Autonomous AI agents
on your infrastructure.
Plexo is the agent harness — tasks, schedules, channels, and one-way-door safety. Connect Claude, GPT, or your own Ollama. Self-hosted. AGPL-3.0.
17
Providers
6
Extension types
5
Channels
AGPL-3.0
License
agents:
- name: research
model: claude-sonnet-4-6
memory: persistent
tools: [web, github, notion]
- name: ops
model: gemini-2.5-flash
schedule: "*/30 * * * *"
tools: [docker, ssh, deploy]
- name: writer
model: gpt-4.1
channels: [slack, telegram]
tools: [docs, calendar]
routing:
fallback: [anthropic, openai, groq]
strategy: cost-optimizedCapabilities
What Plexo does
A self-hosted platform for running autonomous AI agents with full control over models, data, and cost.
Intelligent Model Routing
Fallback chains across providers. If primary fails or hits rate limits, Plexo tries the next. Per-model reliability scoring learns which providers work best.
Bring Your Own Keys
17 providers: Anthropic, OpenAI, DeepSeek, Groq, Mistral, Google, xAI, Ollama, OpenRouter, Cerebras, Cohere, Fireworks, Together, Perplexity, SambaNova, Cloudflare, and any OpenAI-compatible endpoint.
Multi-Channel
Same agent everywhere. Web, Telegram, Slack, Discord, REST API, embeddable widget. Voice transcribed via Deepgram. Images via vision-capable models.
Independent Quality Judge
Separate model evaluates every task output. Ensemble mode runs multiple local judges via Ollama with weighted consensus. Cross-provider judging prevents self-evaluation bias.
Self-Extending Agent
Need an integration? The agent scrapes API docs and generates a working PEX extension -- credential UI and sandboxed execution included. No manual plugin dev.
Project Decomposition
Describe a project. Plexo decomposes it into parallel tasks with dependency-aware wave scheduling. Each task gets its own branch, agent, and draft PR.
Safety
Built-in guardrails
One-Way Door Approvals
Irreversible actions require human approval. Standing approvals for trusted ops. Risk-level classification from low to critical.
Context Intelligence
Stale tool results auto-compressed. Per-model output ceilings prevent truncated tool calls. Truncation triggers automatic retry.
Cost Ceilings
Per-task and per-project cost ceilings halt execution before runaway spending. Budget checks before every wave.
Audit Trail
Every tool invocation, extension activation, and approval decision logged with SHA-256 payload hashing.
Semantic Context Lattice
Knowledge that compounds.
SCL is a persistent knowledge graph built from completed work. Every task extracts structured concepts -- entities, events, actions, claims -- embeds them as vectors, and writes them into a Golden Record. The lattice self-modifies after every task.
Concept Attractors
Vector-positioned nodes with salience scoring, mutation tracking, and depth classes: spirit (drift-protected) and mechanics (freely evolving).
Domain Regions
Semantic clusters with centroids, radius, density. 28+ relation types between regions.
Budget-Aware Retrieval
Token-budget-constrained context retrieval. Priority-sorted, region-aware, budget-packed.
Extensions
Six extension types. One runtime.
Skill
Composite capability -- registers tools, schedules, widgets, and prompts.
Channel
Messaging bridge with onMessage and healthCheck. Telegram, Slack, Discord, or custom.
Tool
Stateless, single-purpose function. Called on demand by the agent or executor.
Connector
Bridges an external MCP server into the PEX sandbox. Translates tool definitions.
Agent
Autonomous actor with plan, executeStep, verifyStep, and escalation contract.
MCP Server
Model Context Protocol server -- stdio or SSE transport, standard tool discovery.
Entity-scoped permissions (memory:read:transaction), three compliance levels (Core/Standard/Full), mandatory escalation for irreversible actions. Install from the Hub or build your own.
Self-hosted
Why self-hosted
Your data never leaves
Task history, agent memory, conversation logs -- all on your infrastructure. No telemetry phones home.
No model lock-in
Switch providers via dropdown. Fallback chains ensure uptime. Run local models via Ollama with zero external calls.
AGPL-3.0 open source
Every line inspectable. Free forever for self-hosted use. Commercial licensing for modified network services.
One-command deploy
Docker Compose. Install script generates secrets, writes env, running in 60 seconds. No Kubernetes required.
Community
Early adopters
“Plexo lets me delegate multi-step infra tasks to agents I can audit. It's the first AI tool I'd run on my own servers without hesitation.”
“We use it for internal ops: deploys, log reviews, PR triage. Having an agent loop we fully control is the key for us.”
“A2A support means our agents can delegate to each other. We're building workflows that would've taken weeks of engineering otherwise.”
Ready to deploy
One command. Your server. Full control over models, data, and cost.