Changelog
1.1.0 (2026-09-26)
- test(apparmor): build the #1305 space-in-path workspace under tmp_path
- feat: keep a daemon-resolved app:// credential through the secret scrub (#1228)
- Revert "Let a profile-less workspace exempt GH_TOKEN from secret scrubbing"
- Let a profile-less workspace exempt GH_TOKEN from secret scrubbing
- Quote workspace_path in every AppArmor rule (#1305)
- docs(scaffold): name a profile-set after what varies, not who selects it
- server: tool_class, diff fields and auto_allow_housekeeping (jaato/#1304 §2 prerequisites)
- file_edit: steer writeNewFile/updateFile/removeFile over shell equivalents
- Fix WS sessions running unconfined: JaatoWSServer had no _event_loop
- Treat reasoning tokens as one quantity inside the output (#1047)
- Fix complexity ratchet violations from the #1294 diagnostics command
- Add self-service session diagnostics live view (#1294)
- Add observability to the WS AppArmor pre-init hook's silent exits (#1296)
- Fix stale WS-server reference and CREATE-path config_root gap (#1293)
- Persist reasoning parts instead of replaying them as text (#1290)
- fix(scaffold): Pi follow-ups after adopting #1093
- Re-anchor two reversions Phase 3 moved in session_manager.py
- Session group messaging phase 3: files on a message
- Session group messaging phase 2: the durable inbox
- fix: one attachment decoder for session.wake and session.message
- feat: session group messaging via the courier plugin (phase 1)
- Memories rail in the web client, backed by quiet daemon verbs (#1232)
- fix(runner): thread the managed workspace root through the SessionManager spawn path (#1280)
- fix(sandbox): fail closed on the SessionManager spawn path when confinement is required (#1253)
- notebook: detect, diagnose, and recover from a crashed subprocess kernel (#1275)
- Model subprocesses: jaato tools and installed tools on PATH, managed tool venv
- Staging: a file over 1 MiB closed the socket and hung for 120 s
- chore(release): jaato-sdk 0.28.0 — the cut absorbs #1263's protocol 1.21
- Web coder: bootstrap the jaato-sdk skill into every workspace (#1263)
- chore(release): cut jaato-sdk 0.27.1, jaato-server 1.1.0, jaato-tui 1.0.1, jaato-eval 0.3.3
- integration: add --refresh (overwrite only when nothing local is lost)
- Frame prompt-library skills as reference, list with descriptions (#1262)
- Co-locate #1253 WS reversions with the tests they name (#1065)
- fix(sandbox): WS session must be confined before its runner serves work (#1253)
- fix: recover double-JSON-encoded tool-call arguments (#1242)
- fix(sandbox): revived runner denies cli exec of /bin/sh and gh (#1251)
- feat(scaffold): add Pi skill integration
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 1.1.0
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-1.1.0.tar.gz | 6.7 MB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| jaato_server-1.1.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 14.4 MB
Release files / jaato_server-1.1.0.tar.gz
| Download URL | jaato_server-1.1.0.tar.gz |
|---|---|
| Size | 6.7 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
27bfcfc4a574dd055e7a0cbffc83060e46d3071286633f3185a86f94afd186a0
|
|
BLAKE2b-256 checksum How to use checksums |
e5ea5aa399e203d7c0dfb5b87c02291309b0ea8de094dd78017abd8340343de4
|
| 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 26, 2026.
Transparency logRelease files / jaato_server-1.1.0-py3-none-any.whl
| Download URL | jaato_server-1.1.0-py3-none-any.whl |
|---|---|
| Size | 7.7 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
3c961a3b5f6c08e936bd7a9776cf40d102991ba2ac80135e09003368c13c3464
|
|
BLAKE2b-256 checksum How to use checksums |
e8608fca7d0407068177197a2f314f4771e3fb9c038b86dc4a6bb5297c0f0bf5
|
| 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 26, 2026.
Transparency log