The best open-source AI agent frameworks in 2026
The best open-source AI agent frameworks right now are LangGraph, Browser Use and Pydantic AI, out of 80 we have reviewed. Each is ranked by our rating, with its price, our verdict and who it suits.
By ReviewRamble editors · Updated Nov 18, 2026 · How we review
How we ranked: by our rating out of 5, from reviews built on each product's own pricing, documentation and policies plus independent benchmarks where they exist, each checked on the date shown in the review. We take no affiliate commission.
- 1
Explicit control over every step an agent takes, paid for in code and learning time, and worth it once an agent must survive restarts and wait for a human.
Best for: Engineering teams building long-running or multi-step agents that must be reliable and auditable.
Price: Free and open source under the MIT licence. LangSmith (tracing and deployment) is separate: Developer free for one seat, Plus $39/seat/month plus compute, Enterprise custom.
Read our LangGraph review4.6/5 - 2
Browser Use is a sound base for agents that must work websites with no API, so long as someone switches off the telemetry before it sees a customer's data.
Best for: Engineering teams building agents that must operate websites without an API.
Price: Library free under MIT. Cloud is pay-as-you-go: browsers $0.02 per hour, residential proxy $5/GB, BU2 model $0.60/$3.50 per million input/output tokens, hosted agents at model price plus 20%.
Read our Browser Use review4.5/5 - 3
Pydantic AI brings the discipline of typed Python to agents, at the cost of a release train you have to pin.
Best for: Python engineers building production agents who want strong typing, testability and freedom to switch models.
Price: Free and open source under the MIT licence. Logfire observability: Personal free up to 10M records/month, Team $49/month plus $2 per extra million records. Pydantic AI Gateway adds a 5% markup on built-in providers.
Read our Pydantic AI review4.5/5 - 4
Temporal is the safest place to run an agent whose half-finished work would be expensive to lose, provided you accept its rules about what may run inside a workflow.
Best for: Teams running agents that last minutes to days, wait on people, or take actions that must not be repeated after a crash.
Price: Server free under MIT. Temporal Cloud: $50 per million Actions, falling in tiers to $25 per million between 100M and 200M on Business and above; active storage $0.042/GBh, retained $0.00105/GBh. Developer plan adds 10% of usage; Business is the greater of $500/month or 10% of usage, with 2.5M Actions included; Enterprise and Mission Critical priced annually by sales. $150 free credits for 90 days.
Read our Temporal review4.5/5 - 5
TypeScript's most complete agent toolkit, released at a pace that rewards teams who pin versions and read changelogs.
Best for: TypeScript and Next.js teams building agents and AI workflows.
Price: Core is free under Apache 2.0; code in ee/ folders needs a Mastra Enterprise licence in production. Mastra's hosted platform: Starter $0, Teams $250/month, Enterprise custom; self-hosted Enterprise is a flat annual fee.
Read our Mastra review4.4/5 - 6
The spec-first way to build MCP in Python; version 2 is cleaner, and it breaks every v1 import on the way.
Best for: Python developers building MCP servers or clients who want the official, spec-tracking implementation.
Price: Free and open source under the MIT licence.
Read our MCP Python SDK review4.4/5 - 7
A small, readable agent SDK whose defaults quietly assume OpenAI: tracing ships your runs to its dashboard, and the local sandbox trusts your machine.
Best for: Developers building agents mainly on OpenAI models who want a minimal, well-documented SDK.
Price: Free and open source under the MIT license. You pay for model usage.
Read our OpenAI Agents SDK review4.4/5 - 8
Hermes is the open-source agent to put on a $5 server if you want one that remembers you, with a security model you need to read before you turn on YOLO mode.
Best for: Developers and small teams who want a self-hosted assistant that runs scheduled jobs and answers in their chat apps.
Price: Free under the MIT licence; you pay for models and hosting. Optional Nous Portal: Free, Plus $20/month, Super $100/month, Ultra $200/month.
Read our Hermes Agent review4.3/5 - 9
OpenClaw is the personal agent everyone else is measured against, and it hands a model your shell by default, so it rewards owners who read the security guide first.
Best for: Technical users and small teams who want one assistant across their chat apps and devices and will run it sandboxed and patched.
Price: Free and open source (MIT). No paid tier; you pay your model provider.
Read our OpenClaw review4.3/5 - 10
Stagehand suits engineers who want to decide, line by line, how much of a browser script a model controls, and it quietly rewards running that script on Browserbase.
Best for: Engineering teams that want AI help on the brittle parts of browser automation while keeping the rest as ordinary, testable code.
Price: SDK free under MIT. Browserbase browsers: Free 1 hour; Developer $20/month (100 hours, then $0.12/hour); Startup $99/month (500 hours, then $0.10/hour); Scale custom. Model Gateway billed at provider prices.
Read our Stagehand review4.3/5 - 11
No faster route to an agent that already edits files and runs commands, provided you accept Claude Code as a closed engine inside it.
Best for: Developers who want a Claude Code-grade coding or file-working agent inside their own product and are standardised on Claude.
Price: Python wrapper free under the MIT licence; the bundled Claude Code binary and the TypeScript package are under Anthropic's Commercial Terms. You pay for Claude API usage, or draw on your own Claude plan's limits.
Read our Claude Agent SDK review4.2/5 - 12
A serious, batteries-included agent server from ByteDance that does far more than research now, provided you treat it as the code-executing service it is.
