SuperAGI

Toolkits let agents take authenticated action in outside systems across many classes: search (Google Search, DuckDuckGo, Searx, SerpAPI), developer tools (GitHub including Fetch PR and Review PR, Jira, Coding toolkit), productivity (Notion, Google Calendar, File Manager), messaging (Email with attachments), social (Twitter, Instagram), image generation (DALL-E, Stable Diffusion), sales data (Apollo.io) and analytics (Google Analytics).

Credentials are configured per toolkit in the GUI with OAuth where the service requires it. Custom toolkits are documented as a first-class extension path through a linked GitHub repository, a cookie-cutter template and a community tools repository, so the catalogue is not a fixed list.

Sourcegithub.com/TransformerOptimus/SuperAGI README and release notes v0.0.9 through v0.0.12read 2026-09-01

The runtime carries the control flow rather than leaving it to the developer. Agents execute a ReAct reasoning-and-acting loop, and predefined agent workflows are selectable at provisioning as sequential instruction sets, with Sales Engagement and Recruitment shipped as worked multi-step examples. A wait block lets a workflow pause for a specified duration before continuing.

A task queue provides prioritisation, a configurable max-iteration limit bounds each run, looping detection heuristics flag an agent that is stuck, and provider errors such as rate limits, authentication failures and exceeded quotas are surfaced to the operator rather than failing silently. Multiple agents run concurrently.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.10 and v0.0.13, and README architecture diagramsread 2026-09-01

The buyer points the agent at their own content and it is retrieved at run time. Files in pdf, docx, pptx, csv, txt and epub are uploaded through the resource manager, converted to vector embeddings by the LlamaIndex integration and queried by the agent through the Resource Query tool.

Separately, Knowledge Embeddings can be created from the customer's own sources or installed from the marketplace, hosted in a Pinecone, Qdrant, Weaviate or Redis index the customer connects by URL, and reached at run time through the KnowledgeSearch toolkit by semantic search. A Knowledge Console logs embedding performance, queries and token consumption. Upgraded from Partial, which predated any check against the vendor's own release notes.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.7, v0.0.9 and v0.0.11, superagi.com knowledge-embeddings articleread 2026-09-01

Oversight is a configurable control rather than a property of one workflow. Restricted mode is selected per agent as a permission type at provisioning, and an agent running under it must seek human approval before critical actions, with writing files, sending emails and moving to the next iteration named. Requests surface in the Action Console where the operator can Approve, Give feedback or Deny, so the buyer decides which class of action needs sign-off, who signs it off and what happens while it waits. A configurable max-iteration limit bounds a run independently of approvals.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.4, v0.0.6 and v0.0.10read 2026-09-01

One of the four controls a security review asks for is documented and three are not. Present: identity integration, through GitHub OAuth used both for authentication and to protect the APIs, alongside a configurable auth secret and an option to restrict list creation, and code scanning through CodeQL in the project's own pipeline.

Absent: any named access model, so nothing states who inside an organisation can make an agent do what; any audit surface framed as governance rather than telemetry; and any compliance page a buyer could read, since the framework publishes no certification of its own. Downgraded from Full, which credited the GDPR, SOC 2 and HIPAA claims made by the company's separate commercial platform; those belong to that product and are not creditable here. Self-hosting means the operator supplies the surrounding controls themselves.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.4 and v0.0.7, config_template.yamlread 2026-09-01

Run-level visibility is inspectable after the fact and reaches the steps and tools, not just a counter. The Agent Performance Monitoring dashboard reports organisation metrics for agents, runs and tokens, per-model breakdowns of runs and consumption, and agent-level analytics covering total runs, average tokens per run, total calls and run time.

Model, Tool and Knowledge Consoles each expose their own logs, so a run can be traced through which model was called, which tools executed and which embeddings were queried. A real-time activity feed shows execution as it happens, APM retains data for deleted agents and runs, and the resource manager keeps a per-run output directory downloadable as a zip.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.8, v0.0.9 and v0.0.12read 2026-09-01

The memory layer is documented with a stated structure and lifetime. Short-term memory is a rolling window sized to the model's token limit; context falling outside it is condensed into a Long Term Summary, and the two combine into an Agent Summary passed into each reasoning step at a stated seventy-five to twenty-five weighting.

