Skip to content
The Agents Index

Buyer's guide Agent platforms

Agent platforms that pause for your approval

OpenClaw and Letta gate agent actions with policy modes, n8n pauses each tool for review, and Letta's paid plan starts at $20 a month.

The Agents IndexPublished Updated 19 min readEditorial policy
Logos of OpenClaw, Salesforce Agentforce, Letta, n8n, Dify, Botpress, GC AI, Glean. 8 agents compared, from $20 a month.
On this page
  1. The short answer
  2. Pick by what you need
  3. 8 agents, side by side
  4. The picks
  5. How we ranked
  6. Questions buyers ask
  7. Sources

The short answer

OpenClaw leads with an exec policy that runs from deny through ask to full, with a deny fallback when no approval screen is reachable. Letta follows with four modes from strict to unrestricted plus allow and deny tool patterns, and GC AI asks Approve, Always approve or Deny per action type.

Full permission controls come with 4 of the 8, with n8n pausing at each tool that has a review step. Salesforce Agentforce, Dify, Botpress and Glean pause at fixed points instead, such as a Human Input node in Dify or an admin opt-in for write tools in Glean.

On 26 October 2026 we listed 192 agents: 34 of them under Agent platforms and 104 with a free plan.

  • Full agent permission controls come with 4 of the 8: OpenClaw, Letta, n8n and GC AI.
  • Letta Pro costs $20 a month for up to 20 agents.
  • n8n pauses a workflow at each tool with a review step, and tools without one run unattended.
  • Glean is sized for large buyers, with typical enterprise minimums around 100–250 seats.

8 agents, side by side

Table 1 Price, free plan and the questions that separate them, checked 24 October 2026
ProductFromFree planAgent permissionsIn your editorOpen modelsBest for
Free, as of 21 September 2026
Yes; 3-day trial
Yes
Yes
Yes
self-hosters who want exec policy modes
Free, then $2 per outcome
Yes
Partly
Yes
Partly
Salesforce teams acting on CRM records
Free, then $20 a month
Yes
Yes
Yes
Yes
stateful agents with tool allow rules
about $27 a month
Yes
Yes
Partly
Yes
per-tool human approval in workflows
Free, then $59 a month
Yes
Partly
Partly
–
workflows that stop at a human step
Free, then $189 a month
Yes
Yes
Partly
Partly
support agents with human-approved replies
$500 a seat a month
14-day trial
Yes
Partly
No
legal teams gating actions by type
On quote
–
Partly
Yes
Yes
enterprises with admin-approved write tools

How we check

The picks

1. OpenClaw

Best for self-hosters who want exec policy modes: Its exec policy runs from deny through allowlist, ask and auto to full, with a deny fallback when no approval screen is reachable.

Price Free, as of 21 September 2026

OpenClaw is an open-source agent runtime you host yourself and message from WhatsApp, Telegram, Slack, Signal, iMessage and 29 chat channels in total. It writes scripts, runs code and installs packages from its own command line. A cron-like heartbeat wakes it to check its task queue and message you first.

Exec policy has deny, allowlist, ask, auto and full modes with approval prompts and allowlist bindings. An audit ledger records runs, tool actions and approval decisions as metadata, and richer execution-identity recording is off by default. Single sign-on (SSO) comes through an identity-aware reverse proxy such as Pomerium or oauth2-proxy rather than a built-in SAML login.

It has a free plan. You bring your own model keys or subscription sign-in, or run models on your own hardware. Pick it over Letta for one personal agent on chat channels rather than many stateful agents behind an API.

What it does well

  • Exec modes from deny to full let you choose how much runs unattended, with approval prompts and allowlist bindings.
  • Reaches 29 chat channels, including WhatsApp, Slack and iMessage.
  • Connects to 64 model and media providers, including local runtimes such as Ollama and vLLM.
  • Works as an MCP client and as an MCP server.

Where it falls short

  • One reviewer’s setup took about 6 hours on a rented server, with no polished installer.
  • No built-in task view, and the reviewer saw it drop some assigned tasks.
  • The audit ledger holds metadata, and execution-identity recording stays off until you enable it.
  • Free plan: yes; 3-day free trial
  • Who it’s for: Developers who already run their own servers and want one agent they message from chat apps, under exec rules they set.
  • Skip it if: Non-technical users should pick something else, since one reviewer’s setup took about 6 hours on a rented server.

2. Salesforce Agentforce

Best for Salesforce teams acting on CRM records: Agents act on CRM records, and their actions and outputs are logged.