Best for: Teams that want a self-hosted, long-running agent service for research, reports and code jobs and can run it behind proper network controls.
Price: Free and open source (MIT); you pay for the model API and the server it runs on.
Read our DeerFlow review4.2/5 - 13
A serious, fast-moving agent framework whose shortest road always leads to Gemini and Google Cloud, which is either the point or the catch.
Best for: Teams building agents on Gemini or deploying to Google Cloud.
Price: Free and open source under the Apache 2.0 licence. You pay for the models and Google Cloud services you use.
Read our Google Agent Development Kit (ADK) review4.2/5 - 14
Still the widest integration library for LLM apps, much tidier since 1.0, and no longer the automatic default it once was.
Best for: Teams building agents that need many provider and data integrations and may grow into custom LangGraph workflows.
Price: Free and open source under the MIT licence. LangSmith is separate: Developer free for one seat, Plus $39/seat/month, Enterprise custom.
Read our LangChain review4.2/5 - 15
Headless Chrome's memory bill cut by an order of magnitude for agents and crawlers, as long as the pages you need sit inside what a young engine supports.
Best for: Teams running many concurrent scraping or browsing agents who are paying for headless Chrome's memory and can fall back to Chrome for the odd site.
Price: Free and open source (AGPL-3.0). Hosted cloud: Explorer free (10 browser hours, 5 concurrent sessions); Builder $19/month (300 hours, then $0.08/hour, 30 concurrent); Enterprise on request.
Read our Lightpanda review4.2/5 - 16
The obvious home for .NET and Azure teams building agents, and a capable but less proven choice for Python teams elsewhere.
Best for: .NET and Azure teams, and anyone migrating from AutoGen or Semantic Kernel.
Price: Free and open source under the MIT licence. You pay only for the models and Azure services you use.
Read our Microsoft Agent Framework review4.2/5 - 17
Light enough to run beside a single agent and sturdy enough to carry work between hundreds of them, provided you treat its security patches as urgent.
Best for: Teams running several independent agent or tool workers that need durable job queues and fan-out without operating Kafka.
Price: Free and open source under Apache 2.0. Synadia Cloud (hosted NATS): Personal free (10 connections, 10 GiB network data, 5 GiB storage); Starter $49/month (100 connections); Pro $199/month (1,000 connections, 1 TiB network data); Enterprise by quote. Overage from $0.01 to $0.09 per GiB; Remote System add-on $49/month to manage your own cluster.
Read our NATS review4.2/5 - 18
The framework to start an MCP App with today, as long as you are comfortable that its maker also sells the hosting it nudges you towards.
Best for: TypeScript teams building interactive apps for ChatGPT and Claude who want types and a fast local loop instead of the raw SDKs.
Price: Framework free under the MIT licence. Optional Alpic hosting: Free with 10,000 requests a month, Pro $50/month with 200,000 requests, Enterprise custom.
Read our Skybridge review4.2/5 - 19
Skyvern is the open browser agent most clearly built for back-office form work, though its licence and its cloud-only anti-bot layer push serious users towards paying.
Best for: Operations and RPA teams automating form filling, logins and downloads across many websites that have no API.
Price: Self-hosted code free under AGPL-3.0. Skyvern Cloud: Free $0 (5,000 credits, about 170 actions); Hobby $29/month (30,000 credits, about 1,200 actions); Pro $149/month (150,000 credits, about 6,200 actions); Enterprise custom.
Read our Skyvern review4.2/5 - 20
Strix is a capable, permissively licensed way to let AI agents test your own apps, provided you treat it as an active attacker on a budget and aim it only at systems you are authorised to test.
Best for: AppSec and platform teams with staging environments and written authorisation who want repeatable, agent-driven testing of their own web apps and APIs without per-test pentest fees.
Price: The CLI is free under Apache-2.0; you pay the model provider per scan. Strix Cloud Pro is $29 per seat per month with 50 PR reviews per developer, then $1 each; pentests are extra at $100 to $300 each, and a Full Audit starts at $2,000. Enterprise is custom-priced.
Read our Strix review4.2/5 - 21
Agno turns a Python agent into a served API with very little glue, but that API ships with authentication switched off.
Best for: Python teams that want to run production agents on their own infrastructure with minimal plumbing.
Price: Framework and AgentOS free under Apache 2.0. Control plane: free against a local AgentOS; Pro $150/month (one live connection, three seats, extra seats $30, SAML SSO $300); Enterprise custom.
Read our Agno review4.1/5 - 22
ARTEMIS is the free way to let a coding agent work a real Android phone, provided you treat it as an explorer and not as your regression suite.
Best for: Android teams who want their coding agent to reproduce bugs and explore new builds on a real device before anyone writes the regression test.
Price: Free under Apache-2.0. You pay for the vision-language model calls (Gemini, OpenAI, Anthropic, OpenRouter or xAI), billed per step by your provider, plus an optional Google Cloud Vision OCR key.
Read our ARTEMIS review4.1/5 - 23
Fast to a working multi-agent demo and harder to steer the further it travels from one, so production teams end up writing Flows rather than Crews.
Best for: Python developers who want a multi-agent workflow running quickly and can accept less low-level control.
Price: Free and open source under the MIT licence. CrewAI AMP: Basic free with 50 workflow executions a month; Enterprise (more executions, SSO, RBAC, VPC or on-prem deployment) by quote.
