MEEPT
Production-Grade
Autonomous Agent Platform

A self-hosted autonomous agent platform with intelligent agent orchestration, memory-backed conversations, and tool-driven autonomy.

$ meept chat "fix the auth bug" <enter>
● classifying intent...
● delegating to coder agent
● done. PR #42 opened.
Meept mascot
🧡

Multi-Agent Orchestration

Dispatcher routes tasks to specialist agents: coder, debugger, planner, analyst, committer.

💬

Memory-Backed Chat

Episodic and task memory via SQLite+FTS5. Agents remember context across sessions.

🛠

Built-in Security

Input sanitization, TIRITH scanning, tool gating, TLS--defense in depth.

⚙︎

Tool Ecosystem

Built-in tools, MCP server integration, skill discovery, and custom tool registration.

🔄

Collaborative Pairs

Channel-based agent pairing with editor token, round-robin, and turn management.

📈

Self-Improvement

Detect, analyze, generate, validate, and apply improvements autonomously.

implemented  ·  partial  ·  not implemented
Architecture & Orchestration
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Multi-agent specialists(Agent Team)(rlm children)(subagent)
Intent classification
DAG workflow/plans(PipelineSpec)
Async handoff between agents(agent_message)
Dynamic subagent spawning(request_handoff)(AgentBus)(rlm())
Recursive programmatic subagentsdelegate + handoff; depth-capped recursion unwired(rlm + passivation)
Agent-to-agent messaging(AgentBus)(send + receipts)
Constitution-bound employees
Autonomy tiers (reactive/propose/autonomous)(bounded auto)
Goal loop with enforcement(persistent goals)(/goal + autonomous)
Quality-gated goal completion
Plan approve/reject lifecycle(approval gate)
Collaboration engine (pair/diff)
Memory & Context
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Short-term conversation store
Long-term memory (SQLite/FTS5)(6 tiers)(JSONL)(SQLite+FTS5)(JSONL)(9 providers)(SQLite FTS)
Vector/semantic search(sqlite-vec HNSW)(optional embeddings)
Knowledge graph (PageRank)(bounded links)
Epistemic memory (claims/trust)
Context compaction / summarization(3-layer firewall)(guarded_mid projection)(summarize+compact)
Session persistence(SQLite, tree)(checkpoint)(append-only JSONL)(SQLite)(JSONL)
Conversation branching(/revert)(/rewind checksum-guarded)
Steering / follow-up queues
Tools & Execution
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Tool count40+~1518~30~1586~20~10~20
MCP client(20 preconfigured, 6 enabled)Limited
MCP server mode
ACP client (drive external agents)(opt-in, default off)(ACP server for editors)(opencode acp)
Parallel tool execution(semaphore)(resource-class)
Tool streaming progress
Browser automationopt-in [browser] · SSRF-guarded Chrome(Playwright)
Computer use (CUA)opt-in cua-driver MCP(macOS)
GBNF grammar-constrained toolsopt-in gbnf_constrained
Vision / image describemultimodal + pre-flight describe(mmproj)
Office doc read/writeMCP only(PDF/DOCX/etc.)MCP
Security & Sandboxing
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Permission system(SecurityEngine)(allowlist)(JSON policy)(approval ladder)(allowlist)
Input sanitization(prompt injection)
Command scanning (Tirith)
Output scanning
Taint tracking(lattice)
Path fencing
OS-enforced sandbox(docker opt-in)(bwrap/container)(bwrap/macOS/Windows)
Network egress policy(proxy+CIDR)
Secret-backed credential injectionopt-in secrets broker + placeholders(placeholder proxy)
Diff preview + write approval(PendingChanges, FileEdit only)(unified diff all mutations)(checksum rewind)
Fail-closed policy
Scheduling & Persistence
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Cron jobs
Job queue with priorities
Agent-targeted jobs
Resident / always-on mode(gateway service)(resident supervisor)(gateway)
Crash recovery / resumestartup reclaim + persisted parks(--resume)(worker restart)
P2P cluster mesh(gossip+WireGuard)
Distributed task queue
Model & Cost
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Provider count10+~530+~10~1015+2-3limited~3
Capability-based routing
