github.com/Muvon/octomind @0.35.0

Install agents, not frameworks.

Open-source runtime for specialist AI agents. One command. Any model. Any domain.

Documentation · Tap Registry · Website

Table of Contents

• The Problem
• Five Pillars
• Pillar 1 — Zero Config, Full Flexibility
• Pillar 2 — Sessions That Stay Sharp at Hour 4
• Pillar 3 — Cost as a Control Plane
• Pillar 4 — Guardrails: Policy as Code, Not Approval Clicks
• Pillar 5 — Intent-Driven Context: Skills Activate on What You Mean
• Quick Start
• How It Works
• Power Users — Roles, Workflows, Layers
• Embedders — ACP, WebSocket, Daemon
• Installation
• Configuration
• Architecture
• Contributing
• License

The Problem

Building a specialist AI agent in 2026 means stitching together three different tools, writing glue code nobody wants to own, and praying it holds. You spend 45 minutes wiring MCP servers, system prompts, tool configs, and credentials — every domain, every machine, every time.

• Config Wars. No central registry. Skills here, MCP servers there, agent configs nowhere. The community calls it "solidarity in frustration."
• Generic AI hallucinates in expert domains. ChatGPT writes wrong drug dosages. Lawyers cite cases that don't exist. Multi-agent specialization is now the default architecture for serious work.
• One generic assistant for every task. Rust debugging, blood-test interpretation, contract review — same prompt, same tools. Drift compounds.
• Sessions break at hour 4. Naive truncation drops the decisions you need. Quality collapses. You restart.
• Bills surprise you. Cursor users posting $7K daily overages. No per-task budget, no kill switch.

Octomind ships specialist agents ready to run — and a runtime that grows with you.

Five Pillars

Pillar 1 — Zero Config, Full Flexibility

Most agent tools force a tradeoff: zero-config (Lindy, no-code) or fully customizable (Mastra, LangGraph). Octomind gives you both. octomind run lawyer:sg works out of the box. Need a different model, a custom MCP server, a guardrail pipe — all live in TOML, no framework code.

octomind run developer:general    # general dev, language skills auto-activate
octomind run doctor:blood         # blood-test interpretation specialist
octomind run doctor:nutrition     # nutrition specialist

What happens when you run a specialist

→ Fetches the agent manifest from the tap registry
→ Installs required binaries automatically (skips if already present)
→ Prompts once for any credentials, saves permanently
→ Spins up the right MCP servers for this domain
→ Loads specialist model config, system prompt, tool permissions
→ Ready in ~5 seconds, not 45 minutes

This is packaged expertise — not a prompt file, not a skill injection. The full stack, configured by the community, ready to run.

Specialists grow at runtime

Every agent has built-in power tools that let it acquire new capabilities and spawn sub-agents mid-session, without restart:

User: "Cross-reference our Postgres metrics with the deployment log"

Agent:
  → mcp.enable(postgres-mcp)        # auto-detected need, no user prompt
  → agent.spawn(log_reader)         # delegates log parsing
  → results merge mid-session
  → mcp.disable(postgres-mcp)       # cleans up
  → presents the analysis

Most agent harnesses pre-load every available tool into context. Octomind starts focused for the domain and grows only when needed. Smaller context, lower cost, faster responses, no surprise tools. See Pillar 5 for how activation actually works.

Add your own taps

octomind tap yourteam/tap                 # clones github.com/yourteam/octomind-tap
octomind tap yourteam/internal ~/path     # local tap for private agents

octomind run finance:analyst              # available immediately
octomind run security:owasp

Each tap is a Git repo. Each agent is one TOML file. Pull requests are contributions.

Want to publish your expertise? A doctor:medications, a lawyer:us, a devops:terraform. One file, and everyone with that problem gets a specialist instantly. How to write a tap agent →

Pillar 2 — Sessions That Stay Sharp at Hour 4

Every coding agent degrades after a few hours. Context fills. Decisions get truncated. The agent forgets why it started.

Octomind's adaptive compaction engine runs automatically:

• Cache-aware — calculates if compaction is worth it before paying for it. Never breaks the prompt-cache hit by accident.
• Pressure-tiered — compacts more aggressively as context grows.
• Structurally preserving — keeps decisions, file references, errors, dependencies; drops noise.
• Plan-aware and free-form-aware — works whether you use the plan tool or have a free-form chat.
• Fully automatic — you never think about it.

The second-order benefit: smaller context means fewer tokens, faster responses, lower cost every turn after compaction fires. The three pillars compound.

Work on a hard problem for 4 hours. The agent still knows what it decided in hour one.

Pillar 3 — Cost as a Control Plane

Pick the right model for each step. A cheap one for routine research, a frontier one for review — per-role, per-step, mid-session swap. Real-time cost tracking and hard spending caps come for free.