Read our CrewAI review4.1/5 - 24
The open deep-research agent with the longest track record, cheap per report and easy to embed, if you accept that its default stack runs on two paid APIs.
Best for: Developers building research or report generation into their own product, and analysts who want deep research over their own documents with a model they choose.
Price: Free and open source (Apache-2.0 on GitHub). You pay for the LLM and search APIs it calls.
Read our GPT Researcher review4.1/5 - 25
OpenWork is the open Cowork alternative most likely to survive contact with a real team, because it treats shared skills and admin policy as the product, though its open-core split means the team half is not MIT.
Best for: Teams leaving Claude Cowork, or standardising on one set of agent skills and connections across several tools, who want an open desktop client and can live with an open-core admin layer.
Price: Desktop app and core MIT, free with your own keys; self-hosted control plane free for up to 5 users. Team $10/user/month; Enterprise custom. Add-ons: Cloud Computer $50/member/month, managed OpenWork models $10/user/month.
Read our OpenWork review4.1/5 - 26
Rivet is the most direct way to give each agent its own durable process and database without paying for an idle sandbox, provided you can live with a 3.x line that changed under its users this month.
Best for: TypeScript teams building agents or collaborative apps that need durable per-user state and live connections without running a sandbox per session.
Price: Free to self-host under Apache-2.0. Rivet Cloud: Free $0, Hobby $18/month, Team $156/month, Enterprise $740/month, plus $18 to $327 per actor node per month.
Read our Rivet review4.1/5 - 27
For anyone with a program and a score to push up, this is the open alternative to AlphaEvolve to start with, provided they sandbox the code it writes.
Best for: Researchers and performance engineers who have a scored program, such as a heuristic, kernel or solver, and want LLMs to search for better versions within a fixed budget.
Price: Free under Apache 2.0. You pay for the LLM calls (or a Claude or ChatGPT subscription through Headless) and the compute that runs your evaluator.
Read our ShinkaEvolve review4.1/5 - 28
Giving the agent its own Linux desktop is a genuinely different idea, and Agent Zero builds it out well, as long as you lock down the container yourself.
Best for: Individual operators and small consultancies who want one general agent that can use a real desktop, browser and office files under their supervision.
Price: Free and open source (MIT); you pay for the model provider and wherever Docker runs.
Read our Agent Zero review4.0/5 - 29
Dapr pays off for a platform team already running many services on Kubernetes; for a single agent it is a lot of runtime to adopt for one workflow engine.
Best for: Platform teams on Kubernetes that want agents and ordinary services to share one model for durable workflows, messaging and secure service calls.
Price: Runtime and Dapr Agents free under Apache 2.0. Diagrid Catalyst: Cloud free tier; Cloud Plus $31/month billed annually ($372/year; 1M API requests/month, 100 concurrent workflows per project); Dedicated Cloud from $1,199/month billed annually; BYOC from $1,599/month; Enterprise Server (self-managed, air-gapped) custom.
Read our Dapr review4.0/5 - 30
nanobot is the personal agent to choose when you want to read the code you are trusting, smaller and easier to follow than OpenClaw, with a younger and spikier security record.
Best for: Python developers who want a self-hosted personal or team agent they can read, extend and call from their own code.
Price: Free and open source (MIT). You pay your model provider, or run a local model.
Read our nanobot review4.0/5 - 31
TradingAgents is the reference design for multi-agent stock analysis, and its recent work on look-ahead bias makes it one of the few you can backtest without fooling yourself.
Best for: Researchers and technical investors who want a well-documented multi-agent baseline for single-stock decisions and a fair way to backtest it.
Price: Free and open source (Apache-2.0). You pay your LLM provider or use a ChatGPT Plus or Pro plan; FRED and SEC EDGAR data are free, Alpha Vantage is optional.
Read our TradingAgents review4.0/5 - 32
Dograh is the most practical open-source answer to Vapi so far, provided you are ready to run a WebRTC server with dozens of open ports and to read what its quickstart switches on.
Best for: Engineering teams that want Vapi-style voice agents on their own servers, with call data and provider choice under their control.
Price: Free under BSD 2-Clause to self-host. Dograh Cloud charges a 1¢/min platform fee with 10 concurrent calls included and credits from $5, plus model costs unless you bring your own keys. Committed-volume and on-premise enterprise plans are priced on request.
Read our Dograh review3.9/5 - 33
Eigent is the most ambitious of the open Cowork clones, splitting work across parallel agents that can drive your desktop, but its shell and GUI access run with your full user rights and its security record is short and already includes a 1-click RCE.
Best for: Technical individuals and small teams who want parallel agents for research, reporting and file work on their own machine, and are willing to supervise them.
Price: Free and open source (Apache 2.0) with your own API keys or local models. Hosted credits: Plus $19.99/month (2,000 task credits), Pro $99.99/month (10,000 task credits), both with a 7-day trial; Teams announced; Enterprise custom.
Read our Eigent review3.9/5 - 34
Nod is the personal agent that treats your approval as a hard rule rather than a prompt, but it is a 0.8 research preview that sends app traffic through n8n's service by default.
Best for: People who want an agent working in their email, calendar and browser and care more about approvals and sandboxing than about reach across chat apps.
Price: Free and open source (Apache-2.0). You pay your model provider directly, or sign in with a ChatGPT Plus or Pro plan.