Model failover chain
Token budgeting
Dollar cost tracking(OpenRouter live)
Reasoning effort support
Local inference management(RuntimeManager)(TurboQuant llama.cpp)
Subscription CLI provideropt-in ACP (codex-acp shipped)(claude/codex)
Self-Improvement & Learning
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Skill evolution (closed loop)(/refine)
Shadow training (LoRA/DPO)
Reflection collector
Routing decision log
Q Agent meta-optimization
Automated code fixing
Observability & Ops
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Metrics store(SQLite TSDB)(artifacts)(NDJSON traces)(cost)
Structured logging(slog)
Health endpoints
Benchmark harnessmeept-bench + internal eval(14 benchmarks)(context policy)(GAIA published)
Trace replay / prompt drift(NDJSON+hash)
UI & Channels
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
TUI(Bubbletea)(Textual)(Ink)(prompt_toolkit)
CLI
GUI (Flutter web+desktop)
macOS MenuBar app
Web API + REST
WebSocket events
Telegram bot
Chat channels (30+)(30+ channels)
STT / TTS
Desktop notifications
Multi-user auth(per-channel allowlists)
Code Intelligence
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
AST tools (tree-sitter)
LSP client
RepoMap / PageRank
Auto-lint + reflection
Other
Feature Meept FrontierAgent duckagent atomic-agent prime-agent Hermes OpenCode oh-my-pi Claude Code
Evidence pipeline (claim-evidence)
Hallucination detection
Context propagation to subtasks
Token cache (L1+L2)
Session branching / forking
Change journal with revert(git checkpoints)(WorkspaceJournal)(checksum rewind)
Multi-platform installers(static binary)
Google Calendar integration
Video generation
Image generation
Where Meept Leads
Category Meept Advantage
Multi-agent orchestrationOnly framework with 22 specialists + 6 reviewers, intent classification, DAG planning, async handoff, and a collaboration engine covering pair programming, differential A/B, and parallel-team mixture-of-agents (lead + N specialists, presets)
Memory depth6-tier system: episodic (FTS5), task, knowledge graph (PageRank), semantic (vector HNSW), distributed (memvid), epistemic (claims/trust)
AI EmployeesOnly framework with constitution-bound autonomous agents, three autonomy tiers, enforcement engine, and optional quality-gated goal completion
Evidence pipelineOnly framework with claim-evidence matching, validation gates, and needs_info routing for human review
Defense-in-depth securityOnly framework combining taint lattice, input sanitizer, Tirith shell scan, adversarial boundary markers, and fail-closed policy
Self-improvement loopsOnly framework with closed-loop skill evolution (usage tracking + LLM-judge verifier + versioning), shadow LoRA/DPO training with eval gate, reflection collector, and Q-Agent meta-optimization
Always-on platformOnly framework with resident platform + five frontends (TUI/Flutter/MenuBar/Telegram/HTTP+WS) + voice (STT/TTS)
Cluster meshOnly framework with P2P gossip networking, WireGuard sync, and distributed task queue
MCP server modeOnly framework that can be consumed BY other agents as an MCP server
ACP clientDrive Codex/OpenCode/other Agent Client Protocol agents as full peers (acp_agent); [acp] enabled defaults false

Matrix generated 2026-08-29 from docs/features.md and the agent parity audit — see docs/feature-comparison-matrix.md in the repo.

Install & run in 3 commands

# 1. Clone and install the platform
$ git clone https://github.com/bhodgens/meept.git
$ cd meept
$ make install

# 2. Start the platform
$ meept-daemon
[meept] listening on :8081
[meept] agents loaded: dispatcher, coder, debugger, planner, analyst, committer
[meept] memory store initialized: sqlite://~/.meept/episodic.db

# 3. Chat with your agent
$ meept chat "fix the auth bug in handlers.go"
[dispatcher] intent: CODE -> routing to coder agent ...
[coder] analyzing handlers.go...
[coder] found 2 vulnerability points -> opened PR #42
[coder] done. 4 files changed, 23 additions, 7 deletions.

See it in action

demo video — coming soon