Muster — the AI agent harness you can audit
Open-source agent runtime with a token-waste ledger, leak-proof scoped memory, eval-gated learning, and integrity verification. Works with Claude, OpenAI, Gemini, Grok, Kimi, DeepSeek, and 20+ providers. TypeScript, MIT, self-hosted.
Self-improving agents are easy. Provably governed agents are Muster: every memory scoped, every skill eval-gated, every token on a ledger. Does your agent pass muster?
pnpm dlx @musterhq/cli init && muster demoSee it work — muster demo
One command provisions a throwaway workspace and a local model service, then runs the full governed pipeline: scoped-memory recall → token ledger → integrity check.
muster 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).
run model in out est cost$ waste session
----------------------------------------------------------------------------------------------
287bde9c-eb19- demo/demo-model 38 17 ~ - - -
653b434a-0924- demo/demo-model 7 18 ~ - - -
totals by model runs in out cost$ waste-runs
--------------------------------------------------------------------------------
demo/demo-model 2 45 35 - 0
integrity check at 2026-06-12: OK
store lines corrupt
episodes 2 0
memory 3 0
tokens 2 0Proof, not promises — muster benchmark
The Token Waste Index measures what Muster's immutable-transcript renderer and never-wedge compactor actually save versus a naive replay-everything harness. Deterministic — no model calls, fully reproducible.
scenario turns naive muster reduction replay-overhead
--------------------------------------------------------------------------------------
codebase-refactor-20 21 82.6k 40.7k 50.7% 90.5%
incident-triage-30 31 140.4k 56.2k 59.9% 93.6%
erp-data-audit-40 41 197.8k 72.4k 63.4% 95.1%
research-synthesis-25 26 156.8k 64.6k 58.8% 92.3%
long-support-thread-50 51 268.8k 93.8k 65.1% 96.1%
--------------------------------------------------------------------------------------
AGGREGATE 170 846.4k 327.9k 61.3% 94.2%~61% fewer tokens on long agent sessions, and the saving grows with session length. Full methodology + table: benchmark/RESULTS.md.
Features
| Token ledger | Every run recorded; replay-waste flagged with the exact ratio. muster tokens |
| Scoped memory | Tenant / workspace / user / role / session lanes. Cross-user leakage is a failing test, not a hope. |
| Eval-gated skills | Skills promote only after an eval suite converges — no self-certified learning. |
| Integrity verify | Corruption, duplicate runs, silent model drift, stale-narrative poisoning. muster verify |
| Never-wedge compactor | A session can always take a turn — no compaction deadlock. |
| Recursive self-test | muster evolve runs real tasks, adjudicates against evidence, converges. |
| Flow engine | Tool/agent/gate steps, preflight, durable runs, replay/diff, flow loop --cron. |
| One gateway, every chat app | Telegram · Slack · Discord · WhatsApp · Google Chat · Teams + a zero-dep web client. |
| MCP client | Per-server isolation, circuit breakers, capped results. |
| 20+ providers | Claude (Fable 5), OpenAI, Gemini, Grok, Kimi, DeepSeek, Groq, vLLM… zero lock-in. |
| Pulse scheduler | Heartbeat that feels alive at ~5% of the token cost — zero-LLM preflight + daily budget. |
| Pull-based subagents | Durable run store, exactly-once results, no zombie processes. |
Everyday commands
muster provider add anthropic # or kimi / add-openai-compatible <any-url>
muster run "where do we deploy?" # governed run: memory recall + ledger + evidence
muster tokens # per-run cost table, replay-waste flags
muster verify # store integrity
muster sessions search "leave balance" # FTS search across past sessions
muster evolve evolve-suites/core-capabilities.json # recursive self-test
muster pulse add "0 9 * * 1-5" --kind task --prompt "summarize open work"
muster benchmark # the Token Waste Index, liveEverything renders plain-text tables in your terminal. No web dashboard required.