Read our Nod review3.9/5 - 35
OpenWorker puts the approval gates and audit trail where they belong, in the architecture, but it is a three-month-old beta whose roadmap is set by an internal list rather than its contributors.
Best for: Security engineers and individual knowledge workers who want a local agent that asks before it acts and keeps a record of who approved what.
Price: Free and open source (MIT). No paid plan; you pay your model provider, or nothing with a local model. An optional free account brokers OAuth for connectors.
Read our OpenWorker review3.9/5 - 36
Sellbird is the rare cold email agent that treats the law and the recipient as part of the job, which is why it is worth running and why autopilot is worth leaving off.
Best for: Founders and small sales teams selling to local or mid-sized businesses who want researched, personal cold email with a human approving what goes out.
Price: Free under AGPL-3.0. You supply a Claude Pro or Max plan or a ChatGPT plan with Codex, and a mailbox (about $7 a month for Google Workspace). Company search is free on OpenStreetMap or about $0.37 per 1,000 on the cheapest paid source. The hosted sellbird.ai has no published price.
Read our Sellbird review3.9/5 - 37
Steel wraps real Chrome in a sensible agent API and prices its cloud fairly, but the open-source build runs one session at a time and leaves the features people buy it for in the cloud.
Best for: Teams building browser agents that need logged-in, long-running Chrome sessions and would rather pay per browser hour than run a Chrome fleet.
Price: Free to self-host (Apache-2.0). Steel Cloud: Launch $0 plus usage, with $30 of one-time credits valid 90 days, browser hours at $0.10, proxies $10/GB, CAPTCHA solves $3 per 1,000, 10 concurrent sessions; Scale $250/month plus usage, with $100 of credits a month, $0.08 an hour, $6/GB, $1 per 1,000, 100 concurrent sessions; Enterprise custom.
Read our Steel review3.9/5 - 38
The reference project for automated ML research and now a usable harness for coding agents, but its papers still need a human to catch errors, and every run executes code an LLM wrote.
Best for: ML research groups studying or using automated experimentation who can sandbox agent runs and will review every result before it is published.
Price: Free. v1 and v2 under the AI Scientist Source Code License 1.0 (use restrictions and mandatory AI disclosure); v3 under Apache 2.0. You pay for LLM API calls or subscriptions and GPUs.
Read our The AI Scientist (v1, v2 and v3) review3.9/5 - 39
Vibe-Trading packs a quant desk's worth of backtesting and data plumbing behind a chat box, and its unusually careful changelog is the best reason to trust it more than the average trading bot.
Best for: Retail quants and finance researchers who want to test strategy ideas in plain language and audit how each result was produced.
Price: Free and open source (MIT). You pay your LLM provider; market data is free by default, with optional paid sources (QVeris, Finnhub, Tiingo and others) on your own keys.
Read our Vibe-Trading review3.9/5 - 40
AgentRun makes a persuasive case that most agent jobs should become checkable workflows, and it ships the approvals and caps to back it, but it is a 52-day-old beta whose cost claims come only from its makers.
Best for: Teams running the same agent job many times a week who want each step visible, costs capped and the routine cases moved off the expensive agent loop.
Price: Free under Apache-2.0; you pay your model provider and any remote sandbox vendor. The one-click Render deploy costs about $25/month for the instance plus $2.50 for the disk, per the README.
Read our AgentRun review3.8/5 - 41
Avibe is a thoughtful home for an always-on assistant on your own computer, with a better secrets design than most, but it is young and its security policy is still a stub.
Best for: Individuals and small teams who want a persistent assistant that works on their own files and accounts from one computer, reachable from a phone or chat app.
Price: Free and open source (MIT). You pay your own Claude or ChatGPT subscription or an API provider directly; Avibe adds no markup or second subscription.
Read our Avibe review3.8/5 - 42
BAML makes LLM output dependably typed, but today you are choosing between a legacy DSL and a new language that still breaks every few weeks.
Best for: Teams that extract typed data from LLMs at volume, often on cheaper models, and will own a code generation step and a fast-moving toolchain.
Price: Free and open source under Apache-2.0; the toolchain runs locally without an account. BAML Cloud (hosted observability and team controls) is announced but has no published pricing.
Read our BAML review3.8/5 - 43
A sound idea, that agents work better through a command line than through screenshots, delivered as a catalogue whose quality depends on who wrote each harness.
Best for: People who want an AI agent to operate desktop applications such as Blender, GIMP or LibreOffice from the command line.
Price: Free and open source under Apache 2.0. The wrapped applications and the agent you run it in are separate.
Read our CLI-Anything review3.8/5 - 44
Eleven turns Ghidra and REA into a workbench you share with an agent, a real gain in usability from a three-week-old app that one person builds.
Best for: Security researchers and developers who already reverse-engineer software they are authorised to examine and want an agent working in the same window.
Price: Free and open source (Apache 2.0), with optional GitHub Sponsors. You need your own agent subscription, and Hopper or IDA need their own licences if used.
Read our Eleven review3.8/5 - 45
Moli gets more real pages working than the other from-scratch headless browsers, on its maker's numbers, at a fraction of Chrome's memory, but it is three months old, mostly one developer's work and slower than Lightpanda.
Best for: Agent and evaluation teams whose tasks fail on lighter engines and who want to cut Chrome's memory without losing as many pages.
Price: Free under Apache-2.0 or MIT. Lexmount Browser (hosted, built on Moli) has a free allowance; prices not published.
