10xGraph API

Formerly 10xscale-agentflow-cli. The server and CLI half of 10xGraph: write your AI logic, and this package runs the API around it.

10xGraph is open-source, self-hosted infrastructure for AI applications. You write the agents, tools and tasks; 10xGraph runs them behind an API server with authentication, role-based access, tenant isolation, durable state in Postgres and Redis, background jobs, metrics and tracing, and generates Docker and Kubernetes deploy files. MIT licensed, by 10xScale. Formerly Agentflow.

10xgraph-api is the server and the 10xgraph command. Point 10xgraph.json at your compiled graph and it serves REST, NDJSON streaming, WebSocket and realtime-audio endpoints behind authentication, roles and tenant-aware access control, then writes your Docker and Kubernetes files. Built for teams going through SOC 2 and ISO 27001 audits: see the control mapping for what the platform supports and what stays with you.

---

What the server gives you

• Endpoints generated from the graph. Invoke, NDJSON stream, stop, WebSocket and realtime audio; threads, state and messages; the memory store; file uploads for multimodal input; and an optional AG-UI endpoint for CopilotKit and other AG-UI clients. See Invoke and stream.
• Authentication. "auth": "jwt", or your own BaseAuth subclass. See Authentication.
• Roles with reach. Each role carries permissions such as graph:invoke or threads:delete, and a reach of own, tenant or all. Reach widens reads, lists, deletes and stop; runs and writes always stay with the owner. The server refuses to boot if a non-public route has no permission guard. See Tenancy and access.
• Tenant isolation. Ownership is enforced in storage: a thread, message, memory, file or run outside the caller's reach answers 404. You declare the tenant columns; tenant_claims must match them.
• Cache store. One cache_store object holds run locks and caps, idempotency, token revocation and rate-limit counters: in memory, on Redis, or your own BaseCacheStore. See Cache store and Rate limiting.
• Background jobs. A thin TaskIQ layer: 10xgraph worker, 10xgraph scheduler and /v1/tasks routes, run under the same auth. See Background tasks.
• Metrics, logs and audit. Prometheus metrics from the metrics block, JSON logs with LOG_FORMAT=json, and a metadata-only audit log from the audit block that you ship to your own sink. 10xGraph stores no audit data. See Observability, Metrics with Prometheus and Grafana and Audit log.
• Hardened defaults. Request size limits, security headers, error and log sanitization, SSRF checks on external media URLs, startup warnings for insecure settings. See the Production checklist.
• Deploy files. 10xgraph build --docker-compose --k8s writes a Dockerfile, docker-compose.yml and k8s.yaml. See Deploy and Kubernetes.

---

Installation

pip install 10xgraph-api

Optional extras; install only what you configure:

pip install "10xgraph-api[jwt,redis,websocket]"  # JWT auth, Redis cache store, WebSocket and realtime audio
pip install "10xgraph-api[jobs]"      # background jobs (TaskIQ worker and scheduler)
pip install "10xgraph-api[metrics]"   # Prometheus scrape or OTLP push metrics
pip install "10xgraph-api[otel]"      # OpenTelemetry tracing
pip install "10xgraph-api[ag-ui]"     # AG-UI endpoint
pip install "10xgraph-api[media]"     # document text extraction (multimodal)
pip install "10xgraph-api[images]"    # image resize, EXIF, format conversion
pip install "10xgraph-api[snowflakekit]"  # Snowflake ID generation
pip install "10xgraph-api[sentry]"    # Sentry error reporting
pip install "10xgraph-api[gcloud]"    # Google Cloud Logging
pip install "10xgraph-api[firebase]"  # Firebase Admin SDK

jobs-rabbitmq and jobs-kafka add a RabbitMQ or Kafka broker for jobs.

Requires Python 3.12 or newer. Installs the core 10xgraph package as a dependency.

---

Quick start

# 1. Scaffold a project (asks for a name, a template and an AI provider)
10xgraph init

# 2. Run the dev server and open the playground (127.0.0.1:8000)
10xgraph dev

Auth is off in the Quick Start template. To turn it on, add the auth object to 10xgraph.json and set a secret in .env:

{ "auth": { "method": "jwt" } }

JWT_SECRET_KEY=<32+ characters>

"auth": "jwt" is a shortcut for the same thing. Every authenticated user then gets the default permission set with reach own. Add roles and tenant_claims for the setup below, or pick the Production template in 10xgraph init, which writes JWT auth, a Redis cache store with rate limits and token revocation, a prompt-injection validator, evals and tests.

---

Configuration

10xgraph.json defines your agent and the infrastructure around it. A realistic example:

{
  "agent": "graph.react:app",
  "env": ".env",
  "auth": {
    "method": "jwt",
    "roles_claim": "roles",
    "tenant_claims": ["org_id"],
    "roles": {
      "member": {
        "permissions": ["graph:*", "threads:*", "store:*", "files:*", "tasks:trigger", "tasks:read", "tasks:cancel"],
        "reach": "own"
      },
      "org_admin": { "inherits": ["member"], "permissions": ["admin:read"], "reach": "tenant" }
    },
    "default_role": "member"
  },
  "cache_store": {
    "type": "redis",
    "rate_limit": { "enabled": true, "requests": 100, "window": 60, "by": "user", "fail_open": false },
    "token_revocation": true
  },
  "audit": { "enabled": true, "sinks": ["log"] }
}

• The Redis URL comes from the REDIS_URL environment variable when cache_store.url is unset.
• tenant_claims must match the tenant columns declared on your checkpointer and store. A mismatch refuses to boot in production and warns in development.
• Audit sinks are log (JSON lines from the tenxgraph_api.audit logger), publisher (your events publisher) or module:Class (a BaseAuditSink).

