Changelog
0.18.0 (2026-09-20)
- fix(ci): the four checks #1165 turned red, each at its own layer
- feat(scaffold): ask the daemon about a topic this venv does not have
jaato-server
The runtime and daemon behind jaato — a multi-provider agentic tool orchestrator. This package holds the session runtime, the plugin system, ~24 model-provider adapters, and the daemon that clients connect to over a Unix socket or WebSocket.
You install this on the machine where agents should run. To talk to it from Python, install jaato-sdk; for a terminal client, jaato-tui.
Installation
pip install jaato-server
# or: uv pip install jaato-server
The base install carries the daemon, the session runtime and the plugins that need no third-party SDK. Provider SDKs and the heavier tool plugins are extras, so you install only the ones you use:
| Extra | Brings |
|---|---|
google |
Google GenAI / Vertex AI |
openai |
OpenAI (Chat Completions + the Responses API) |
azure-openai |
Azure OpenAI, including Microsoft Entra ID auth |
bedrock |
AWS Bedrock (boto3) |
openrouter, nim, ovhcloud |
OpenAI-compatible gateways |
github-models |
GitHub Models (azure-ai-inference) |
web |
web_fetch / web_search — extraction, PDF and search backends |
interactive |
interactive_shell — PTY sessions (pexpect / wexpect) |
ast |
ast_search — structural code search |
notebook, diagrams, templates, kaggle |
the matching tool plugins |
telemetry |
OpenTelemetry tracing |
kerberos |
SPNEGO/Negotiate proxy auth |
all |
everything above, plus the dev tooling |
pip install 'jaato-server[all]'
Anthropic's SDK, mcp, websockets and jaato-sdk are base dependencies — Claude, MCP servers and the WebSocket transport work out of the box. Requires Python 3.12+.
Quick start
Start the daemon:
# Local clients over a Unix domain socket
jaato-server --ipc-socket /tmp/jaato.sock --daemon
# Add a WebSocket port for remote or browser clients. A bearer token is
# generated at ~/.jaato/ws.token on first WS start if you supply none.
jaato-server --ipc-socket /tmp/jaato.sock --web-socket :8080 --daemon
jaato-server --status
jaato-server --stop
Then drive it from Python:
import asyncio
import jaato # from jaato-sdk
async def main():
async with jaato.session(mode="ipc", profile="researcher") as s:
print(await s.ask("What changed in this repo last week?"))
asyncio.run(main())
The same code runs against an embedded runtime (mode="in_process", no daemon), a local daemon ("ipc"), or a remote one ("ws").
Model providers
One session interface over every backend below; a profile picks the provider and model, and enter_tier can switch bindings mid-session.
| Hosted APIs | Anthropic, OpenAI, Azure OpenAI, Google GenAI / Vertex AI, AWS Bedrock, GitHub Models, Zhipu AI, MiniMax, Moonshot Kimi, Xiaomi MiMo |
| Gateways | OpenRouter (300+ models), NVIDIA NIM, Nebius Token Factory, OVHcloud AI Endpoints, Doubleword |
| Self-hosted | Ollama, LM Studio, vLLM, NVIDIA TensorRT-LLM, Triton |
| On-device | Chrome built-in AI (Gemini Nano, over the DevTools Protocol) |
| Subscription-backed | Claude Code CLI, Google Antigravity |
Several ship an interactive auth command (anthropic-auth login, github-auth login, openrouter-auth key …) so a credential need never be pasted into a shell profile.
