Future AGI

What is Future AGI?

Future AGI is an end-to-end platform for building, testing, monitoring, and improving LLM applications and AI agents. It integrates simulation, evaluation, safeguards, tracking, gateways, datasets, and prompt optimization into a single feedback loop.

The platform offers managed Cloud services, as well as the core code that can be used for self-hosting. Engineers, product managers, and industry experts can view information related to quality, cost, latency, errors, and manual annotations within the same project.

A one-sentence summary

Future AGI helps teams simulate AI agents before deployment, monitor and evaluate them after deployment, and continuously refine prompts and agent processes using real-time safeguards and production feedback.

Six major product modules

AI agent simulation

Simulate allows for the use of preset or custom characters to conduct multiple rounds of text and voice conversations before going online. The scenarios can be defined using scripts, datasets, or graph structures, and it provides functionality for storing transcripts, comparing conversations, and tracking results.

What is suitable for testing using simulation?

• The complete process for a regular user, from submitting a question to completing the task.
• Vagueness, contradictions, overreach of authority, and prompt injection inputs.
• Tool failures, timeouts, repeated calls, and missing data.
• Characters with different languages, tones, backgrounds, and levels of knowledge.
• Voice interruptions, silence, accents, and recognition errors.
• Memory, state, and goal drift in long sessions.

Multi-type AI evaluation

Evaluate supports heuristic metrics, code evaluation, LLM-as-judge, and agent evaluation agents. The public repository lists over 50 metrics, including factual accuracy, hallucinations, correct use of tools, personal information, and tone.

Comparison of evaluation methods

Production tracking and observability

Observe, which is based on OpenTelemetry, enables the tracking of LLMs and agents; it displays span charts, sessions, end users, latency, tokens, and costs. The system supports 11 types of spans, over 70 filtering conditions, and full-text search.

TraceAI integration

TraceAI offers automatic instrumentation for over 50 AI frameworks, including LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, DSPy, and others. There are corresponding integration methods available for Python, TypeScript, Java, and C#.

Real-time guardrail protection

Protect can detect personal information, keys, jailbreaks, prompt injection, malicious content, and other unsafe elements. 15 rules-based safeguards are available for free forever, while the machine learning version of Protect is charged based on AI Credits.

Guardrail operation mode

Agent Command Center

Command Center is an LLM gateway that is compatible with the OpenAI interface and can route requests between more than 100 model providers. It offers automatic failover, retry mechanisms, virtual keys, budget controls, rate limiting, cost tracking, as well as precise or semantic caching.

Problems that gateways are suitable for solving

• Unified management of keys and requests from different model suppliers.
• Route based on latency, cost, weight, or availability.
• Automatic switching and retry in case of supplier failure.
• Reduce redundant calls and model costs through caching.
• Set budgets for organizations, teams, users, and virtual keys.
• Implement real-time safeguards in the request or response path.

Datasets, experiments, and manual annotation

The platform supports the creation of datasets in various formats such as manual entry, CSV, PDF, Hugging Face, and SDK, and offers 14 different column types. There are no limits on the number of datasets or experiments, and datasets along with their annotations are stored permanently across all Cloud solutions.

The free and pay-as-you-go plans offer 3 tagging queues, Boost provides 10, while the Scale and higher-tier plans have no limit. Tags can be of numerical, textual, categorical, star-rated, or like/dislike type.

Prompt words and Agent Playground

The Prompt module supports templates, versions, tags, and folders; Prompt Playground can be used for free with the built-in model keys. Agent Playground offers visual graph construction, tool invocation, and output format validation.

Falcon AI

Falcon AI can analyze and track data, evaluate errors, generate insights, apply automatic labeling, and produce performance reports. It shares 2,000 free AI credits per month with other hosted AI services.

Prompting and Agent Optimization

Optimize allows the feedback from production tracking and evaluation results to be used for improving the prompt texts, and it provides algorithms such as GEPA, PromptWizard, ProTeGi, Bayesian, Meta-Prompt, and Random. The improved versions still need to be validated on separate test sets.

Create the first tracking tutorial

1. Register for Cloud or deploy self-hosted instances and create a project.
2. Choose a Python, TypeScript, Java, or C# instrumentation package.
3. Configure the project name, endpoints, and securely saved authentication details.
4. Enable the corresponding Instrumentor for the model or framework being used.
5. Run a test request and check the span chart, tokens, and latency.
6. Add business attributes such as users, sessions, versions, and environments.
7. Configure error filtering, dashboards, and alerts.

Create a tutorial for evaluation

1. Organize the representative inputs, standard answers, and business conditions into a dataset.
2. First, use regular expressions or heuristic criteria to check the determinism requirements.
3. Create clear scales and examples of failures for open quality.
4. Choose either the BYOK model for judging or hosted Turing evaluation.
5. Add tools, RAG, and prompt error localization agents.
6. Use manually labeled samples to calibrate automatic scores and thresholds.
7. Include key evaluations in continuous integration and production sampling.