Other keys include checkpointer, store, injectq, thread_name_generator, id_generator, ag_ui, websocket, runs and media. Edit them in a browser with 10xgraph config, or read the configuration reference and Configure the server.

For a custom auth class, set { "auth": { "method": "custom", "path": "auth.custom:MyAuth" } } and subclass tenxgraph_api.BaseAuth. See Authentication and Extensibility.

---

CLI

Run 10xgraph --help or 10xgraph COMMAND --help for options. Reference: CLI.

Command	What it does
10xgraph init	Scaffold a project. Asks for name, template (Quick Start or Production) and AI provider.
10xgraph dev	Run the dev server with reload and open the hosted playground. --no-open runs the server only.
10xgraph config	Edit 10xgraph.json in a local browser editor, with validation and a .bak of the previous file.
10xgraph build	Write a Dockerfile, and with --docker-compose and --k8s a Compose file and a Kubernetes manifest.
10xgraph worker	Run the TaskIQ worker for background jobs.
10xgraph scheduler	Run the TaskIQ scheduler. Run exactly one.
10xgraph revoke	Revoke one token (--jti) or all tokens of a user (--user).
10xgraph audit	Read-only check of the interpreter, installed packages, 10xgraph.json and the default port. Exits 1 on failure, so it works as a CI gate.
10xgraph test	Run the project tests with pytest.
10xgraph eval	Run agent evaluations and write JSON and HTML reports to eval_reports/.
10xgraph skills	Install the bundled 10xGraph skill for coding agents (Codex, Claude, GitHub Copilot).
10xgraph demo	Preview the terminal animations without touching the project.
10xgraph version	Show the CLI and core versions.

Machine-readable output is available with --format json or --format jsonl, and --no-animation gives a static screen.

---

Deploy

10xgraph build --docker-compose --k8s
docker compose up --build -d

This writes a multi-stage, non-root Dockerfile that runs Gunicorn with MODE=production and LOG_FORMAT=json, plus docker-compose.yml and k8s.yaml. Liveness is /ping; readiness is /health/ready and checks the checkpointer, store and cache store. The manifest has a fixed replica count, so autoscaling is yours. Run more than one worker or replica only with a Redis cache store. See Deploy, Kubernetes and Scaling.

---

Honest limits

• Auth is off until you configure it. The bare server has hardened defaults, not authentication. The Production template of 10xgraph init turns it on.
• Tenant isolation needs columns you declare. Declare tenant_columns on the checkpointer and store; it is not automatic.
• Permissions apply per route, not per tool. There is no built-in rule such as "user A may call refund, user B may not". Tools can read the caller's verified identity from the run config and decide for themselves.
• No hosted service. You run it, or nobody does.
• No audit evidence of its own. 10xGraph supports your controls; it does not hold a SOC 2 report. See Limits.

---

One stack

Package	What it does	Install
10xgraph	Core: graphs, state, replay-safe tools, checkpointing, memory	pip install 10xgraph (pulled in by this package)
10xgraph-api (this package)	Server and the 10xgraph command	pip install 10xgraph-api
10xgraph-client	Typed TypeScript client for every endpoint	npm install 10xgraph-client
Docs	Guides, reference and the control mapping	10xgraph.com

---

Migrating from 10xscale-agentflow-cli

10xGraph is the new name of 10xScale Agentflow. The server, CLI, license and maintainers are the same, and existing projects keep working.

	Before	Now
PyPI package	10xscale-agentflow-cli	10xgraph-api
Python import	agentflow_cli	tenxgraph_api
Command	agentflow	10xgraph
Config file	agentflow.json	10xgraph.json
Core package	10xscale-agentflow (import agentflow)	10xgraph (import tenxgraph)
CLI env vars	AGENTFLOW_NO_FULLSCREEN, AGENTFLOW_NO_SPINNER, AGENTFLOW_ASCII	TENXGRAPH_NO_FULLSCREEN, TENXGRAPH_NO_SPINNER, TENXGRAPH_ASCII
Website	agentflow.10xscale.ai	10xgraph.com
GitHub	github.com/10xHub	github.com/10xGraph

pip uninstall 10xscale-agentflow-cli 10xscale-agentflow
pip install 10xgraph-api

Do not install 10xscale-agentflow next to 10xgraph: both ship an agentflow module.

Kept working until 2.0:

• The agentflow command runs the same CLI and prints a deprecation notice.
• from agentflow_cli import ... resolves to tenxgraph_api with a DeprecationWarning.
• agentflow.json is read when no 10xgraph.json sits in the same directory; 10xgraph.json wins when both exist. The keys are unchanged.
• AGENTFLOW_* CLI env vars are read when the TENXGRAPH_* variable is unset.
• The agentflow-bearer WebSocket subprotocol and agentflow://media/ URLs are still accepted. New references use 10xgraph-bearer and graph://media/.

More: Coming from Agentflow.

---

Security

Report vulnerabilities privately; see SECURITY.md. Do not open a public issue for a security problem. Production settings are listed in the Production checklist and the security model.

---

Development

git clone https://github.com/10xGraph/10xgraph-api.git
cd 10xgraph-api
uv sync --dev
pre-commit install

pytest                         # tests (coverage gate: 80%)
ruff check . && ruff format .  # lint and format
uv run mypy

Contributing

Pull requests are welcome, and drafts are fine: open early and ask questions in the PR. Good first contributions:

• Type one module that mypy still skips (listed in pyproject.toml).
• Add a test for an edge case you hit: a request that should be rejected, a config that should fail at startup.
• Write the deployment recipe you needed on 10xgraph.com; every docs page has an "Edit this page" link.

For bigger changes, start a thread in the core repository's Discussions first.

License

MIT. Made by 10xScale, which runs its own AI products on 10xGraph.

Links

Documentation · PyPI · Issues · Changelog