Plugins
Tools reach the model through plugins, discovered from this package and from any installed distribution declaring a jaato.plugins entry point. Built-ins include:
- Execution —
cli(shell),interactive_shell(PTY sessions for REPLs, debuggers, wizards),notebook,mcp(any stdio MCP server) - Code and files —
file_edit,filesystem_query,ast_search,lsp,references - Web —
web_fetch,web_search,webmcp,service_connector(call an OpenAPI service by alias),webhook(inbound events) - Orchestration —
subagent,todo,clarification,waypoint,reliability,template - Context management — four GC strategies (
gc_truncate,gc_summarize,gc_hybrid,gc_budget) plus provider prompt-caching plugins - Control —
permission,sandbox_manager,telemetry
Agent profiles
A profile is a YAML file in .jaato/profiles/ binding a model, a provider, a plugin set, a GC strategy and resource ceilings:
name: researcher
description: Deep research profile
provider: anthropic
model: claude-sonnet-4-20250514
plugins: [cli, web_search, memory, todo(preload)]
default_agent: researcher # the persona in .jaato/agents/researcher.md
runtime_limits:
max_parallel_tools: 2
max_orphan_seconds: 300
budget_control:
limits: {usd: 5.0, tool_calls: 200}
degrade:
- {at: 95, action: finalize}
- {at: 100, action: abort}
jaato-scaffold interrogates the installed framework rather than a copy of the docs, so it cannot go stale:
jaato-scaffold explain profile # every profile key, with its vocabulary
jaato-scaffold explain providers # providers, their knobs and caveats
jaato-scaffold explain plugins # what is installed, and where it came from
jaato-scaffold validate # check a workspace before you run it
jaato-scaffold new client --profile researcher
explain covers 20-odd scopes — env, tiers, completion, runtime, paths, services among them.
Isolation and secrets
Sessions run in a subprocess runner, and the daemon confines it where the host allows:
- AppArmor — a per-boundary profile confines the runner to its workspace when the LSM is available. Opt in per client, or require it with
JAATO_REQUIRE_APPARMOR=1so an unconfined host refuses to start rather than degrading silently. - cgroup v2 —
runtime_limitscaps memory, pids and CPU; tool timeouts, output size, tool-pool width and two wall-clock bounds are enforced by the framework itself. - Secret scrubbing — provider keys and tokens are stripped from the environment handed to every model-driven subprocess by default;
scrub_secret_envnarrows or waives it, and waiving is announced. - Permissions — every tool call passes a policy that can allow, deny or ask, per session, with the decision recorded.
Configuration
Workspace configuration lives under .jaato/ (profiles, agents, GC settings, permissions) with ~/.jaato/ as the fallback tier. MCP servers are declared in .mcp.json. Credentials resolve from a profile knob, then the environment, then a stored auth file — pass:// and vault:// URIs are resolved daemon-side so a secret never sits on disk in the clear.
jaato-scaffold explain env prints every environment variable the installed tree reads, each tagged with its scope and the typed profile key that supersedes it.
Documentation
- Project README
- Architecture · Sequence diagrams
- Design philosophy
- Permission system
- AppArmor setup
- OpenTelemetry
- Issues
License
BUSL-1.1. See LICENSE.
Metadata
Release files for jaato-server 0.18.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| jaato_server-0.18.0.tar.gz | 6.3 MB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| jaato_server-0.18.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 13.6 MB
Release files / jaato_server-0.18.0.tar.gz
| Download URL | jaato_server-0.18.0.tar.gz |
|---|---|
| Size | 6.3 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
c3ad71ec134932b4a215febf2dadb3e9e40f85228c5d96fdb65f71ae7a81ffe6
|
|
BLAKE2b-256 checksum How to use checksums |
fcfcf7f87c3c2cb53b0194da61c7e1323d6a473ec867d1c1c5645f1cd2c2c124
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 20, 2026.
Transparency logRelease files / jaato_server-0.18.0-py3-none-any.whl
| Download URL | jaato_server-0.18.0-py3-none-any.whl |
|---|---|
| Size | 7.2 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
8310c6922a2bbdf94396ce37eb718d90ae9edc70244c159829ff5497daae3a52
|
|
BLAKE2b-256 checksum How to use checksums |
bac242beb63bcd37e00d959f40cdc5688a9c9fe3291b1c102ac1312e2521e9b2
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 20, 2026.
Transparency log