Letta (formerly MemGPT) is a stateful AI agents platform built around long-term memory, tool execution, and multi-agent coordination. The Letta REST API exposes 239 endpoints across 36 public resource categories — agents, memory blocks, archival memory, sources (RAG), custom tools, MCP servers, multi-agent groups, identities, runs, scheduled messages, and a streaming voice-mode endpoint. The open-source server (Apache-2.0, 22.9k+ stars) is available on GitHub; Letta Cloud is the managed offering; the Agent Development Environment (ADE) is the web UI.
APIs.json: apis.yml · raw on GitHub
- x-type: company
AI, Agents, Stateful Agents, Memory, MemGPT, Continual Learning, MCP, Multi-Agent, RAG, Open Source
Managed REST API. 239 paths under /v1/. Bearer-token auth (Authorization: Bearer $LETTA_API_KEY). Covers:
- Agents (56 ops) — create, list, modify, export, message, stream.
- Tools (21 ops) — sandboxed Python, client-side, MCP-delegated.
- Identities (18 ops) — per-user agent/memory scoping.
- Sources (14 ops) — file/URL ingestion for archival memory.
- Groups (11 ops) — multi-agent (supervisor / round-robin / dynamic / sleep-time).
- MCP Servers (10 ops) — register external MCP servers (streamable_http / sse / stdio).
- Runs (10 ops) — inspect agent executions, steps, usage.
- Blocks (9 ops) — labeled in-context memory blocks, shareable across agents.
- Messages (8 ops) — send, list, search, batches.
- Providers (8 ops) — LLM provider credentials and routing.
- Plus: archives, passages, jobs, scheduled messages, voice (beta), models, sandboxes, folders, projects, templates, conversations, environments, and more.
Open-source Letta server (Apache-2.0). Same REST API surface as Letta Cloud. Pay only for hosting and your own LLM API usage. Agents move between self-hosted and cloud via the open .af agent file format.
"The memory-first coding agent that remembers and learns." CLI + desktop app + GitHub action, built on the Letta Code SDK.
- Website · Blog · LinkedIn
- Documentation · llms-full.txt · Pricing
- GitHub org: letta-ai · Python SDK · TypeScript SDK
- Sign in: app.letta.com · Agent File (.af) format
| Artifact | Count | Folder |
|---|---|---|
| OpenAPI specs | 2 (JSON + YAML) | openapi/ |
| Spectral rulesets | 1 | rules/ |
| Naftiko capabilities (one per public tag) | 36 | capabilities/ |
| JSON Schemas (per entity) | 14 | json-schema/ |
| JSON Structures | 14 | json-structure/ |
| JSON-LD context | 1 | json-ld/ |
| Examples (operations + payloads) | 28 | examples/ |
| Vocabulary | 1 | vocabulary/ |
| Plans (pricing) | 1 | plans/ |
| Rate limits | 1 | rate-limits/ |
| FinOps | 1 | finops/ |
openapi/letta-openapi.json— Letta API v1.0.0, 239 paths, 347 component schemas, 36 public tags, bearer-token auth. Mirrored fromletta-ai/letta/fern/openapi.jsonwith summaries Title-Cased and acronyms (MCP, LLM, API, ID, SSE) normalized.openapi/letta-openapi.yml— YAML rendering of the same spec.
One self-contained Naftiko capability file per public OpenAPI tag — each declares both a REST exposer and an MCP exposer routed through its inline consumes block. All 36 files in capabilities/:
letta-agents, letta-blocks, letta-archives, letta-passages, letta-sources, letta-tools, letta-mcp-servers, letta-groups, letta-identities, letta-runs, letta-jobs, letta-messages, letta-conversations, letta-scheduled-messages, letta-voice, letta-models, letta-llms, letta-embeddings, letta-providers, letta-folders, letta-projects, letta-templates, letta-sandboxes, letta-memory-files, letta-device-storage, letta-feeds, letta-environments, letta-client-side-access-tokens, letta-pipelines, letta-health, letta-telemetry, letta-steps, letta-tag, letta-metadata, letta-chat, letta-admin.
Per-entity Draft 2020-12 schemas: agent, block, tool, source, run, group, identity, message, passage, archive, job, llm-config, embedding-config, provider.
- Self-Hosted: free (you pay infra + LLM).
- API Plan (Cloud): $20/mo base + $0.10/active agent/mo + $0.00015/sec tool execution + LLM pass-through at provider cost.
- Enterprise: custom (RBAC, SAML/OIDC SSO, increased quotas).
See plans/letta-plans-pricing.yml, rate-limits/letta-rate-limits.yml, finops/letta-finops.yml.
| Repo | Stars | Lang | Purpose |
|---|---|---|---|
| letta | 22.9k | Python | OSS server, the platform itself (Apache-2.0) |
| claude-subconscious | 2.7k | TypeScript | Letta-based memory subconscious for Claude Code |
| letta-code | 2.5k | TypeScript | Memory-first coding agent |
| agent-file | 1.1k | TypeScript | Open .af file format for stateful agents |
| sleep-time-compute | 132 | Python | Companion material to the sleep-time compute paper |
| skills | 105 | Python | Shared skills (Letta Code / Claude Code / Codex) |
| letta-chatbot-example | 99 | TypeScript | Reference chatbot built on Letta |
| context-constitution | 74 | — | Letta primitive for memory governance |
| letta-oss-ui | 73 | TypeScript | Open-source demo UI on the Letta Code SDK |
| letta-obsidian | 72 | TypeScript | Obsidian plugin running a Letta agent on your vault |
| letta-evals | 71 | Python | Evaluation kit for stateful agents |
| letta-code-sdk | 68 | TypeScript | Letta Code SDK |
| letta-python | 52 | Python | Python SDK |
| letta-node | 48 | TypeScript | TypeScript SDK |
| learning-sdk | 45 | Python | Drop-in continual learning SDK |
| n8n-nodes-letta | 18 | TypeScript | Official n8n node |
| vercel-ai-sdk-provider | 20 | TypeScript | Official Vercel AI SDK provider |
| letta-voice | 25 | Python | Low-latency voice chat with Letta agents |
| deep-research | 27 | Python | Open-source deep-research agent on Letta |
- MemGPT lineage — Letta is the production framework that grew out of the 2023 MemGPT paper on virtual-context-management for LLM agents.
- Sleep-time agents — background agents that share memory with primary agents and asynchronously consolidate/rewrite it (
enable_sleeptime: true). - Multi-agent groups — supervisor / round-robin / dynamic / sleep-time manager configs; agents can share memory blocks across the group.
- MCP as a first-class consumer — Letta registers external MCP servers (streamable_http, sse, stdio) and exposes their tools to agents alongside sandboxed Python and client-side tools.
- Agent Development Environment (ADE) — "bridges the gap between development and deployment, providing complete transparency, state control, rapid prototyping, and robust debugging."
- Agent File (.af) — open format for serializing a full agent (memory + tools + config) for export, versioning, and migration.
FN: Kin Lane · Email: kin@apievangelist.com