AnkaLoop is for developers who want a useful coding agent now—not another framework to assemble.
It combines a capable tool loop, persistent context, multi-agent delegation, skills, memory, MCP,
hooks, and automation in one Python package that you can run locally or self-host.
Out-of-the-box coding-agent runtime for terminal, IDE, server, and Telegram: built-in tools, subagents, skills, memory, MCP/ACP, hooks, and automation.
ดาว
52
เครื่องมือ
—
อัปไทม์
—
Useful on the first run — read, search, edit, patch, and execute code with built-in tools.
Persistent by design — sessions, searchable history, project rules, and long-term memory
survive beyond a single prompt.
One runtime, multiple surfaces — use the same agent from the CLI, an HTTP/WebSocket server,
or Telegram.
Built for real tasks — context compaction, progress events, cancellation, bounded tool use,
and provider retries make long-running work observable and controllable.
Multi-agent without plumbing — delegate focused work to built-in explorer, planner, and coder
agents.
Open and extensible — add MCP servers, reusable skills, slash commands, hooks, and custom
agent specifications without replacing the built-in experience.
Provider-flexible — connect to OpenAI, Anthropic, GMI, or any OpenAI-compatible endpoint.
Quick start
AnkaLoop requires Python 3.11+ and credentials for a supported model provider.
Install from PyPI
# Install the latest stable release (with pip)
python -m pip install ankaloop
# Or install with uv
uv tool install ankaloop
# Configure a provider, then start in the current project
anka init
anka
# Or run one task and exit
anka --once "summarize this repository and suggest the next test to run"
The package installs both anka (recommended) and ankaloop commands.
All interfaces use the same core runtime. Conversation state is persistent, while project rules,
skills, and memory provide context across requests.
Usage
# Chat and one-shot tasks
anka
anka --once "create a hello.py file"
anka -t explorer --once "find all TODO comments"
anka --agent path/to/agent.yaml
# Sessions and agents
anka --session my-session
anka --list-sessions
anka --list-types
anka --clear
# MCP
anka mcp tools --server custom
anka mcp call --server custom --tool example_tool --args '{"query":"rust async"}'
# Server and remote client
anka serve
anka attach http://localhost:8080
# Telegram
anka telegram setup
anka telegram start
Run anka --help or anka <command> --help for the complete command reference.
Configuration
Run anka init for the interactive provider wizard. Configuration lives in
~/.config/ankaloop/config.toml. Set ANKA_CONFIG_DIR_NAME=amcp to keep the legacy
~/.config/amcp/ path from pre-rebrand deployments.
Keep credentials out of version control and provide the key through the environment:
export OPENAI_API_KEY="your-api-key"
anka --once "explain the architecture of this repository"
You can define multiple [chat.providers.<name>] profiles. Telegram administrators can list them
with /models and switch the active profile with /model use <name>.
[chat]
request_timeout_seconds = 120
max_retries = 2
retry_base_delay_seconds = 0.5
# Reasoning effort for models that support it:
# none | minimal | low | medium | high | xhigh | max (unset or "auto" = provider default).
# Can also be set per provider profile or through ANKA_REASONING_EFFORT.
reasoning_effort = "high"
# Keep model reasoning on a separate channel (default): stream APIs emit `reasoning`
# frames and responses carry `reasoning_content`. Set true to restore the legacy
# behavior of folding reasoning into the message content.
reasoning_as_content = false
# Replay each assistant message's reasoning_content in conversation history.
# Off by default: providers ignore a replayed reasoning trace, so sending it only
# inflates prompt tokens. Enable (usually per provider profile) only for providers
# that reject requests lacking it.
requires_reasoning_echo = false
tool_loop_limit = 300
bash_tool_limit = 100
default_max_lines = 400
mcp_tools_enabled = true
write_tool_enabled = true
edit_tool_enabled = true
sync_tool_settle_timeout_seconds = 2.0
default_agent = "coder"
[context]
progressive_tools = true
progressive_skills = true
response_ratio = 0.30
# HTTP/SSE transport
[servers.remote]
url = "https://example.com/mcp"
# stdio transport
[servers.local]
command = "npx"
args = ["-y", "@some/mcp-server"]
Configured tools are exposed as mcp__<server>__<tool>. Tool names are normalized for providers
with strict function-name requirements.
[tracing]
enabled = true
exporter = "otlp" # or "console"
endpoint = "http://localhost:4318"
capture_content = false # opt in to record prompts / tool arguments
Install the exporter with pip install "ankaloop[tracing]". Standard OTEL_* environment
variables override the TOML values. See the tracing guide.
The server exposes session, prompt, streaming, cancellation, timeline, tool, and agent APIs. Visit
/docs on a running server for its OpenAPI interface.
Non-loopback binds require a server API key. Configure [server.auth] or pass --api-key; deploy
behind TLS or a trusted reverse proxy.
The optional Telegram integration supports direct messages, groups and topics, allowlists,
pairing, streaming responses, bounded queues, cancellation, and session switching.