Read our Moli review3.8/5 - 46
OpenBot gets agent governance right, with a fail-closed gateway in front of every action, but it is a beta template tied to CopilotKit's paid memory service rather than something you can install and run.
Best for: Platform teams that want self-hosted agents with per-action policy and audit, and are prepared to fork and maintain a template.
Price: MIT licence. Requires CopilotKit Intelligence: Free ($0) hosted tier with 1,000 interactions per organisation a month; Pay as you go $49/month with 10,000 interactions included; Annual custom. You supply the model key.
Read our OpenBot review3.8/5 - 47
OpenCLI is a fast way to give an agent your browser and your logins, which is exactly why it needs treating like a credential rather than a utility.
Best for: Developers and power users who want agents to read and act on sites they are already logged into, from a dedicated Chrome profile.
Price: Free and open source (Apache-2.0), including the Chrome extension and desktop app.
Read our OpenCLI review3.8/5 - 48
OpenResearch gives coding agents the experiment memory and compute plumbing that research actually needs, from a team whose business is papers, but it is young and phones home by default.
Best for: ML and computational-science researchers who already use a coding agent and want it to run, log and compare experiments across their own compute.
Price: Free and open source (MIT). Managed OpenResearch compute is billed at cost from a prepaid balance; rates are shown in the dashboard, not published on the site. Agent and model costs are your own.
Read our OpenResearch review3.8/5 - 49
Paperclip keeps a dozen agents on task and on budget better than any open-source rival, but its security record says keep it off the open internet.
Best for: Developers and small teams running three or more coding or business agents who need tasks, spend limits and review in one place.
Price: Free and open source (MIT), self-hosted. You pay for the agents' model usage; Paperclip Cloud pricing not published.
Read our Paperclip review3.8/5 - 50
A lean local browser service that saves agents real tokens, from a young project whose security record so far argues for keeping it strictly on localhost.
Best for: Developers who want one local, token-efficient Chrome service that all their coding agents and scripts can share.
Price: Free and open source (MIT).
Read our PinchTab review3.8/5 - 51
A clean, minimal implementation of a respected search method that does exactly one job, and leaves the hard parts, scoring and choosing a final answer, to you.
Best for: Engineers who already have a verifier for model outputs and want a tested search algorithm to turn a bigger call budget into better answers.
Price: Free under Apache 2.0. Costs are the LLM calls your generation and scoring functions make.
Read our TreeQuest review3.8/5 - 52
A polished open client with a closed cloud behind it: useful if you want your agents on every device, less so if you wanted to own the whole stack.
Best for: Individuals who already pay for Claude Code or Codex and want the same agent sessions on their desktop, laptop and phone.
Price: Client free with your own keys or plans. Managed models: Plus $20, Super $100, Ultra $200 of credit a month ($16, $80, $160 on a recurring plan); top-ups from $10.
Read our Cindy review3.7/5 - 53
DS Bot is the friendliest way yet to keep a standing team of DeepSeek agents on your desktop, but it is a preview build on top of a preview product, with a very powerful bot in charge.
Best for: DeepSeek Harness users who want a few persistent assistants with memory and group chats on their own computer, and are comfortable with preview software.
Price: Free under Apache-2.0. Model calls are billed by your provider; with DeepSeek's API that is pay-as-you-go, and group chats cost more than single bots.
Read our DS Bot review3.7/5 - 54
Letta still has the most considered ideas about agent memory, but rebuilding the product mid-flight left its earlier users to migrate and newcomers with a fast-moving target.
Best for: Developers who want long-running personal or team agents with inspectable, versioned memory and are happy to adopt a whole harness.
Price: Apache 2.0, free to run locally with your own model keys. Letta Cloud: Free (3 agents, 5 MiB per memory repository); Pro $20/month (20 agents, 50 MiB, model quota); API plan $20/month plus $0.10 per active agent a month and $0.00015 per second of tool execution, tokens extra; Teams Pro $20/seat/month; Enterprise custom.
Read our Letta review3.7/5 - 55
SwarmChat is a thoughtfully engineered message bus for agent swarms that takes identity and permissions more seriously than most version 0.1 research code, but it is a research tool first and its security rests on how you deploy it.
Best for: Researchers and engineers running teams of Claude Code, Codex or custom agents who want them to coordinate over channels and DMs and want to watch and replay every exchange.
Price: Free and open source (MIT). No hosted service; you run the server and pay only for the agents' own model usage.
Read our SwarmChat review3.7/5 - 56
WebBrain gives you a capable agent inside your real browser for nothing, but its hosted default keeps your prompts for training until you find the switch, so run it on your own key or a local model.
Best for: Individuals who want a side-panel browser agent on their own API key or a local model and will take ten minutes to audit its defaults.
Price: Extension free under GPL-3.0-or-later. WebBrain Compass, the default hosted model, has a free daily allowance (size not published) and a $5/month subscription per device profile under a fair-use policy. Your own API key or a local model costs only what that provider or your hardware costs.
Read our WebBrain review3.7/5 - 57
AdalFlow's automatic prompt tuning is a serious idea with a paper behind it, but the library around it has fewer users, thinner release notes and a patchier release history than the frameworks it competes with.
Best for: Python teams with labelled examples who want prompts in a multi-step LLM pipeline tuned by measurement rather than by hand.