Price Free, then $2 per outcome

Agentforce builds, runs and monitors AI agents on Salesforce CRM data. Agentforce Builder offers natural language, a visual canvas, simulation and batch testing, and Agent Script adds deterministic conditions, loops and variables next to LLM reasoning. Observability shows sessions, outcomes and credit use per agent.

An approval before every action is documented for the Agentforce Vibes coding agent alone. Agent actions and outputs are logged, and event-level logging with streaming to a security monitoring tool is the paid Salesforce Shield add-on.

Salesforce Foundations is free and includes Agentforce Builder and Agent Script. Paid use is $5 a seat a month, Flex credits at $500 per 100,000 credits, or Conversations at $2 per customer-facing conversation. Flex Credits and Conversations cannot be used in the same org.

What it does well

  • SSO over SAML or OpenID Connect, with the org as identity provider, service provider or both.
  • Agentforce is covered by SOC 2, ISO 27001 and ISO 42001 audits.
  • Salesforce-hosted MCP servers let outside agents query an org under per-user OAuth.
  • The free Foundations entry includes Agentforce Builder and Agent Script.

Where it falls short

  • An approval before every action is documented for the Agentforce Vibes coding agent alone.
  • Flex Credits cost $500 per 100,000 credits, and unused credits do not roll over.
  • Cross-system context needs connected data, Salesforce administration and governance work beyond the credits.
  • Free plan: yes
  • Who it’s for: Salesforce service, sales and field-service teams with admin or developer capacity who need agents to act on CRM records.
  • Skip it if: Teams outside Salesforce, or buyers who want a flat seat price, should look elsewhere, since paid use runs on metered credits or $125 to $550 per-user monthly options.

3. Letta

Best for stateful agents with tool allow rules: It offers four modes from strict to unrestricted, plus allow and deny tool patterns and a cross-agent memory guard.

Price Free, then $20 a month

Letta keeps each agent’s memory blocks, skills and configuration as a durable server-side object across sessions. Agents rewrite their own memory at runtime, and git tracks every change so memory can be diffed and rolled back. The active harness is TypeScript, letta-ai/letta-code, under Apache 2.0.

Permissions come in four modes from strict to unrestricted, with allow and deny tool patterns. No administrative audit log is published, and no SOC 2 or ISO 27001 attestation appears.

The Free tier covers 3 stateful agents with your own keys. Pro is $20 a month for up to 20 agents. The API Plan is $20 a month base plus $0.10 per active agent per month and $0.00015 per second of tool execution. Every plan accepts your own provider keys, and that usage bypasses Letta credits.

What it does well

  • Allow and deny tool patterns on top of four permission modes.
  • Pro is $20 a month for up to 20 agents.
  • Memory changes are tracked in git, so they can be branched, diffed and rolled back.
  • Runs as a custom agent in Zed and JetBrains IDEs through the letta-acp adapter.

Where it falls short

  • No administrative audit log is published.
  • No Letta MCP server is published, since MCP runs inbound alone.
  • Teams on the original Python repo must migrate to the TypeScript harness.
  • Free plan: yes
  • Who it’s for: Developers who need agents whose memory, skills and identity persist across long-running sessions, with tool rules on each agent.
  • Skip it if: Teams that need a stateless workflow graph for a bounded task should use a framework with a larger ecosystem.

4. n8n

Best for per-tool human approval in workflows: Approval is wired per tool: the workflow pauses and a reviewer approves or denies through Chat, Slack, Discord, Telegram, Gmail or Outlook.

Price From about $27 a month

n8n builds agent automations on a visual canvas, and every step of the agent’s reasoning is traceable. It connects to 500+ integrations, with custom HTTP for anything else. JavaScript or Python steps can sit anywhere in a workflow.

A tool with a review step pauses the workflow until a reviewer approves or denies, and a tool without one runs unattended. An instance-level MCP server exposes chosen workflows to outside clients, and the MCP Client Tool calls other servers.

It has a free plan and runs on your own infrastructure, keeping data, credentials and execution in-house. Users are unlimited on every plan and the quota is executions.

What it does well

  • Per-tool approval through Slack, Gmail, Outlook or chat.
  • Unlimited users on every plan, metered by executions.
  • SOC 2 audited, with a pre-signed data processing agreement for GDPR.
  • Connects to 500+ integrations.

