Alternatives to Letta

Where it improves on Letta

• +Pooled budget · YesA Zen workspace draws on one balance with a workspace-wide monthly cap plus optional per-member caps.Evidence ↗
• +Admin controls · YesPermissions are set in config globally or per tool, so an operator can deny edits or shell commands outright.Evidence ↗

Trade-offs against Letta

No disadvantages found in the criteria researched so far.

Where it improves on Letta

• +Writes tests · YesTest generation and coverage expansion is a named workflow; a separate QA agent also exercises the running software rather than the test suite.Evidence ↗
• +MCP server · YesAgent Server exposes an OpenAI-compatible chat-completions endpoint that another agent or client can call, holding conversation state behind a returned header.Evidence ↗
• +Pooled budget · PartialCredits are pooled across all members, but only inside an Organization, which requires a paid Cloud subscription or Enterprise.Evidence ↗

Trade-offs against Letta

• −In your editor · PartialZed, JetBrains and Toad connect natively over the Agent Client Protocol and VS Code via a community extension, but the vendor labels the whole integration experimental.Evidence ↗
• −Seat model · NoEnterprise is explicitly not licensed per seat and carries unlimited users; the two free tiers are each capped at one user.Evidence ↗

Where it improves on Letta

• +MCP server · YesAgentOS can expose registered agents as an MCP server, so another agent can call them; docs also ship as MCP.Evidence ↗
• +Seat model · YesThree seats included on Pro, $30/month per extra seat; no published ceiling.Evidence ↗
• +Pooled budget · PartialThere is no metered balance to pool; seats share access to the same connected runtimes and usage inside a connection is unlimited.Evidence ↗

Trade-offs against Letta

• −Agent permissions · PartialRuns are unattended unless a tool is marked for approval; then the run pauses in your database until an admin resolves it.Evidence ↗
• −In your pipeline · NoThe enumerated serving surfaces are REST, MCP in Claude/ChatGPT and chat apps; no pull-request or CI surface appears.Evidence ↗
• −Open source · PartialSDK and AgentOS are Apache 2.0; the control plane is the commercial part and is not published as source.Evidence ↗

Where it improves on Letta

• +Writes tests · YesTest generation is named as agent work, and Automations can be configured to add tests on every new commit.Evidence ↗
• +Whole-codebase work · YesCloud agents work across multiple repositories in one task and raise a pull request in each repo they touch.Evidence ↗
• +MCP server · PartialMCP is inbound only: Cursor is the client that connects to MCP servers. Another agent drives Cursor over ACP (agent acp) or the SDKs, not by querying an MCP endpoint Cursor exposes.Evidence ↗

Trade-offs against Letta

• −Open models · PartialAll inference is hosted: the catalogue includes GLM 5.2 from Z.ai and Kimi K3 from Moonshot, but there is no path to point Cursor at open weights you run.Evidence ↗
• −Self-hosted · PartialOnly tool execution moves to your machines: Cursor still runs the agent loop, inference and planning, and the Self-Hosted Pool needs Enterprise. The Enterprise FAQ says there is no on-premises deployment.Evidence ↗
• −Open source · NoThe Terms of Service forbid deriving or accessing the source; the /licenses page covers only third-party open source components used inside the IDE.Evidence ↗

Where it improves on Letta

• +Writes tests · YesDocumented CLI workflow for writing a unit test for a named function.Evidence ↗
• +MCP server · PartialCodex is an MCP client for local and remote servers, but exposing Codex itself as an MCP server is deprecated in favour of the app server.Evidence ↗
• +Seat model · YesPer-user pricing with a two-seat floor and no published ceiling; monthly billing costs $25 per user.Evidence ↗

Trade-offs against Letta

• −Claude · NoSign-in and model support are OpenAI models only; no Anthropic model appears in any documented model list.Evidence ↗
• −Model choice · PartialYou pick the model per run and can define custom providers, but the first-party catalogue is OpenAI models only.Evidence ↗
• −Self-hosted · PartialThe agent binary runs on your machine or runner; inference is OpenAI-hosted unless you route it through your own AWS Bedrock account.Evidence ↗

Where it improves on Letta

• +Writes tests · YesThe /install-qa skill generates QA sub-skills and executes end-to-end test flows per diff, with evidence.Evidence ↗
• +Whole-codebase work · YesMissions decompose multi-feature work into milestones and are documented for brownfield migrations.Evidence ↗
• +MCP server · PartialDroid is an MCP client over stdio/http/sse and can be driven by another program over JSON-RPC or ACP, but no Factory-hosted MCP server is documented.Evidence ↗

Trade-offs against Letta

• −Self-hosted · PartialThe agent always runs on your machine, and a dedicated airgap build exists, but airgap builds are Enterprise-only and not publicly distributed.Evidence ↗
• −Open source · PartialThe CI action is a public repository; the Droid agent itself ships as binaries, with no source offered.Evidence ↗
• −Free tier · NoThe published plan list starts at Pro $20/mo; no free tier or trial is offered, though Droid Core and a BYOK allowance are free within a paid plan.Evidence ↗