Price: Free and open source (MIT). Optimisation runs call your own model providers, typically a strong teacher model, so training costs tokens.
Read our AdalFlow review3.6/5 - 58
A capable free attack surface scanner with unusually honest docs, but it is one week old and runs non-GET Nuclei templates by default, so keep it to assets you own and a model you host.
Best for: Security engineers and bug bounty hunters who want a free, self-hosted view of which exploited CVEs touch their own internet-facing hosts, and who can run a local model and read a fast-moving changelog.
Price: Free and open source under GPL-3.0, with no seat or asset limits. You pay for the server, any hosted model tokens and optional paid data keys such as WPScan.
Read our cvENGINE review3.6/5 - 59
Evenfire puts serious guardrails around agents that touch production systems, at the cost of running a Kubernetes platform that is still in beta.
Best for: Platform teams already running Kubernetes who need several agents with real system access, shared across users, under approvals, budgets and audit logs.
Price: Free under MPL-2.0 for self-hosting, one organisation per deployment. Hosted multi-tenant service with support and SLA; price not published.
Read our Evenfire review3.6/5 - 60
Holographic is the careful, offline choice for Hermes users who want a fact store they can audit, but it is a one-maintainer plug-in with no standard benchmark and little use outside Hermes.
Best for: Hermes Agent users who want inspectable, offline fact memory and are willing to have the agent write facts deliberately.
Price: Free and open source under the MIT licence. Optional search by meaning and reranking call endpoints you choose and pay for.
Read our Holographic review3.6/5 - 61
JiuwenSwarm is the serious multi-agent choice for teams that work in Feishu, DingTalk or WeCom, but the config it ships leaves shell and file tools on allow, whatever the README says.
Best for: Engineering teams on Feishu, DingTalk or WeCom who want several agents to split long research or coding jobs and will lock down permissions themselves.
Price: Free and open source (Apache-2.0). You pay your model provider, or run a local model.
Read our JiuwenSwarm review3.6/5 - 62
LangAlpha is the most carefully engineered open agent for stock research we have read, and its self-hosted mode is too loosely secured to point at a live brokerage account.
Best for: Analysts and technically minded investors who want an agent to keep long-running equity research in one place, with any live trading kept behind approvals.
Price: Self-hosted free (Apache-2.0); hosted Free 1,000 credits/month, Plus $20/month, Pro $40/month, Max $100 or $200/month; 15% off yearly; your own model keys use no credits.
Read our LangAlpha review3.6/5 - 63
A clean reference for tool retrieval that does one job well, from a LangChain repo that has had no feature work since its third release in June 2025.
Best for: Python teams on LangGraph whose agents have outgrown the number of tools a model can sensibly see at once.
Price: Free and open source under the MIT licence; you pay for the embedding and chat model calls.
Read our LangGraph Bigtool review3.6/5 - 64
MiroFish is the easiest way to watch a crowd of LLM personas react to a document, but nothing yet shows its predictions beat a thoughtful analyst.
Best for: Comms, policy and research teams who want to stress-test how a public might react to a draft, and who will read the output as a scenario rather than a forecast.
Price: Free and open source (AGPL-3.0). You pay for LLM tokens and for Zep Cloud memory beyond its free 10,000 credits a month; Zep Flex is $125/month.
Read our MiroFish review3.6/5 - 65
OMA's approvals, journal and egress controls are well thought through, but a few hundred weekly installs behind 7,000 stars says it has more admirers than users.
Best for: TypeScript teams that must show who approved each agent action and where model traffic went, and can live with a young, single-company project.
Price: Free and open source (MIT). No hosted service; the maintainer sells custom builds on request.
Read our Open Multi-Agent (OMA) review3.6/5 - 66
A thoughtful control room for a team of CLI agents, let down by full-access defaults and a licence that now charges for work use.
Best for: Individuals running several CLI agents on a dedicated machine who want them to work as a coordinated team, and businesses willing to buy a licence.
Price: Free for personal, noncommercial use. Commercial licence from €10 a month (Starter, 1 member) or €70 a month (Team, 2 members), free until 1 December 2026; hosted servers €20, €50 or €100 a month.
Read our OpenBot (Nightly Labs) review3.6/5 - 67
OpenMuse is a carefully documented blueprint for an open Meta Muse, but it is a developer's template, it cannot run without CopilotKit's thread service, and it reports your usage to CopilotKit until you switch that off.
Best for: Developers who want to build their own Muse-style personal agent on a phone and web app they control, and who accept CopilotKit Intelligence as part of the stack.
Price: MIT licence. Requires CopilotKit Intelligence: Free ($0) hosted tier with 1,000 interactions per organisation a month; Pay as you go $49/month with 10,000 interactions included; Annual custom. You supply the model key and the hosting.
Read our OpenMuse review3.6/5 - 68
An ambitious agent office for one desktop, open in its code but with the collaboration features kept behind the packaged app.
Best for: Solo founders and small teams who want research, content and store tasks handled by agents from one desktop app, and already use Claude Code or Codex for the code.
Price: Source build free (MIT) with your own API keys. Packaged app: Free with 500 lifetime credits; Lite $5.99/month (1,000 credits); Pro $16.99/month (5,000 credits); Max $29.99/month (10,000 credits); Enterprise custom. Own-key usage does not consume credits.
Read our Orkas review3.6/5 - 69
AutoGPT is now a commercial agent platform with a famous repository attached, and its licence, its open sign-up default and a heavy 2026 run of security advisories all argue for reading the small print first.