Where it falls short

  • The fair-code license is source-available rather than open source, so check it before commercial resale.
  • Tools with no review step execute unattended.
  • Free plan: yes
  • Who it’s for: Ops and automation teams who build agent workflows visually and keep data on their own infrastructure.
  • Skip it if: Developers who express complex agent control in code should pick a code-first framework, since complex logic gets unwieldy on a canvas.

5. Dify

Best for workflows that stop at a human step: A Human Input node pauses a run for review, approval or edits, with a timeout branch.

Price Free, then $59 a month

Dify wires model calls, tools, retrieval, logic and human-input nodes into an agent app on a canvas. Its retrieval-augmented generation (RAG) pipeline ingests files, sites, docs and drives, then chunks and indexes them. Apps ship as a hosted web app, REST API, embed or MCP tool.

There is no per-action permission prompt. Control comes from the Human Input node, which pauses a run for review. Keys you add bill to your own provider account.

Sandbox and the self-hosted Community edition are free. Seats are a plan ceiling: 1 on Sandbox, 3 on Professional and 50 on Team.

What it does well

  • A Human Input node pauses a run for review, approval or edits.
  • A built-in retrieval-augmented generation (RAG) pipeline ingests, chunks and indexes files, sites and drives.
  • Any app can be exposed as an MCP server from its Access Point tab.
  • Logs, feedback and usage analytics, with Langfuse and Arize Phoenix integrations.

Where it falls short

  • Self-hosting Community means you own upgrades, scaling and ops.
  • The license adds conditions to Apache-2.0, such as multi-tenant and branding limits.
  • Professional caps seats at 3.
  • Free plan: yes
  • Who it’s for: Developers and teams who want an open-source, self-hostable platform for document-grounded apps and agents with full data control.
  • Skip it if: Non-technical operators who want a fully managed agent with zero infrastructure should pick something else.

6. Botpress

Best for support agents with human-approved replies: A human can approve AI output before it is sent, and an Autonomous Node cannot read or write variables until access is granted.

Price Free, then $189 a month

Botpress pairs Agent Studio, a visual flow builder with custom JavaScript, API calls and webhooks, with a code-first TypeScript agent development kit and a helpdesk. Its LLMz engine plans and runs multi-step actions in place of a fixed decision tree. Knowledge Bases index PDFs, websites, Notion, Google Drive and Confluence.

Approval covers output. No approval gate exists before a tool call executes. Helpdesk operations are exposed as MCP tools for Claude Code, Claude Desktop and Cursor, and no SAML single sign-on is offered.

It has a free plan. The cheapest paid managed-cloud tier is $150 a month (Plus), with a 250-conversation cap before overages of $65 per 100 conversations. Teams with serious support volume sit on Team ($750 a month) or Enterprise.

What it does well

  • A human can approve AI output before it is sent.
  • One agent definition deploys to Webchat, WhatsApp Business, Slack, Microsoft Teams and Telegram.
  • MCP in both directions: the agent consumes outside MCP servers and exposes its own.
  • SOC 2 and GDPR are claimed, with outside penetration testing.

Where it falls short

  • No approval gate exists before a tool call executes.
  • No SAML single sign-on.
  • The Twilio voice channel is Enterprise-only, so Plus and Team get no voice.
  • Free plan: yes
  • Who it’s for: Engineering-led teams building one support agent for Webchat, WhatsApp Business, Slack and Microsoft Teams.
  • Skip it if: Non-technical teams who want a no-code chatbot builder should skip it, since wiring tool calls needs at least one developer.

7. GC AI

Best for legal teams gating actions by type: Permissions are set per action type, so searching an inbox and sending mail are separate grants, and an in-chat card offers Approve, Always approve or Deny before anything leaves.

Price From $500 a seat a month

GC AI runs research agents in parallel across jurisdictions, agencies and courts, then reconciles the findings into one cited analysis. It searches 13M+ US court opinions, statutes, regulations and agency guidance. A Microsoft Word add-in brings drafting and research into Word.

Agent Connectors link 20+ work apps such as Gmail, Outlook, Slack and HubSpot, with approval per action. Its remote MCP server exposes twenty tools over OAuth and stays off until an org admin turns it on.

Individual is $500 a month with a 14-day free trial. Case Law research is included on Team and Enterprise but is a paid add-on at Individual. The public API allows 60 requests per minute for inference endpoints.

What it does well

  • Separate grants per action type, with Approve, Always approve or Deny before anything leaves.
  • SOC 2 Type II and SOC 3 are claimed, with a data processing agreement in the standard terms.
  • Searches 13M+ US case-law and statute records, with source links on every finding.
  • A Microsoft Word add-in puts drafting and research inside Word.