Architecture
prompt ──> router ──> [agent rules + recalled scoped memory] ──> runtime
├─ Pi SDK (embedded)
scoped memory lanes ├─ Claude Code CLI
tenant/workspace/user/role/session ├─ Codex CLI
│ └─ any HTTP provider
▼
episode store ──> token ledger ──> feedback adjudication ──> eval fixtures
│ │ │
└──── muster verify (integrity) ◄──── muster evolve (self-test loop)Built on the pi.dev coding-agent SDK as bedrock — embedded sessions, tools, and TUI — with the governance layer Muster adds on top.
How it compares
| Muster | OpenClaw | Hermes | crewAI | |
|---|---|---|---|---|
| Token ledger + waste detection | ||||
| Scoped memory (leak = CI failure) | partial | (single MEMORY.md) | ||
| Eval-gated learning | (promotes on use) | |||
| Governed fallback (evidence, never silent) | (#65646) | |||
| Session integrity verification | (#75235) | (#5563) | ||
| Channels & web embeds (one governed envelope) | Slack, Discord, Telegram, WhatsApp, GChat, Teams, any web app | 20+ bespoke | ||
| Maturity / ecosystem | v0 | huge | large | large |
Honest table: they have breadth and ecosystems we don't (yet). We have the governance core they demonstrably lack — each above links to their own issue tracker.
Use cases
- AI agents for business systems: the Frappe/ERPNext capability pack ships permission-scoped tools where every action executes as the real user — see
capability-packs/frappe/. Built from a production deployment serving thousands of employees. - Cost-controlled agent fleets: per-profile ledgers, per-flow budgets, waste alerts.
- Regulated / BFSI / air-gapped: local models (vLLM, SGLang), no cloud required, full audit trail.
- Agent CI:
muster evolveas a pipeline gate — your agent's behavior is regression-tested like code.
Keywords
AI agent framework · LLM agent harness · agent memory · token cost tracking · agent observability · eval-driven development · agentic workflows · Claude agent SDK · OpenAI agents · self-hosted AI agent · AI governance · agent audit trail · ERPNext AI · Frappe AI assistant · multi-provider LLM routing
Maturity — v0.1, feature-complete core
Muster is v0.1: the governed core is feature-complete and test-covered, the public API may still shift before 1.0. (For reference, the largest open agent frameworks still version in the v0.x / date-based range — v0.x here means "pre-1.0 stability," not "incomplete.")
Mapped against the mid-2026 production bar for agent harnesses:
| Production-bar capability | Muster |
|---|---|
| MCP client | per-server isolation, circuit breakers, capped results |
| Eval-gated learning | skills promote only through a converged suite |
| Per-run cost / token tracking | token ledger with replay-waste detection |
| Layered, deterministic permissions | scoped-memory lanes + hook bus, leak = failing test |
| Memory: working / episodic / scoped | scoped lanes + SQLite session store (FTS5) |
| Strategic (not reactive) compaction | immutable transcript renderer + never-wedge compactor |
| One protocol for CLI / desktop / web | JSON-RPC gateway with ledger.tick live cost |
| OpenTelemetry tracing | planned |
| Desktop apps | Tauri over the RPC protocol |
Claude Fable 5 ready: the Anthropic preset defaults to claude-fable-5 (1M context, adaptive thinking via effort). The token ledger and scoped tool exposure align with Fable 5's deferred-tool-loading and task-budget direction. First-class stop_reason: "refusal" handling is on the roadmap.
Independence: Muster is operator-governed and MIT — no foundation, no single-vendor entanglement. You run it, you audit it.
Next: OTEL tracing, Tauri desktop apps, channel approval round-trips, npm publish, and a Token Waste Index benchmark. See docs/SDLC_KANBAN.md and docs/FEATURE_PARITY_PLAN.md.
License
MIT. Open source, community-driven. Contributions welcome — start with good first issue.