Best for: Small business teams that want a hosted agent for recurring content, sales and admin work and accept a source-available licence.
Price: Self-hosting free (platform under PolyForm Shield, the rest MIT); you pay for infrastructure and model usage. Hosted Pro $50/month with a 7-day trial, Max $320/month; workflow runs billed from a separate credit wallet. Team plan not yet priced.
Read our AutoGPT review3.5/5 - 70
Dodaios packs voice, coding-agent panes, desktop control and memory into one local app with unusually honest docs, but it is a five-month-old, one-maintainer project whose safety gate still lets a lot through.
Best for: Developers who want one local, voice-driven cockpit for several coding agents and accept a young project's rough safety edges.
Price: Free and open source (Apache 2.0), with no account or paid tier. You pay your model provider, or reuse a Claude or ChatGPT subscription or a local model.
Read our Dodaios review3.5/5 - 71
Interpreter Workstation has the most carefully argued security and distribution design of the open Cowork clones, but it is three months old, has 44 stars of its own and puts a computer-use driver and browser bridge on your desk from day one.
Best for: Open Interpreter users and organisations that want to build their own branded desktop agent from an auditable open source tree.
Price: Free and open source (Apache 2.0); bring your own provider key or a local model. Official builds offer optional Interpreter-hosted services; their prices are not published in the repository.
Read our Interpreter Workstation review3.5/5 - 72
A clear demonstration that agents can act in code, now drifting into maintenance mode with 590 pull requests waiting.
Best for: Researchers and learners exploring code-writing agents, especially with open models.
Price: Free and open source under the Apache 2.0 license.
Read our smolagents review3.5/5 - 73
UFO shows how far deep Windows integration can push a desktop agent, but it still fails most benchmark tasks while driving your real mouse and keyboard.
Best for: Automation engineers in Windows-heavy organisations who want to prototype agents for desktop software with no API, inside a test virtual machine.
Price: Free and open source under the MIT licence; you pay your model provider for every step.
Read our UFO review3.5/5 - 74
MagenticLite is a worthwhile look at what small open models can do as a browser agent, but it is a research prototype that needs two GPU endpoints and has gone quiet since July.
Best for: Researchers and platform engineers evaluating small open models for computer-use agents on hardware they control.
Price: Free under MIT, including the MagenticBrain and Fara1.5 weights. You host the two models yourself or on a GPU cloud.
Read our Magentic-UI (MagenticLite) review3.4/5 - 75
A friendly front end for DeepSeek Harness that gets Office work and scheduled jobs onto a non-technical desktop, resting on a runtime that is itself still a release candidate.
Best for: Office workers and small teams who want a free, open desktop agent for documents and recurring jobs and already pay for a model API.
Price: Free and open source (Apache-2.0). No paid tier; you pay your model provider directly, and built-in web search costs nothing.
Read our PawWork review3.4/5 - 76
OpenCompany has one of the more thought-through designs for agents that work as a team, but it is a self-declared work in progress that leans on its maker's paid key to do anything useful.
Best for: Developers and solo founders curious about deliberating agent teams who are happy to pay TinyHumans for orchestration while they experiment.
Price: Software free under GPL-3.0, but agents need a TinyHumans API key to do real work. TinyHumans Basic $20/month ($21 credits, 1 company, 3 users); Pro $200/month ($210 credits, 5 companies); pay-as-you-go credits from $5; Enterprise custom. Your own Claude Code, Codex or OpenRouter accounts can be used, billed by them.
Read our OpenCompany review3.3/5 - 77
A real idea with headline numbers that compare against the wrong baseline, maintained by two people and installed by almost nobody.
Best for: LangGraph teams with large, frequently checkpointed state who want to measure whether serialisation is their bottleneck.
Price: Free and open source under the MIT licence.
Read our Fast-LangGraph review3.0/5 - 78
OpenComputer is a tidy, honestly documented experiment in fully local computer use, but one commit, no releases and 31 stars make it a project to read rather than one to trust with your desktop.
Best for: Developers who want to study or extend a small, fully local computer-use agent on a Linux X11 machine with a 16 GB GPU.
Price: Free and open source (MIT). No paid plan; you supply the model, locally on your own GPU or through any compatible hosted API.
Read our OpenComputer review2.9/5 - 79
A capable one-person agent desktop with real scheduling and memory, but it is seven weeks old, has no outside users yet and runs on an Electron release a year out of support.
Best for: Windows power users comfortable reading TypeScript who want a free local host for several scheduled agents and accept an early single-maintainer project.
Price: Free and open source (MIT). Model costs are yours: Cursor SDK runs bill your Cursor plan (Plus $8/month covers Cursor and Grok models; other models at API list price), or your own provider key is billed by that provider.
Read our TecAdRiseBot review2.9/5 - 80
AutoGen is archived and unmaintained, so it belongs in legacy code you are migrating and in the history of agent frameworks, not in a new project.
Best for: Teams keeping existing AutoGen code alive while they migrate, and researchers studying multi-agent patterns.
Price: Free and open source: code under the MIT licence, documentation under CC BY 4.0. You pay only for the models you call.
Read our AutoGen review2.6/5
Sources
The main official source behind each ranked review; every review lists its full set.
