Scoped memory
Tenant, workspace, user, role, and session lanes backed by SQLite/FTS retrieval and leakage tests.
control layer for AI agents
One governed point for agents, tools, memory, models, and evals.
Muster is the control layer where independent AI systems assemble, route through policy, and leave an auditable trail.
Many moving parts. One controlled runtime.
pnpm --package=@musterhq/cli dlx muster demo
Muster does not try to be the AI. Muster controls how AI behaves.
20-second tour
Prompt, action, memory receipt, ledger.
The CLI demo shows the loop Muster is built around: a prompt enters the harness, scoped memory is recalled, a governed run completes, tokens are recorded, and integrity can be checked from the terminal.
the agent harness you can audit $ muster demomuster demo — provisioned an isolated workspace and a live stub model service. > Where do we deploy? (recalled 1 scoped memory) Muster deploys to uat-erp.example.com (recalled from scoped memory). > Summarize the day's work. Demo run complete. Every token above is real, recorded to the ledger. run model in out est cost$ waste session------------------------------------------------------------------------------------------------run_2f9c41a8 demo/demo-model 48.1k 412 ~ 0.0061 6.2x ! oneshotrun_5b1e07d3 demo/demo-model 1.8k 128 ~ 0.0004 - oneshot totals by model runs in out cost$ waste-runs--------------------------------------------------------------------------------demo/demo-model 2 49.9k 540 0.0065 1 ! integrity check: OK — 2 runs, 0 corrupt, 0 duplicate, 0 drift, 0 poisoned.That was a real run loop: scoped memory recall, token ledger, integrity verification. $ muster benchmarkscenario turns naive muster reduction replay-overhead--------------------------------------------------------------------------------------codebase-refactor-20 21 84.6k 42.7k 49.6% 90.5%incident-triage-30 31 144.9k 60.5k 58.2% 93.6%erp-data-audit-40 41 205.5k 79.5k 61.3% 95.1%research-synthesis-25 26 160.0k 67.7k 57.7% 92.3%long-support-thread-50 51 280.8k 104.9k 62.7% 96.1%--------------------------------------------------------------------------------------AGGREGATE 170 875.8k 355.2k 59.4% 94.2% Muster reduced naive token cost by 59.4% across these scenarios.Deterministic — no model calls.$
The problem
01 Most agents fail after the demo.
Long-running agents accumulate state. Without a harness, context grows, token costs disappear, memory becomes unsafe, tools get called without governance, and learning happens without tests.
- Context grows
- Token costs disappear
- Memory boundaries blur
- Tools run without policy
- Learning skips evals
- Fallback becomes silent
- Automation gets brittle
The solution
02 Five controls around every run.
Muster keeps governance outside the model provider. The agent can use tools, memory, MCP servers, browsers, and app surfaces, but the harness owns the ledger and boundaries.
Token ledger
Every run records usage, estimates cost where pricing is known, and flags replay waste.
Eval-gated learning
Feedback becomes replayable fixtures before it is promoted into behavior.
Capability packs
Typed tool bundles with declared permissions, secrets, setup guidance, and tests.
Governed execution
Tools, flows, subagents, browser actions, and channel adapters share one evidence trail.
Definitions
What people mean by Muster agent harness.
A governed agent harness is the control layer around an AI agent: memory scope, token accounting, tool policy, provider routing, channel boundaries, and eval evidence. Muster is an AI agent harness for production systems where a long-running assistant may use MCP tools, browser automation, Frappe / ERPNext context, Slack or Telegram channels, and private model routes without losing the audit trail.
Muster agent harness
The open-source TypeScript harness for scoped memory, token ledger, MCP policy, channel adapters, and eval-gated learning.
MCP agent harness
MCP servers become safer when install, auth, allowlists, result caps, and token visibility sit outside the model.
Frappe / ERPNext agent harness
Frappe AI needs DocType, field, workflow, role, and permission context instead of generic ERP answers.
Use cases
03 Built for production automation, not toy chats.
Use Muster where agents need persistent context, governed tools, and an audit trail across real systems.
Browser operators
Route browser-capable work through setup, evidence capture, and approval-aware tool policy.
Frappe / ERPNext
Build module, DocType, field, workflow, and docs context through a plugin instead of bloating the core binary.
MCP ecosystems
Attach MCP servers with include/exclude policy, result caps, circuit breakers, and OAuth setup.