Long-term memory persists in a vector database, with Qdrant added as an LTM backend alongside Pinecone, Weaviate and Redis. Tool Memory stores every tool response from a run as vectorised chunks in Redis for semantic retrieval, though the vendor states only the Thinking Tool can read it.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.8, v0.0.10 and v0.0.11read 2026-09-01

Deployment goes well beyond shared multi-tenant software as a service. The framework is MIT-licensed and self-hosted by Docker Compose on the customer's own machine or infrastructure, with a separate GPU compose file supporting local large language models and multi-GPU execution so neither the platform nor the model need leave the environment.

A one-click Digital Ocean deployment runs the image on a self-hosted cloud machine and auto-updates with each release. Storage is configurable between local disk and S3, and the database and vector store are the operator's own. A hosted SuperAGI Cloud is offered alongside for those who want it.

Sourcegithub.com/TransformerOptimus/SuperAGI README installation section and release notes v0.0.4, v0.0.12 and v0.0.14read 2026-09-01

A browsable marketplace carries toolkits, agent templates, knowledge embeddings and models, reachable from the product and at its own subdomain.

Named templates aimed at named jobs are shipped rather than merely possible: SuperCoder builds end-to-end coding projects from goals and instructions, and predefined agent workflows cover Sales Engagement, fetching prospects from Apollo and drafting personalised outreach per row, and Recruitment, comparing uploaded CVs against a job description and sending shortlist or rejection mail.

Customers can save templates from any run and publish them to the marketplace after a validated run, so the library grows from the community as well as the vendor.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.6, v0.0.7, v0.0.10 and v0.0.12, marketplace.superagi.comread 2026-09-01

The trigger half is genuinely met and the channel half is not. Scheduled runs are documented in detail: a one-time future run, or a recurring run at a set interval with optional expiry by deadline or by total run count, and schedules can be stopped, edited and saved, so an agent starts without a person typing.

Webhooks fire outward on agent running, paused, completed, terminated and max-iterations-reached, which notifies external systems rather than starting work. What is absent is reach: the operator meets the agent only in the framework's own web interface at localhost or the hosted console, with no inbound channel from email, chat or a team tool. Email, Twitter and Instagram are outbound toolkits the agent acts through, counted on Integrations.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.8 and v0.0.12read 2026-09-01

Model choice is an explicit product surface. Providers are configured in account settings and a model is selected per agent from a dropdown at provisioning, with GPT-4-32k, gpt-3.5-turbo-16k and Google PaLM 2 Bison named and a model factory class built to admit more.

HuggingFace and Replicate integrations let the customer bring any public or self-hosted model by entering the model name and inference endpoint, a models marketplace offers preconfigured entries, and local large language models run against the customer's own weights with multi-GPU support. Upgraded from Partial, which predated any check against the vendor's own release notes.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.7, v0.0.8, v0.0.11 and v0.0.14read 2026-09-01

Public APIs are documented and reachable from cloud or local base URLs, letting developers create, pause, resume and update agents, start runs and retrieve run details and input or output files, with API keys generated in settings and the full reference published as a Postman collection.

Python and NodeJS SDK client libraries are published and installable by pip or npm with their own documentation section. User-defined HTTP callbacks are supported as webhooks on five named events: agent running, paused, completed, terminated and max iterations reached. Custom tools can be added by linking a GitHub repository from the frontend, with a cookie-cutter template and a community tools repository.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.11, v0.0.12 and v0.0.13, and documenter.getpostman.com SuperAGI collectionread 2026-09-01

The Coding toolkit ships a Write Unit Tests tool, and the SuperCoder sequence runs WriteSpec, WriteCode, ImproveCode then WriteTest, so tests are authored against the customer's generated project. What is not documented is the tests being run and the result gating anything: nothing states that a failing test blocks delivery or is reported back, which is the difference between authoring a criterion and applying one.

Everything else on this axis is after the fact, since Agent Performance Monitoring, the Model, Tool and Knowledge consoles and the activity feed report what happened and are graded on Observability. No evaluation harness, scored comparison between agent versions or preview run against a held-out set appears in any release note. Downgraded from Full, which was never checked against a vendor source.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.6 and v0.0.9read 2026-09-01

A Web Scraping tool ships in the toolkit set for pulling data from webpages, alongside a Web Interaction toolkit, which is the published Partial case for this axis: scraping and wired-in web access rather than control of a real interface the platform drives. No hosted or local browser, desktop session or remote computer control appears in any release note or in the README toolkit list, and the search tools are API-based. Upgraded from none, which treated scraping as zero; the bar places it at Partial.

Sourcegithub.com/TransformerOptimus/SuperAGI release notes v0.0.4 and README toolkit listread 2026-09-01