- LangGraph: langchain-ai/langgraph on GitHub
- Browser Use: browser-use/browser-use on GitHub (README)
- Pydantic AI: pydantic/pydantic-ai on GitHub
- Temporal: temporalio/temporal on GitHub
- Mastra: mastra-ai/mastra on GitHub
- MCP Python SDK: modelcontextprotocol/python-sdk on GitHub
- OpenAI Agents SDK: openai/openai-agents-python on GitHub
- Hermes Agent: NousResearch/hermes-agent on GitHub
- OpenClaw: openclaw/openclaw on GitHub
- Stagehand: browserbase/stagehand on GitHub
- Claude Agent SDK: anthropics/claude-agent-sdk-python on GitHub
- DeerFlow: bytedance/deer-flow on GitHub
- Google Agent Development Kit (ADK): google/adk-python on GitHub
- LangChain: langchain-ai/langchain on GitHub
- Lightpanda: lightpanda-io/browser on GitHub
- Microsoft Agent Framework: microsoft/agent-framework on GitHub
- NATS: nats-server on GitHub
- Skybridge: alpic-ai/skybridge on GitHub
- Skyvern: Skyvern-AI/skyvern on GitHub (README)
- Strix: usestrix/strix on GitHub
- Agno: agno-agi/agno on GitHub
- ARTEMIS: google/artemis on GitHub
- CrewAI: crewAIInc/crewAI on GitHub
- GPT Researcher: assafelovic/gpt-researcher on GitHub
- OpenWork: different-ai/openwork on GitHub
- Rivet: rivet-dev/rivet on GitHub
- ShinkaEvolve: GitHub: SakanaAI/ShinkaEvolve
- Agent Zero: agent0ai/agent-zero on GitHub
- Dapr: dapr/dapr on GitHub
- nanobot: HKUDS/nanobot on GitHub
- TradingAgents: TauricResearch/TradingAgents on GitHub
- Dograh: Dograh on GitHub
- Eigent: eigent-ai/eigent on GitHub
- Nod: n8n-io/nod on GitHub
- OpenWorker: andrewyng/openworker on GitHub
- Sellbird: ethanplusai/sellbird on GitHub
- Steel: steel-dev/steel-browser on GitHub
- The AI Scientist (v1, v2 and v3): GitHub: SakanaAI/AI-Scientist
- Vibe-Trading: HKUDS/Vibe-Trading on GitHub
- AgentRun: Parcha-ai/agentrun on GitHub
- Avibe: avibe-bot/avibe on GitHub
- BAML: BoundaryML/baml on GitHub
- CLI-Anything: CLI-Anything on GitHub
- Eleven: Zfinix/eleven on GitHub
- Moli: lexmount/moli on GitHub (README)
- OpenBot: CopilotKit/OpenBot on GitHub
- OpenCLI: jackwener/OpenCLI on GitHub
- OpenResearch: alphaXiv/OpenResearch on GitHub
- Paperclip: Paperclip on GitHub
- PinchTab: pinchtab/pinchtab on GitHub
- TreeQuest: GitHub: SakanaAI/treequest
- Cindy: makecindy/cindy on GitHub
- DS Bot: Fei2411/DS-Bot on GitHub
- Letta: letta-ai/letta-code on GitHub
- SwarmChat: Pranjal2041/swarmchat on GitHub
- WebBrain: webbrain-one/webbrain on GitHub (README)
- AdalFlow: SylphAI-Inc/AdalFlow on GitHub
- cvENGINE: hevnsnt/CVEngine on GitHub
- Evenfire: evenfire-ai/evenfire on GitHub
- Holographic: itpartypattaya/hermes-plugin-holographic on GitHub
- JiuwenSwarm: openJiuwen-ai/jiuwenswarm on GitHub
- LangAlpha: ginlix-ai/langalpha on GitHub
- LangGraph Bigtool: langchain-ai/langgraph-bigtool on GitHub
- MiroFish: 666ghj/MiroFish on GitHub
- Open Multi-Agent (OMA): open-multi-agent/open-multi-agent on GitHub
- OpenBot (Nightly Labs): nightly-labs/openbot on GitHub
- OpenMuse: CopilotKit/openmuse on GitHub
- Orkas: Orkas-AI/Orkas on GitHub
- AutoGPT: Significant-Gravitas/AutoGPT on GitHub
- Dodaios: Dodaios on GitHub
- Interpreter Workstation: openinterpreter/interpreter-workstation on GitHub
- smolagents: huggingface/smolagents on GitHub
- UFO: microsoft/UFO on GitHub
- Magentic-UI (MagenticLite): microsoft/magentic-ui on GitHub (README)
- PawWork: Astro-Han/pawwork on GitHub
- OpenCompany: tinyhumansai/opencompany on GitHub
- Fast-LangGraph: neul-labs/fast-langgraph on GitHub
- OpenComputer: andykr1k/OpenComputer on GitHub
- TecAdRiseBot: tecadrise-ai/TecAdRiseBot on GitHub
- AutoGen: microsoft/autogen README on GitHub
More guides
- Best open-source tools to run LLMs locally
- Best open-source RAG tools
- Best open-source LLM fine-tuning tools
- Best open-source AI chat apps
- Best open-source AI coding agents
- Best open-source image and video AI tools
- Best open-source LLM evaluation tools
- Best open-source LLM gateways and observability tools
- Best Claude Code plugins and agent skills
- Best open-source tools for coding agents
- Best free courses for learning AI