Framework
comparison
A concise comparison of Agent RT, LangChain, and LlamaIndex across runtime features plus checked-in installation footprint, latency, and memory measurements, and of the Agent RT CLI against six other agent CLIs, including Pi and OpenHuman.
How to read this table✅ means built in, ◐ means available or integration-dependent, and — means it is not a primary built-in focus. Benchmark figures are scenario-specific measurements from the repository comparison suite, not universal framework rankings. Python fresh-install footprint is the 2026-10-03 Python 3.13.13 clean-
uv venv measurement above the empty-venv baseline. TypeScript fresh-install footprint is the 2026-10-03 Node.js 22.22.2/npm 10.9.7 clean-project measurement of logical node_modules bytes. Both include transitive dependencies.⚖️ Quick comparison
| ⚡ Agent RT | 🦜 LangChain | 🗂️ LlamaIndex | |
|---|---|---|---|
| 🎯 Primary focus | 🧠 Agent runtime | 🧩 App/orchestration framework | 📚 Data & RAG framework |
 Python +  TypeScript | ✅ | ✅ | ✅ |
| 🛠️ Tools & agent loop | ✅ Built-in | ✅ | ✅ |
| 🔐 Permissions & approvals | ✅ Built-in | ◐ | ◐ |
| 🧰 API & CLI tools | ✅ Built-in (optional packages) | — | — |
| 🌐 OpenAI + Anthropic compatible API | ✅ Built-in server adapters | ◐ Integration-dependent | ◐ Integration-dependent |
| 🛡️ Agent Action Guard | ✅ Built-in integration (opt-in) | — | — |
| 🧠 Decision-model control gates | ✅ Tool pruning, memory/retrieval gates, failure classification | — | — |
| 🧠 Harmful-memory guard | ✅ Decision-model screening before memory persistence and retrieval exposure; detects harmful-action influence, misleading authority, instruction hijacking, and safeguard bypass attempts | ◐ Application/integration-defined | ◐ Application/integration-defined |
| 🔎 Tool/skill pre-registration safety scan | ✅ Deterministic scan of names, descriptions, schemas/metadata, skill frontmatter/instructions, embedded tools, and string resources | ◐ Application/integration-defined | ◐ Application/integration-defined |
| 🚧 Decision approval for external tools/skills | ✅ Fail-closed Decision-model scan before filesystem skill activation, repeated whenever a trusted skill's files change; checked registration path for untrusted tools | ◐ Application/integration-defined | ◐ Application/integration-defined |
| ✂️ Tool catalog pruning before model call | ✅ Built-in | ◐ Integration-dependent | ◐ Integration-dependent |
| 📦 Sandboxing | ✅ Fail-closed backend selection | ◐ | ◐ |
| ♻️ Budget-preserving checkpoints | ✅ Resume with prior turn/tool/token budgets | ◐ | ◐ |
| 💾 Memory & checkpoints | ✅ Built-in | ✅ | ✅ |
| 🗃️ Switch vector DBs | ✅ Simple env/registry switch | ◐ Integration-dependent | ◐ Integration-dependent |
| 🔌 Provider-neutral | ✅ | ✅ | ✅ |
| 🪶 Lightweight runtime focus | ✅ | ◐ | ◐ |
| Fresh install footprint | 46.43 MiB | 69.87 MiB | 219.91 MiB |
| Warm latency | 1.56 ms | 9.03 ms | 4.26 ms |
| Runtime memory | +5.86 MiB | +19.77 MiB | +22.42 MiB |
| Fresh install footprint | 1.24 MiB | 82.44 MiB | 60.38 MiB |
| Warm latency | 1.85 ms | 2.60 ms | 2.23 ms |
| Runtime memory | +14.33 MiB | +17.27 MiB | +16.79 MiB |
📊 For benchmark methodology, environment, raw measured outputs, startup results, tool-loop measurements, schema-token analysis, and CLI harness results, see the reproducible framework benchmarks ↗.
⌨️ Agent CLI comparison
The agent-rt terminal client compared with other agent CLIs. Each CLI was called as its plain one-shot prompt command with no extra flags and pointed at a deterministic localhost mock model, so no real model API was called. The figures are 30 clean sequential runs per CLI plus an 8-way parallel batch, measured on 2026-10-29 on 16-thread Linux x86_64.