# Per-role model selection — pay Opus only where it's worth it
[[roles]]
name = "researcher"
model = "openrouter:google/gemini-2.5-flash"   # cheap broad context

[[roles]]
name = "reviewer"
model = "anthropic:claude-opus-4-7"            # precision where it counts

# Hard spending limits — enforced, not advisory
max_request_spending_threshold = 0.50    # USD per request
max_session_spending_threshold = 5.00    # USD per session

• Per-role and per-layer model selection across many providers via octolib — different roles can run on different vendors.
• Mid-session model swap with /model anthropic:claude-haiku-4-5.
• Real-time cost tracking per request and per session.
• Cache-aware token accounting (cache_read_tokens, cache_write_tokens separated from input/output).
• Hard spending thresholds with enforcement. On a session cap Octomind prompts you to continue (interactive) or auto-stops (non-interactive); on a request cap it stops execution outright — before the bill, not after.

Cursor users get $7,000 surprise bills. Octomind agents trip a budget and stop — interactive sessions ask before continuing, non-interactive runs halt automatically.

Pillar 4 — Guardrails: Policy as Code, Not Approval Clicks

Other agent CLIs make the human the safety layer: every dangerous tool call pops a modal, every file write waits for a click. That works for one developer at the keyboard. It breaks the moment you point an agent at a long-running task, a CI job, or an autonomous loop.

Octomind takes the opposite position. Policy lives in scripts, not prompts. Drop a .agents/guardrails.toml in your repo and the runtime enforces it deterministically — pre-call, post-result, post-turn.

# Pre-call deny — block a class of calls before they execute
[[guard]]
match   = "shell(command=^rm\\s+-rf?)"
message = "rm -rf blocked."

# Conditional rule — only fires after the agent ran git status this session
[[guard]]
match   = "shell(command=git push)"
when    = ["+shell(command=git status)"]
message = "Review changes before pushing."

# Post-result hook — non-zero exit injects feedback into the agent's inbox
[[hook]]
match  = "text_editor(path=src/.*\\.rs)"
on     = "success"
script = ".agents/check-clippy.sh"

# Post-turn validator — fires only over the call slice since it last ran
[[validator]]
name   = "tests-pass"
roles  = ["developer"]
script = ".agents/run-tests.sh"

• Guards — pre-call deny rules. Match by capability(arg_name=regex), gate by history (+used / -unused), require loaded capabilities (has = [...]). The agent never even attempts a blocked call.
• Hooks — post-result scripts. Run after each tool returns. Non-zero exit injects stdout into the agent's inbox as a user message — clippy errors, lint failures, format diffs become automatic corrections without restarting the turn.
• Validators — post-turn scripts. Fire only over the new call-log slice (cursor-based, never re-fires on old activity). Filter by role. Output is wrapped in <validation> blocks the agent reads on its next turn. This is what replaces "approve this change?" prompts in autonomous loops.

The DSL is richer than competitor lifecycle hooks: capability+arg-regex+history+role+result-regex in one declarative file. No code to compile, no plugin to install. Designed for full automation: fits CI, daemons, scheduled runs, ACP sub-agents.

The world is going autonomous. The choice isn't "ask vs auto" — it's "auto with deterministic policy" vs "auto with hope." Octomind ships the former.

Pillar 5 — Intent-Driven Context: Skills Activate on What You Mean

Every other agent CLI loads every tool, every skill, every instruction pack into context up front. The model sees fifty tool definitions and a wall of system prompts before the user types a single character. Token bills follow. Cache misses follow. Confused tool selection follows.

Octomind inverts this. Skills and capabilities sit dormant until the user's intent matches them — then they activate, inject their content, and stay only as long as they're relevant. Context is a function of what the user is actually trying to do.

How activation works (no keyword guessing)

Skills describe what they're for. The runtime matches your prompt against those descriptions and only loads the skills that fit:

• Meaning, not keywords. A dedicated embedding model — trained on activation traffic — scores how well your request matches each skill. "Help me refactor this auth flow" and "the login is broken" both find the same skill; "what's the weather" finds none.
• Hand-authored rules where precision matters. Skill authors can pin activation to file names, file contents, or exact phrases when they know better than a similarity score.
• Abstain on near-ties. When two skills score close, neither fires. Better to load nothing than the wrong thing.
• Calibrated to skip, not guess. Wrong activations bloat context and waste tokens. The system defaults to silence when in doubt.

Why this matters

``` Other agent CLIs: Octomind: ───────────────────── ────────── 1. User starts session 1. User starts session 2. Load 50 tools into context 2. Load 5 core tools into context 3. Load 30 skills into system prompt 3. Skills sit dormant 4. User types one sentence 4. User types one sentence 5. Model picks tool from a wall 5. Embed model scores → 1 skill matches → skill content injected → 6 tools in context, not 80 6. Skill goes silent again when no longer relevan