

Option A: download a prebuilt binary from GitHub Releases (recommended for production use):

curl -fsSL -o /tmp/install-mistermorph.sh https://raw.githubusercontent.com/quailyquaily/mistermorph/refs/heads/master/scripts/install-release.sh
sudo bash /tmp/install-mistermorph.sh

The installer supports:

• bash install-release.sh <version-tag>
• INSTALL_DIR=$HOME/.local/bin bash install-release.sh <version-tag>

Option B: install from source with Go:

go install github.com/quailyquaily/mistermorph@latest

mistermorph install
# or
mistermorph install <dir>

The install command installs required files and built-in skills under ~/.morph/skills/ (or a specified directory via <dir>).

When config.yaml does not already exist in the install target, install first tries to find a readable config in this order:

1. --config path
2. <dir>/config.yaml
3. ~/.morph/config.yaml

If none is found, install runs an interactive setup wizard (TTY only) before writing config.yaml:

1. select LLM provider (openai|gemini|cloudflare)
2. fill provider-specific required fields (api_key for openai/gemini; account_id + api_token for cloudflare)
3. set model
4. set Telegram bot_token + group_trigger_mode
5. optionally set Slack bot_token + app_token + group_trigger_mode

Use mistermorph install --yes to skip interactive prompts.

You can run without a config.yaml by using environment variables:

export MISTER_MORPH_LLM_API_KEY="YOUR_OPENAI_API_KEY_HERE"
# Optional explicit defaults:
export MISTER_MORPH_LLM_PROVIDER="openai"
export MISTER_MORPH_LLM_MODEL="gpt-5.2"

Mister Morph also supports Azure OpenAI, Anthropic Claude, AWS Bedrock, and others (see assets/config/config.example.yaml for more options). If you prefer file-based config, use ~/.morph/config.yaml.

mistermorph run --task "Hello!"

Model support may vary by specific model ID, provider endpoint capability, and tool-calling behavior.

Run a Telegram bot (long polling) so you can chat with the agent from Telegram:

Edit the config file ~/.morph/config.yaml and set your Telegram bot token:

telegram:
 bot_token: "YOUR_TELEGRAM_BOT_TOKEN_HERE"
 allowed_chat_ids: [] # add allowed chat ids here

mistermorph telegram --log-level info

Notes:

• Use /id to get the current chat id and add it to allowed_chat_ids for allowlisting.
• Use /ask <task> in groups.
• In groups, the bot also responds when you reply to it, or mention @BotUsername.
• You can send a file; it will be downloaded under file_cache_dir/telegram/ and the agent can process it. The agent can also send cached files back via telegram_send_file, and send voice messages via telegram_send_voice from local voice files in file_cache_dir.
• The last loaded skill(s) stay "sticky" per chat (so follow-up messages won’t forget SKILL.md); /reset clears this.
• telegram.group_trigger_mode=smart runs addressing LLM on every group message; acceptance requires addressed=true, confidence >= telegram.addressing_confidence_threshold, and interject > telegram.addressing_interject_threshold.
• telegram.group_trigger_mode=talkative also runs addressing LLM on every group message, but does not require addressed=true (it still uses the same confidence/interject thresholds).
• Use /reset in chat to clear conversation history.
• By default it runs multiple chats concurrently, but processes each chat serially (config: telegram.max_concurrency).

Run a Slack bot with Socket Mode so you can chat with the agent in Slack:

Edit the config file ~/.morph/config.yaml and set your Slack tokens:

slack:
 bot_token: "YOUR_SLACK_BOT_TOKEN_HERE" # xoxb-...
 app_token: "YOUR_SLACK_APP_TOKEN_HERE" # xapp-...
 allowed_team_ids: [] # optional allowlist
 allowed_channel_ids: [] # optional allowlist

mistermorph slack --log-level info

Notes:

• Requires both xoxb bot token and xapp app token.
• Group trigger and addressing controls mirror Telegram style (strict|smart|talkative + confidence/interject thresholds).
• By default it runs multiple conversations concurrently, but processes each team_id:channel_id conversation serially (slack.max_concurrency controls global concurrency).
• See docs/slack.md for setup and thread behavior details.
• See docs/bus.md for bus routing and ordering semantics.

Run a local HTTP daemon that accepts tasks sequentially (one-by-one), so you don’t need to restart the process for each task.