| CLI | Command | Median time | Peak RAM | Input tokens | Cost / prompt | 8 parallel |
|---|---|---|---|---|---|---|
| ⚡ Agent RT | agent-rt -p "hi" | 185 ms | 36.5 MiB | 851 | $0.0026 | 338 ms |
| Codex | codex exec "hi" | 283 ms | 194.1 MiB | 8,420 | $0.0253 | 318 ms |
| Claude Code | claude -p "hi" | 275 ms | 202.7 MiB | 19,229 | $0.0577 | 1,105 ms |
| opencode | opencode run "hi" | 683 ms | 386.3 MiB | 5,604 | $0.0169 | 2,188 ms |
| Copilot CLI | copilot -p "hi" | 1,595 ms | 341.1 MiB | 13,985 | $0.0420 | 2,796 ms |
| Pi | pi -p "hi" | 312 ms | 125.5 MiB | 1,330 | $0.0040 | 561 ms |
| OpenHuman | openhuman-core headless turn¹ | 492 ms | 95.9 MiB | 4,208 | $0.0127 | 688 ms |
What this measuresStartup and per-prompt harness overhead, not task quality or capability. Input tokens are a same-tokenizer estimate (
tiktoken o200k_base) of the system prompt, messages and tool schemas each CLI sends. Part of the gap comes from each CLI's default tool set (Agent RT sends 8 tools; the others 4–16). ¹OpenHuman's release one-shot command cannot run a turn without a booted core, so each run boots openhuman-core run --headless-api, sends one JSON-RPC turn and stops, as OpenHuman's own benchmark does; it is built to serve many agents from one warm core, which this cold-start test does not measure. Cost assumes $3 per million input and $15 per million output tokens, with no cache discount. Peak RAM covers each CLI's whole process tree.🧭 Pi and OpenHuman features
Pi ↗ and OpenHuman ↗ are open-source agent harnesses with very different scopes. This comparison comes from each project's documentation and source as of 2026-10-29. ✅ built in · ◐ partial, via extension or add-on · — not provided.
| Agent RT | Pi | OpenHuman | |
|---|---|---|---|
| Language / runtime | Python and TypeScript (aligned) | TypeScript (Node ≥ 22.19) | Rust (single process) |
| License | MIT | MIT | GPL-3.0 |
| Primary shape | Agent runtime library + CLI + web UI | Minimal coding-agent CLI + TS SDK | Desktop app + headless core + Rust library |
| One-shot CLI | ✅ agent-rt -p | ✅ pi -p, JSON mode, RPC mode | ◐ openhuman inference agent_chat (needs a booted core in 0.64.15) |
| Embeddable library | ✅ Python + TS packages | ✅ TS SDK (pi-agent-core, pi-ai) | ✅ Rust (openhuman-embed) |
| Providers | OpenAI, Anthropic, any compatible endpoint, custom | 30+ built in, plus any compatible endpoint (models.json) | 26 BYOK presets, local runtimes (Ollama, LM Studio, MLX), managed subscription |
| Drop-in SDK migration | ✅ One-import swap for OpenAI, Anthropic, LangChain, LlamaIndex, OpenAI Agents and others; agent-rt migrate | — | — |
| Permissions / approvals | ✅ Built in | — None built in; relies on project trust, containers or extensions | ✅ Approval gate |
| Action blocking | ✅ Guard: composable blockers, incl. Agent Action Guard | ◐ Via extension hooks | ◐ Prompt-injection screening on inputs; approval for irreversible actions |
| Sandboxing | ✅ Fail-closed backend selection | ◐ Via containers or the Gondolin extension | ✅ OS jail or Docker |
| Tool catalog pruning | ✅ Decision-model (Jev-compatible) pruning + tool_search | ◐ codemode (a script calls tools; the model sees only its output) | ✅ Jev tool search |
| Context reduction | Smaller default catalog; Decision-gated memory/retrieval | Compaction | RLM tool-result compression |
| Memory | ✅ mem0, Zep, Letta, Supermemory, memU, framework stores | — (sessions only) | ✅ Built-in recall every turn, with citations; swappable engines |
| Checkpoints / resume | ✅ Budget-preserving checkpoints | ✅ Sessions with branching | ✅ Durable workflows that resume after a pause |
| Sub-agents / plan mode | ✅ Agent-as-tool, plans | — By design (build via extensions) | ✅ Sub-agents, agent teams, Plan vs Build |
| MCP | ✅ stdio + streamable HTTP; CLI reads Codex and Claude Code server configs | ✅ stdio + streamable HTTP | ✅ |
| Skills / extensions | ✅ Skills from .agents/, .claude/, .codex/; re-scanned when files change | ✅ Extensions, skills, prompt templates, themes, npm/git packages | ✅ Skills, loadable Rust modules |
| Messaging channels | ◐ Telegram, Discord (Python, beta) | — (separate pi-chat project) | ✅ 14 channels |
| Browser / desktop control | — | — | ✅ Browser and desktop apps |
| OAuth app integrations | — | — | ✅ 119 apps via Composio |
| Voice | — | — | ✅ Live voice agent, dictation, TTS |
| Web UI | ✅ agent-rt web (time-travel debugger) | — (terminal UI) | ✅ Desktop app |
| OpenAI/Anthropic-compatible API server | ✅ | — | ◐ Own JSON-RPC API |
Pi is the minimalist option: a small core with few tools and no permission system, extended by the user. OpenHuman is the maximalist one: a personal-assistant platform with memory, channels, browser and desktop control, voice and agent fleets. Agent RT sits between them as a runtime with production controls (Guard, approvals, sandboxing, Checkpoints) and Python and TypeScript migration paths. It has none of OpenHuman's end-user assistant features (browser control, voice, OAuth app catalog).