Letta
Listed at https://www.letta.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 按用量：Free $0（最多 3 个有状态 agent、BYOK）、Pro $20/月（最多 20 个 agent、Letta Auto 周/月额度 + 按量超额）、Teams Pro 按席位（共享 agent、权限）、Developer Plan 纯 API 调用按信用（无 agent 上限）、Enterprise 商务报价（自托管/定制模型）。服务端工具按 CPU 时间 $0.00015/秒计费。 | per million tokens |
| Output | Letta Auto 用量超出 Pro 套餐限额后按 pay-as-you-go 模型价（参考 platform.letta.com/models）。Mods、Skills、Conversations 等功能均包含在套餐内，不单独计费。远程 MCP 工具由 MCP 供应商执行，不消耗 Letta 信用。 | per million tokens |
| Cache Read | BYOK 用量从你自带的 API key 走账（OpenAI/Anthropic/Google/Kimi/Z.AI 等），不计入 Letta 信用消耗，缓存/限速策略继承供应商账户。Free 档无 Letta Auto 缓存，超额即限速。 | Discounted |
🤖 Supported Models (7)
✨ Pros
- ✓Open-source Letta Agents SDK (Apache-2.0) + letta-ai/letta main repo with 24,000+ GitHub stars — one of the major open-source agent runtimes
- ✓First-class persistent memory (core / archival / recall blocks + runtime self-editing tools) so agents retain experience across sessions and machines
- ✓Skills API + Mods (Jun 2026): reusable workflows plus harness-level extensions that let agents self-improve their tool stack
- ✓Built-in Channels for Slack / Telegram / Discord plus an always-on cloud VM deployment, so AI coworkers stay reachable
- ✓BYOK across every plan — bring OpenAI / Anthropic / Google / Kimi / Z.AI keys and route usage through your provider account, no Letta credit burn
- ✓CLI `@letta-ai/letta-code` sees 64K+ npm weekly downloads; desktop, CLI, and Cloud Agents share the same runtime
- ✓Conversations API (Jan 2026) supports shared memory across concurrent experiences for multi-user collaborative agents
⚠️ Cons
- ×Focused on the agent runtime and memory layer — not a general LLM API provider; you bring your own model key or route through Letta Auto
- ×No mainland-China direct endpoint — platform.letta.com is a single global endpoint and production traffic from China needs a proxy or relay
- ×Letta Auto usage above Pro plan's quota falls back to pay-as-you-go — heavy agent users can hit $200+/mo
- ×Server-side tools bill CPU time at $0.00015/sec — long-running tool calls (heavy MCP, long bash) make single-task costs non-trivial
- ×No permanent free tier — Free is capped at 3 agents and Pro ($20/mo) is the basic production floor, so the budget-sensitive indie developer tops out fast
- ×Agent-framework semantics have a learning curve: the core / archival / recall blocks, Mods, Skills, and Channels vocabulary is new to most developers
🎯 Best For
Teams building stateful AI coworkers, coding agents, or personal assistants that need cross-session memory, runtime memory editing, Skills-as-workflows, and Channels-based deployment to Slack / Telegram / Discord — and that value the agent framework being Apache-2.0 open source as a long-term hedge against vendor lock-in.
💰 Pricing & Plans
| Plan | Price | Best for | Highlights |
|---|---|---|---|
| Free | $0 / mo | Trying Letta with BYOK | Limited agents (3), limited Letta Auto usage; bring your own API keys for ChatGPT/Codex, Kimi, Z.AI |
| Pro | $20 / mo | Personal users running agents daily | Up to 20 stateful agents, Letta Auto weekly + monthly quota, pay-as-you-go overage, image generation credits |
| Teams Pro | Per seat | Organizations with shared agent workflows | All Pro features + shared agents across the team, permissions, and inter-agent communication |
| Developer Plan | Usage-based (API key) | Teams building on the Letta API in production | Unlimited agents, pure pay-as-you-go credit pricing, programmatic billing dashboard |
| Enterprise | Sales-quoted | Self-hosted or large-scale deployments | Custom models, higher limits, self-hosting options, dedicated support |
🔧 API & Developer Experience
- •Surface: Letta Agents SDK + REST API on platform.letta.com, plus an open-source CLI (`@letta-ai/letta-code` on npm) and a desktop/web app; the same agent runs across desktop, CLI, and Cloud Agents sandboxes.
- •Memory model: First-class persistent memory with core/archival/recall blocks and self-editing memory tools — the agent can edit its own memory block files at runtime instead of stuffing everything into context.
- •Skills: Skills API (configurable in docs.letta.com/configuration/skills) packages reusable workflows the agent can invoke on demand — directly complementary to the Anthropic Skills API for Agent builders comparing both.
- •Channels: Built-in integrations to ship an AI coworker into Slack, Telegram, or Discord (docs.letta.com/configuration/channels) plus an always-on cloud VM deployment option (`platform/computers/byom`) so agents stay reachable from phones.
- •Mods (Jun 2026): Mods extend and adapt the Letta Code harness at the harness layer (not just plugins) so the agent can self-improve and reshape its own tool stack.
- •BYOK: All plans support bring-your-own-API-key (OpenAI/Anthropic/Google/Kimi/Z.AI) — `/connect` in the Letta CLI routes usage through your provider account instead of Letta credits.
- •Tool billing: Server-side Letta API tools bill CPU time at $0.00015/sec; remote MCP tools run on the MCP provider (no Letta credit cost); Letta built-in tools are free except web search/fetch.
🧠 Memory-First Agents: Why the Harness Owns Memory
Letta's bet is that memory is not a plugin but the responsibility of the agent harness. Memory cannot be bolted on as a vector store or a RAG call; the agent has to own its own state, edit its own memory files, and decide what to recall and when. Letta implements this in three concrete ways. (1) Persistent memory blocks (core/archival/recall) survive across sessions and across machines, so a coding agent that learned your repo structure last month still knows it today. (2) The agent can rewrite its own memory at runtime using Letta's memory tools, rather than letting the developer hand-tune every prompt. (3) Mods (introduced June 2026) extend the Letta Code harness itself, letting the agent shape its own tool stack as it learns. Combined with a Skills API for reusable workflows and Channels for Slack/Telegram/Discord deployment, the result is one open-source runtime that ships AI coworkers — and a clean counterpoint to closed agent platforms where memory and skills live behind a vendor.
🌐 Regional Availability & Latency
Letta Cloud is hosted in the US (legal entity Letta, Inc., San Francisco) with engineering in Berkeley and San Francisco. The Letta API at platform.letta.com is a single global endpoint backed by standard cloud infrastructure, and the agent harness itself is open source (Apache-2.0) so teams that need regional residency or self-hosting can run the full Letta Agents SDK on their own infrastructure. For North America and Europe, Cloud-hosted first-byte latency is low; for latency-sensitive workloads, the Enterprise plan offers custom hosting in additional regions. From mainland China, direct access to platform.letta.com is not guaranteed; production traffic typically needs a proxy or relay with cross-Pacific latency similar to other US-based AI platforms (~150-250ms first-byte). The CLI/desktop/web client work over the same global endpoint and share the same constraints. Storage of agent state is region-bound to the deployment you choose (Cloud, Enterprise custom region, or self-hosted) — pick self-hosting if you need China residency or strict compliance requirements that mandate in-region storage.