Start the daemon:

export MISTER_MORPH_SERVER_AUTH_TOKEN="change-me"
mistermorph serve --server-port 8787 --log-level info

Submit a task:

mistermorph submit --server-url http://127.0.0.1:8787 --auth-token "$MISTER_MORPH_SERVER_AUTH_TOKEN" --wait \
 --task "Summarize this repo and write to ./summary.md"

Run a local Console web UI for runtime inspection and file management.

Console currently includes:

• task list + task detail
• TODO files editor (TODO.md, TODO.DONE.md)
• contacts files editor (ACTIVE.md, INACTIVE.md)
• persona files editor (IDENTITY.md, SOUL.md)
• system diagnostics/config view

Build frontend:

cd web/console
pnpm install
pnpm build

Start daemon (task API source):

MISTER_MORPH_SERVER_AUTH_TOKEN=dev-token \
mistermorph serve --server-auth-token dev-token

Start Console backend + static hosting:

MISTER_MORPH_CONSOLE_PASSWORD=secret \
MISTER_MORPH_SERVER_AUTH_TOKEN=dev-token \
mistermorph console serve --console-static-dir ./web/console/dist

Open:

http://127.0.0.1:9080/console

More details: web/console/README.md.

Two common integration options:

• As a Go library: see demo/embed-go/.
• As a subprocess CLI: see demo/embed-cli/.

For Go-library embedding with built-in wiring, use integration:

cfg := integration.DefaultConfig()
cfg.BuiltinToolNames = []string{"read_file", "url_fetch", "todo_update"} // optional; empty = all built-ins
cfg.Inspect.Prompt = true // optional
cfg.Inspect.Request = true // optional
cfg.Set("llm.api_key", os.Getenv("OPENAI_API_KEY"))
rt := integration.New(cfg)
reg := rt.NewRegistry() // built-in tools wiring
prepared, err := rt.NewRunEngineWithRegistry(ctx, task, reg)
if err != nil { /* ... */ }
defer prepared.Cleanup()
final, runCtx, err := prepared.Engine.Run(ctx, task, agent.RunOptions{Model: prepared.Model})
_ = final
_ = runCtx

Core tools available to the agent:

• read_file: read local text files.
• write_file: write local text files under file_cache_dir or file_state_dir.
• bash: run a shell command (disabled by default).
• url_fetch: HTTP fetch with optional auth profiles.
• web_search: web search (DuckDuckGo HTML).
• plan_create: generate a structured plan.

Tools only available in Telegram mode:

• telegram_send_file: send a file in Telegram.
• telegram_send_voice: send a voice message in Telegram.
• telegram_react: add an emoji reaction in Telegram.

Please see docs/tools.md for detailed tool documentation.

mistermorph discovers skills under file_state_dir/skills (recursively), and injects selected SKILL.md content into the system prompt.

By default, run uses skills.mode=on, which loads skills from skills.load and optional $SkillName references (skills.auto=true).

Docs: docs/skills.md.

# list available skills
mistermorph skills list
# Use a specific skill in the run command
mistermorph run --task "..." --skills-mode on --skill skill-name
# install remote skills 
mistermorph skills install <remote-skill-url> 

1. Install audit: When installing remote skills, Mister Morph will preview the skill content and do a basic security audit (e.g., look for dangerous commands in scripts) before asking for user confirmation.
2. Auth profiles: Skills can declare required auth profiles in the auth_profiles field. The agent will only use skills whose auth profiles are configured on the host, preventing accidental secret leaks (see assets/skills/moltbook and the secrets / auth_profiles sections in the config file).

Recommended systemd hardening and secret handling: docs/security.md.

Known issues and workarounds: docs/troubleshoots.md.

There is an argument --log-level set for logging level and format:

mistermorph run --log-level debug --task "..."

There are 2 arguments --inspect-prompt/--inspect-request for dumping internal state for debugging:

mistermorph run --inspect-prompt --inspect-request --task "..."

These arguments will dump the final system/user/tool prompts and the full LLM request/response JSON as plain text files to ./dump directory.

mistermorph uses Viper, so you can configure it via flags, env vars, or a config file.

