9 Letta alternatives (2026)

Where it improves on Letta

• +Seat model · YesTwo seat types with a hard ceiling: Full Seats are capped at 11 per self-serve workspace, every plan includes 20 Basic Seats, and Enterprise seat counts sit in the contract.Source ↗
• +Pooled budget · YesOne shared workspace balance for everyone; admins carve it up with billing groups that set credit quotas per team and reject requests once a group is exhausted.Source ↗

Trade-offs against Letta

• −Writes code · PartlyElevenLabs agents do not author code themselves; the Cursor integration lets an agent launch Cursor background agents that do, and needs a paid Cursor plan.Source ↗
• −Debugs and fixes · PartlyBug fixing is delegated: the agent starts and steers a Cursor background agent, which diagnoses and lands the change; ElevenLabs itself never touches the repository.Source ↗
• −In your editor · NoThe vendor enumerates its surfaces as web widgets, mobile apps, phone systems and custom integrations; no editor or IDE extension appears in that list.Source ↗

Where it improves on Letta

• +Audit log · PartlyEach run returns its result with sources and completion evidence for review, which is per-run traceability; no account-level log of administrative actions is described.Source ↗
• +Opt out of training · YesThe vendor states flatly that no model is trained on customer data, and its edge boundary is designed to keep identifying values from reaching a model at all.Source ↗

Trade-offs against Letta

• −Agent permissions · PartlyOutward-facing actions are off unless explicitly granted for that process, so the default boundary is read-and-prepare with named human checkpoints before anything leaves the company.Source ↗
• −Open models · PartlyOpen-weight families DeepSeek, Mistral, Qwen, GLM and Llama are listed as usable models, but the vendor never says whether they run in its own cloud or where.Source ↗
• −Model choice · PartlyThe Engine, not the buyer, chooses a model per task; the vendor's claim is portability of the task definition across models rather than user-level model selection.Source ↗

Where it improves on Letta

• +Getting out · YesThere is nothing to export from a vendor: session state, event logs and tool output are plain JSON and JSONL files on your own disk.Source ↗

Trade-offs against Letta

• −Runs autonomously · PartlyReads, writes and ordinary project commands run automatically, but a human must sign in to the web AI first and approve anything sensitive or destructive.Source ↗
• −Agent permissions · PartlyA local policy engine grades every tool call by risk: reads, writes and builds inside the project run unattended, sensitive or destructive calls need approval.Source ↗
• −Open models · PartlyProviders include DeepSeek, Kimi and GLM, whose weights are published openly, but WTAgent drives their hosted web chat rather than running any weights locally.Source ↗

Where it improves on Letta

• +Seat model · YesSeats and teams are unlimited on both plans with no per-seat charge or stated minimum, so cost scales with credit consumption instead of headcount.Source ↗
• +Pooled budget · YesTeam members draw on one shared credit balance on both plans rather than each holding a private allowance, which is what makes unlimited seats workable.Source ↗

Trade-offs against Letta

• −Reviews code · PartlyThrough the GitHub connector an agent can open a pull request's changed files and inline review comments and post its own comment, but there is no dedicated review product or verdict workflow.Source ↗
• −In your editor · PartlyThere is no Gumloop IDE extension; instead its remote MCP server is documented for Cursor, VS Code and Windsurf, so the editor drives Gumloop agents rather than Gumloop living in the editor.Source ↗
• −In your pipeline · PartlyWebhook triggers let GitHub, a cron job or an internal script start an agent, but the vendor documents no CI-native runner or agent that acts on a pull request.Source ↗

Where it improves on Letta

• +Whole-codebase work · YesMissions decompose multi-feature work into milestones and are documented for brownfield migrations.Source ↗
• +MCP server · PartlyDroid 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.Source ↗

Trade-offs against Letta

• −Self-hosted · PartlyThe agent always runs on your machine, and a dedicated airgap build exists, but airgap builds are Enterprise-only and not publicly distributed.Source ↗
• −Open source · PartlyThe CI action is a public repository; the Droid agent itself ships as binaries, with no source offered.Source ↗
• −Free tier · NoEvery priced plan on the page is paid: the ladder runs 01 Pro at $20/mo to 06 Enterprise, with no $0 tier and no trial named anywhere on it, so the cheapest way in is a self-serve $20 monthly subscription. Free usage appears only inside paid plans - a BYOK allowance and a free switch to Droid Core after a rate limit.Source ↗

Where it improves on Letta

• +Seat model · YesSeats are a plan ceiling, not a per-seat price: 1 on Sandbox, 3 on Professional, 50 on Team.Source ↗
• +Pooled budget · YesOne pooled balance: message credits, apps and storage quotas belong to the workspace, not to individual members.Source ↗

Trade-offs against Letta

• −Agent permissions · PartlyAutonomy is bounded by design, not by a permission prompt: a Human Input node pauses a run for review, approval or edits, with a timeout branch.Source ↗
• −Open models · PartlyLocally deployed models are supported on air-gapped Enterprise deployments; on Cloud you install a provider from the Marketplace and supply its credentials.Source ↗
• −In your editor · PartlyNo IDE extension; instead a published Dify app can be attached to Cursor or Claude Desktop as an MCP server.Source ↗

Where it improves on Letta

• +Seat model · YesThree seat types are priced per seat per month and can be mixed and reassigned within one workspace; no seat minimum is published, though SSO on Business is noted at 5+ seats.Source ↗
• +Pooled budget · PartlyEach paid seat has its own monthly credits first; a shared workspace pool exists but needs admin-purchased top-ups on Business and is contract-funded on Enterprise, and Free seats cannot touch it.Source ↗

Trade-offs against Letta

• −Writes code · PartlyAgents author React components (Frames) and run code in a sandbox to produce files, but Dust documents no repository-level code authoring or commits.Source ↗
• −Debugs and fixes · PartlyThe engineering page claims debugging support by surfacing code context, docs and recent issues in the IDE, which stops short of diagnosing and landing a fix itself.Source ↗
• −Runs autonomously · PartlyAgents run unattended on schedules and event triggers, but a root task started from the web app or extension pauses for human confirmation once it reaches 600 credits of AI spend.Source ↗

Where it improves on Letta

• +Pooled budget · YesOne prepaid balance covers the whole workspace, with an optional per-member limit layered on top; a request must fit both ceilings.Source ↗
• +Admin controls · YesAdmins constrain spend and model choice in the workspace, and an enterprise central config can restrict the agent to one approved gateway.Source ↗

Trade-offs against Letta

None found.

Where it improves on Letta

• +Seat model · YesBilling is strictly per user per month; the calculator's slider runs from a single user upward and the vendor publishes no seat minimum or ceiling.Source ↗
• +Pooled budget · YesBudget rules can be scoped to the whole organization, a team, a role, a user or one API key, so spend is pooled or individual as configured.Source ↗

Trade-offs against Letta

• −Reviews code · PartlyA separate agent can be granted a review tool to judge another agent's completed result and its evidence, but the vendor scopes this to task deliverables and says Tale does not verify GitHub state itself.Source ↗
• −Ships it · NoWork stops at collected deliverables and a review verdict; the vendor states plainly that a review does not merge a pull request, and connector writes are refused to agents.Source ↗
• −Runs autonomously · PartlyAn agent completes a task unattended and posts its report, but it must stop at In review and cannot approve its own result, so a person or a separate reviewer agent closes the loop.Source ↗