Tutorial on setting up gateways and safeguards

1. Add a model provider and create a virtual key with the minimum required permissions.
2. Configure the primary model, backup model, retry settings, and timeout.
3. Select the caching backend and identify the sensitive requests that cannot be cached.
4. Enable the safeguards for personal information, keys, and prompt injection rules.
5. First, use Monitor to observe false positives and false negatives, then set the Enforce threshold.
6. Configure budgets and throttling by organization, team, user, and key.
7. Monitor costs, failures, cache hits, and guardrail delays.

Docker self-hosting tutorial

1. Prepare Docker Engine or Docker Desktop along with Compose.
2. Clone the official future-agi repository and read the installation and security instructions.
3. Use the installation script to generate configurations and pull the published image.
4. Generate a separate key for the production environment and fix the image version.
5. Start components such as PostgreSQL, ClickHouse, Redis, RabbitMQ, and Temporal.
6. Restrict network access, enable backups, and set up monitoring.
7. First, verify the upgrade, recovery, and telemetry settings in a non-production environment.

Which users are it suitable for

• AI Application Engineer: Tracks model, tool, and agent calls.
• Evaluation team: Establishes offline, continuous, and production-quality testing.
• RAG team: Detects factual basis, hallucinations, and retrieval issues.
• Voice AI agent team: Simulates characters, scenarios, and conversations.
• Security team: Configure personal information, injection, and content safeguards.
• Platform engineering team: Unifies routing, keys, caching, and costs.
• Product and domain experts: Participate in the review using datasets and annotation queues.
• Regulated enterprises: Control the location of data and infrastructure through self-hosting.

Typical use cases

• Simulate normal and attack sessions with the customer service AI before going live.
• Continuously evaluate whether RAG responses are well-founded and free from hallucinations.
• Track the latency, cost, and failures of each tool in multi-agent systems.
• Failover between multiple model providers is achieved through a gateway.
• Intercept personal information and injection attempts before a production response is generated.
• Generate an annotation queue and a regression dataset from error tracking.
• Compare different models, prompts, and agent versions.
• Deploy a comprehensive evaluation and observability platform within the corporate network.

Product advantages

• It covers the entire pipeline including simulation, evaluation, tracking, safeguards, gateways, and optimization.
• Cloud offers a generous free tier, with no limit on the number of team members or projects.
• The costs for heuristic evaluation, BYOK model judging, and the rule safeguard platform are 0.
• OpenTelemetry and interfaces compatible with OpenAI help reduce access constraints.
• It supports more than 100 model providers and over 50 AI frameworks.
• The core platform is based on Apache 2.0 and can be self-hosted using Docker.
• Datasets and annotations are permanently stored in all Cloud solutions.
• Charging is done separately based on storage, credit points, gateway, caching, and simulation.

Usage restrictions and precautions

• The open-source repository is currently labeled as a Nightly early beta version; a stable version has not yet been released.
• Self-hosting involves multiple databases, queues, and running components, which results in high operational costs.
• Official Kubernetes and Helm support remains on the roadmap.
• AWS Marketplace is still marked as upcoming.
• Automatic evaluation and model judging may be inaccurate; therefore, manual samples are needed for calibration.
• Guardrails can lead to false positives and false negatives; high-risk decisions cannot rely solely on automatic blocking.
• The fact that the fee for the BYOK platform is 0 does not mean that it is free for model providers to use it.
• The tracking data for Cloud’s free and pay-as-you-go plans is retained for only 30 days.
• Permanently storing datasets can increase storage costs, and it is also necessary to proactively delete sensitive data.
• Self-hosting includes deployment telemetry by default; any external connections must be blocked at the network boundary.

Free Cloud quota

Tracking and gateway tiered pricing

Platform add-on packages

The price verification date is August 22, 2026. For the pay-as-you-go plan, it is possible to set up billing alerts and hard limits; in the free plan, service delivery is suspended once the allocated quota is reached, and no additional charges are generated automatically.

Self-hosting costs

Open-source self-hosting eliminates the cost of cloud subscriptions, but it requires investment in servers, databases, object storage, backup solutions, upgrades, monitoring services, security measures, as well as personnel costs. The use of external model and safeguard providers also incurs additional expenses.

Supported languages, SDKs, and integrations

Open source, self-hosting, and licensing

The complete core platform repository for Future AGI is licensed under Apache License 2.0; it includes tools for testing and evaluation, prompt management, tracking, as well as gateway code, and can be hosted using Docker Compose. The repository also contains a LICENSE-EE file, and it is necessary to check the directory structure before making use of the enterprise features.