Enterprise web apps
Connect app events and web clients to governed runs through the gateway and surface packages.
Long-running agents
Keep memory, sessions, token spend, feedback, and evals visible over days or weeks.
Capability packs
04 Operators people can recognize and configure.
Muster is not only a terminal chat. It ships setup surfaces, readiness checks, and policy boundaries for channel operators, MCP servers, providers, browser work, and app-specific packs.
Chat app operators
Bring the same governed run loop behind real team and customer surfaces.
- Telegram
- Slack
- Google Chat
- Teams
- Discord
MCP + OAuth setup
Connect external tools with explicit setup links, env checks, OAuth guidance, include/exclude policy, and result caps.
- MCP
- GitHub
- Notion
- Linear
- Google Drive
- Postgres
Browser and web-app work
Run web QA, browser evidence capture, Playwright-style inspection, and enterprise web-app flows through the same audit trail.
- Browser QA
- Playwright
- Web search
- Visual checks
- Artifacts
Frappe / ERPNext packs
Build context around sites, installed apps, modules, DocTypes, fields, reports, scripts, workflows, and production runbooks.
Model and runtime routes
Keep governance outside the model provider while routing to cloud, open-source, private, local, and CLI-auth runtimes.
- Cloud APIs
- Enterprise APIs
- Gemini
- vLLM
- Self-hosted
- Private endpoints
Artifact and data packs
Use governed runs for reports, dashboards, generated docs, data checks, release notes, and repeatable eval evidence.
- Reports
- Dashboards
- Docs
- Data QA
- Evals
How it works
05 One harness around models, memory, tools, and apps.
The CLI and gateway feed the same run loop. Providers answer; Muster owns the scoped memory, token ledger, MCP policy, capability packs, flows, and eval evidence.
CLI / gateway
Terminal chat, webhook channels, and web surfaces enter one governed envelope.
Runtime router
Routes to cloud APIs, Gemini, open-source/self-hosted servers, aggregators, Pi, private endpoints, and optional CLI-auth runtimes.
Capability layer
Capability packs, MCP servers, browser tools, and flows are policy-wrapped.
Memory + ledger
Scoped SQLite/FTS memory and token records stay local and inspectable.
Verification
Integrity checks and eval fixtures keep learning and automation testable.
Proof
06 The Token Waste Index.
Deterministic measurement — no model is called. Across 5 realistic agent tasks (170 turns), a naive replay-everything harness sends 876k tokens; Muster sends 355k — a 59.4% reduction, up to 62.7% on the longest threads. Reproduce with muster benchmark.
| scenario | turns | naive | muster | reduction |
|---|---|---|---|---|
| codebase-refactor-20 | 21 | 84.6k | 42.7k | 49.6% |
| incident-triage-30 | 31 | 144.9k | 60.5k | 58.2% |
| erp-data-audit-40 | 41 | 205.5k | 79.5k | 61.3% |
| research-synthesis-25 | 26 | 160.0k | 67.7k | 57.7% |
| long-support-thread-50 | 51 | 280.8k | 104.9k | 62.7% |
| AGGREGATE | 170 | 875.8k | 355.2k | 59.4% |
Positioning
07 What Muster is, and what it is not.
Muster is a harness for governed execution. It can work alongside agent frameworks, workflow graphs, coding agents, MCP servers, and business apps.
Just a chatbot wrapper
The value is not the prompt box; it is scoped memory, token accounting, policy, and verification around the run.
Only a workflow graph
Flows exist, but the wedge is the harness around runs, tools, memory, providers, and learning.
Uncontrolled autonomy
Capability packs, MCP policy, and channel setup make boundaries explicit before tools act.
Provider-flexible
Route through cloud APIs, Gemini, open-source/self-hosted servers, aggregators, Pi, private endpoints, and optional CLI-auth runtimes without moving governance into the provider.
Integration-aware
Muster separates setup plans, readiness checks, and executable adapters so teams can see what is configured before an agent acts.
Built for contributors
Good first issues include docs, examples, Frappe packs, provider adapters, MCP workflows, evals, and demo assets.
Star it. Try it. Break it. Improve it.
Muster is open source and early. The useful next contributions are concrete: examples, demo videos, Frappe packs, provider adapters, MCP auth tests, browser automation recipes, and eval suites.
pnpm --package=@musterhq/cli dlx muster demo