• Config file: --config /path/to/config.yaml (supports .yaml/.yml/.json/.toml/.ini)
• Env var prefix: MISTER_MORPH_
• Nested keys: replace . and - with _ (e.g. tools.bash.enabled → MISTER_MORPH_TOOLS_BASH_ENABLED=true)

Global (all commands)

• --config
• --log-level
• --log-format
• --log-add-source
• --log-include-thoughts
• --log-include-tool-params
• --log-include-skill-contents
• --log-max-thought-chars
• --log-max-json-bytes
• --log-max-string-value-chars
• --log-max-skill-content-chars
• --log-redact-key (repeatable)

run

• --task
• --provider
• --endpoint
• --model
• --api-key
• --llm-request-timeout
• --interactive
• --skills-dir (repeatable)
• --skill (repeatable)
• --skills-auto
• --skills-mode (off|on)
• --max-steps
• --parse-retries
• --max-token-budget
• --timeout
• --inspect-prompt
• --inspect-request

serve

• --server-bind
• --server-port
• --server-auth-token
• --server-max-queue

submit

• --task
• --server-url
• --auth-token
• --model
• --submit-timeout
• --wait
• --poll-interval

console serve

• --console-listen
• --console-base-path
• --console-static-dir
• --console-session-ttl

telegram

• --telegram-bot-token
• --telegram-allowed-chat-id (repeatable)
• --telegram-group-trigger-mode (strict|smart|talkative)
• --telegram-addressing-confidence-threshold
• --telegram-addressing-interject-threshold
• --telegram-poll-timeout
• --telegram-task-timeout
• --telegram-max-concurrency

slack

• --slack-bot-token
• --slack-app-token
• --slack-allowed-team-id (repeatable)
• --slack-allowed-channel-id (repeatable)
• --slack-group-trigger-mode (strict|smart|talkative)
• --slack-addressing-confidence-threshold
• --slack-addressing-interject-threshold
• --slack-task-timeout
• --slack-max-concurrency

skills

• skills list --skills-dir (repeatable)
• skills install --dest --dry-run --clean --skip-existing --timeout --max-bytes --yes

install

• install [dir]
• --yes

Common env vars (these map to config keys):

• MISTER_MORPH_CONFIG
• MISTER_MORPH_LLM_PROVIDER
• MISTER_MORPH_LLM_ENDPOINT
• MISTER_MORPH_LLM_MODEL
• MISTER_MORPH_LLM_API_KEY
• MISTER_MORPH_LLM_REQUEST_TIMEOUT
• MISTER_MORPH_LOGGING_LEVEL
• MISTER_MORPH_LOGGING_FORMAT
• MISTER_MORPH_SERVER_AUTH_TOKEN
• MISTER_MORPH_CONSOLE_PASSWORD
• MISTER_MORPH_CONSOLE_PASSWORD_HASH
• MISTER_MORPH_TELEGRAM_BOT_TOKEN
• MISTER_MORPH_SLACK_BOT_TOKEN
• MISTER_MORPH_SLACK_APP_TOKEN
• MISTER_MORPH_FILE_CACHE_DIR

Provider-specific settings use the same mapping, for example:

• llm.azure.deployment → MISTER_MORPH_LLM_AZURE_DEPLOYMENT
• llm.bedrock.model_arn → MISTER_MORPH_LLM_BEDROCK_MODEL_ARN

Tool toggles and limits also map to env vars, for example:

• MISTER_MORPH_TOOLS_BASH_ENABLED
• MISTER_MORPH_TOOLS_URL_FETCH_ENABLED
• MISTER_MORPH_TOOLS_URL_FETCH_MAX_BYTES

Secret values referenced by auth_profiles.*.credential.secret_ref are regular env vars too (example: JSONBILL_API_KEY).

Key meanings (see assets/config/config.example.yaml for the canonical list):

• Core: llm.provider selects the backend. Most providers use llm.endpoint/llm.api_key/llm.model. Azure uses llm.azure.deployment for deployment name, while endpoint/key are still read from llm.endpoint and llm.api_key. Bedrock uses llm.bedrock.*. llm.tools_emulation_mode controls tool-call emulation for models without native tool calling (off|fallback|force).
• Logging: logging.level (info shows progress; debug adds thoughts), logging.format (text|json), plus logging.include_thoughts and logging.include_tool_params (redacted).
• Loop: max_steps limits tool-call rounds; parse_retries retries invalid JSON; max_token_budget is a cumulative token cap (0 disables); timeout is the overall run timeout.
• Skills: skills.mode controls whether skills are used (off|on; legacy explicit/smart map to on); file_state_dir + skills.dir_name define the default skills root; skills.load always loads specific skills; skills.auto additionally loads $SkillName references.
• Tools: all tool toggles live under tools.* (e.g. tools.bash.enabled, tools.url_fetch.enabled) with per-tool limits and timeouts.