Official repositories such as futureagi-sdk and Agent Command Center SDK also use Apache 2.0; some ecosystem packages may employ Apache or MIT licenses. It is still necessary to check the license file of each individual software package.

Deployment options

Telemetry and privacy

With self-hosting, registration and periodic heartbeat telemetry are sent by default; this telemetry may include the instance ID, version, deployment type, administrator’s email address and domain name, as well as anonymized aggregated usage data. Tracking information, prompt strings, or API keys are not sent.

Periodic heartbeats can be disabled through environmental settings, but even when disabled, a minimal set of statistics without email addresses is still sent. An environment that requires complete absence of any external communications should have such connections blocked at the network boundary.

Cloud Privacy and Deletion

The privacy policy lists names, email addresses, organizations, phone numbers, as well as IP addresses, browsers, visited pages, and diagnostic data. This information is used for services, support, analysis, security, and notifications; personal identification data is not sold.

Users may request access, correction, deletion, restriction, objection, and portability of their data; requests for deletion are processed within 30 days. Legal requirements, disputes, and anonymous statistics may constitute exceptions to this rule.

Basic information

Recommendation score

Recommendation score: 4.7 / 5. Future AGI integrates evaluation, observability, safeguards, gateways, simulation, and optimization into a single open-source platform; the free tier along with no limit on the number of users makes it very attractive to development teams.

The main risk is that the complete open-source repository is still in the Nightly stage; the self-hosted architecture is complex, and automatic evaluation mechanisms as well as safeguard systems require continuous calibration. For production use, a stable version must be adopted, along with completion of recovery, security, and capacity tests.

Frequently Asked Questions

Is Future AGI free?

There is a free Cloud option that does not require a credit card; it includes 50 GB of storage per month, 2,000 AI credits, 100,000 gateway requests, 1 million text simulation tokens, and 60 minutes of voice simulation time.

Will a charge be applied once the free quota is used up?

The free plan will be suspended and no bills will be generated automatically; charges will continue to be applied at the published excess rate only after a credit card is added and the pay-as-you-go option is activated.

Is Future AGI open source?

The core platform has been made open source under the Apache 2.0 license, but the repository is currently labeled as a nightly version for early testing; the enterprise directory may still be subject to the LICENSE-EE terms.

Can it be self-hosted?

The entire stack can be deployed using Docker Compose, or it can run on cloud virtual machines. Official support for Kubernetes and Helm is still part of the roadmap.

What SDKs are available?

It offers tracking capabilities in Python, TypeScript, Java, and C#, as well as Python or TypeScript packages related to evaluation, platforms, simulation, and gateways.

Which types of evaluations are supported?

It supports heuristic methods, code, LLM-as-judge, machine learning, agent evaluation agents, and manual annotation; the public ecosystem offers over 50 metrics.

What does BYOK mean?

Users provide their own model keys; Future AGI does not charge any fees for platform calls related to model evaluation, but the model provider will still charge based on the actual amount of usage.

Is there a fee for using the rule-based guardrails?

15 built-in rule-based safeguards are available permanently and free of charge. Machine learning Protect and hosted evaluations utilize shared AI Credits.

Is voice AI supported?

Text and voice simulation are supported; the free plan includes 60 minutes of voice time per month, with a charge of $0.08 per minute for any additional usage.

How long are the data stored?

Free and PAYG tracking data is retained for 30 days, Boost for 90 days, Scale for 1 year, while Enterprise options allow for custom retention periods; datasets and annotations are preserved permanently.

Are HIPAA and SSO available?

Scale offers HIPAA BAA, SAML SSO, and SCIM; Boost provides SOC 2 Type II and OAuth SSO, while Enterprise adds capabilities such as ABAC.

Are there MCP servers?

Yes, the Future AGI platform and its documents can be integrated with compatible clients such as Cursor or Claude Code.

Does self-hosting not send any telemetry at all?

It’s not the case that telemetry is completely absent by default. The periodic heartbeat can be disabled, but a minimum amount of statistics will still be collected; complete silence can only be achieved by blocking it at the network layer.

Is it suitable for Chinese teams?

Technical interfaces and self-hosting are suitable for development teams around the world, but the documentation is primarily in English. When dealing with data compliance requirements in China, it is necessary to choose an appropriate deployment location and conduct a legal assessment.

Summary

Future AGI is suitable for teams that wish to improve the quality of AI agents by utilizing a unified data loop; it offers dedicated modules for everything from pre-deployment simulation and offline evaluation to production monitoring, as well as for optimizing safeguards and prompts.

Small teams can first use the free Cloud and BYOK options for testing; organizations that are subject to regulations or deal with sensitive data can consider self-hosting solutions. In a production environment, it is important to pay attention to the nightly processing steps, the limitations imposed by enterprise licenses, telemetry functions, data retention policies, and the maintenance of complex components.