Where it falls short

  • Priced at $500 a seat a month.
  • US Case Law research is a paid add-on on the Individual plan.
  • The public API is limited to 60 requests per minute for inference endpoints before general release.
  • Free trial: 14 days
  • Who it’s for: In-house legal teams and corporate counsel who need cited research across US case law, statutes and regulations.
  • Skip it if: Solo practitioners and small firms should look elsewhere, since at $500 a seat a month it is priced for in-house legal budgets.

8. Glean

Best for enterprises with admin-approved write tools: Write tools ask for confirmation unless an admin marks them eligible and the agent builder enables that step, so unattended writes need two opt-ins.

Price On quote

Glean indexes 100+ connected workplace systems and scopes every result to what the user may already see. Glean Assistant answers plain-language questions with cited sources. Glean Agents is a no-code builder, and Zillow has built 3,400+ internal agents with it.

An admin must mark a write tool eligible and the agent builder must enable it for the step before it runs unattended. Admin audit logs record configuration and administrative actions with CSV export, but exclude end-user activity. SSO over SAML is required.

A remote MCP server exposes search, chat and agents to 22 documented hosts with OAuth 2.1. You can use Glean’s Universal Model Key or bring your own provider keys. Pick it over n8n or Dify when agents mostly read company knowledge rather than run workflows.

What it does well

  • Unattended write access needs an admin and an agent builder to opt in.
  • Search results are scoped to what each user is already authorized to see.
  • SOC 2 Type II, ISO 27001 and HIPAA are claimed.
  • Plugs into Cursor, VS Code and Windsurf through its remote MCP server.

Where it falls short

  • Chat answers still occasionally hallucinate, and Glean needs its own separate login.
  • There is no built-in way to author documents or take structured notes.

“Yes. All plans support bringing your own API keys (BYOK). When you connect your own keys (via /connect in the Letta CLI), usage goes directly through your provider account”

From Letta’s own page, read on 2 September 2026
  • Who it’s for: Large and mid-market enterprises whose knowledge sits across 100+ internal tools and who want staff building agents on that index.
  • Skip it if: Teams that want a single-purpose customer-facing support agent should pick a support platform instead.

How we ranked

We compared 8 agents on price, free plan and 3 questions a buyer asks, reading each vendor’s pricing page and product pages between 2 September 2026 and 24 October 2026. We did not sign up for or run any of them. The list runs from the lowest starting price up, and agents without a public price come last. In all we answered 37 questions about what they do.

Questions buyers ask

Which agent platforms ask before an agent runs a command?

OpenClaw, Letta, n8n and GC AI let you set full agent permissions, which is 4 of the 8. OpenClaw’s exec policy has deny, allowlist, ask, auto and full modes with approval prompts. Letta has four modes from strict to unrestricted, plus allow and deny tool patterns.

Can a workflow pause until a person approves the next step?

Yes, in n8n and Dify. n8n pauses the workflow at a tool with a review step until a reviewer approves or denies. Dify’s Human Input node pauses a run for review, approval or edits, with a timeout branch.

Which of these can I self-host?

Self-hosting is offered by 4 of the 8, including OpenClaw, Letta and Dify. OpenClaw runs self-hosted alone, on your own machine, a rented server, a container or Kubernetes. Letta’s local runtime keeps all state on-device, or you run the App Server via Docker, Railway or Fly.io. Dify Community Edition runs anywhere Docker runs.

How much does a paid agent platform plan cost?

Letta Pro is $20 a month, and GC AI Individual is $500 a month. Letta’s API Plan adds $0.10 per active agent per month to a $20 a month base. GC AI Individual comes with a 14-day free trial.

Do these platforms have a free plan?

Yes, 6 of the 8 have a free plan, and for GC AI and Glean it is unconfirmed. GC AI Individual comes with a 14-day free trial.

How much does Glean cost?

Glean is sold to large and mid-market enterprises, and typical deals carry minimums around 100–250 seats. You can run it on Glean’s Universal Model Key or bring your own provider keys.

Can I bring my own model API key?

Yes, OpenClaw, Letta and Glean all accept your own keys. With Letta, usage on your own keys bypasses Letta credits. OpenClaw also takes provider subscription sign-in or models on your own hardware, and Glean offers its Universal Model Key as the alternative.

Sources

Top of agent platforms

The category page
Show the whole category

More of our research