switchyard
Threnody is a local-first MCP meta-harness for planning, routing, and coordinating host-native AI coding workflows, with receipts, learning, and memory.
Plan and decompose tasks into subtasks with model tiers, dependencies, and parallel waves (
plan_task,decompose_task,fleet_plan,start_task).Route tasks via heuristics or planner, validate routing guards, and manage bypass exceptions (
route_task,validate_routing_guard,routing_exception_*).Execute delegated subtasks through utility providers and run/monitor/resume/stop them (
execute_subtask,list_subtasks,stop_subtask,resume_subtask).Orchestrate swarms and host-native waves, report wave/workflow completion, expand plans, and resume failed swarms (
execute_swarm,report_host_wave,report_workflow_result,expand_host_plan,resume_swarm_*).Use task packs, workflow blueprints, and preview/approve outside-workspace writes (
plan_task_pack,workflow_blueprint_*,apply_preview).Inspect status, providers, spend, quality, run receipts, write audits, swarms, tasks (
inspect_*,check_providers).Store and search cross-session memory, cache results, record outcomes, tune settings, and manage learning/approval queues (
memory_*,cache_*,record_outcome,tune_show,agent_queue_*,learning_*).
Routes coding tasks to Amazon Q (Kiro) CLI as a secondary adapter, expanding provider options for task execution.
Routes coding tasks to GitHub Copilot CLI for execution, leveraging its free gpt-5-mini tier for low-complexity work.
Routes coding tasks to JetBrains Junie CLI, with medium-tier auto-routing by default for balanced performance.
Routes coding tasks to OpenAI Codex CLI, enabling host shell and execution capabilities for AI-assisted development.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@switchyard refactor the login module and run tests on it"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
What is Threnody?
A local-first MCP meta-harness for developer workflows — small, auditable, and host-native, not a hosted swarm platform. Register it in Claude Code, Copilot CLI, Codex, Cursor, Junie, or OpenCode: Threnody plans and routes in MCP; the host shell executes via host_spawn / host_spawn_waves (Agent or Task subagents). When a handoff includes host_spawn_waves, spawn subagents — do not substitute direct edits on planned files.
execute_subtask is utility delegation only (opt-in): OpenCode, Aider, and local loopback endpoints — never other host CLIs. Same-host work returns HostNativeRequired with a spawn payload; host→host delegation returns HostDelegationBlocked. Claude Code is a router-only host by default.
On Claude Code, an opt-in mode (routing_policy.shells.claude-code.workflow_emit) emits tier-aware Dynamic Workflow scripts for fan-out plans — each agent() routes to its Threnody tier model, where a vanilla workflow runs every agent on the session model. Recurring shapes can be approved and saved as permanent /workflow commands.
For operators who want multi-agent coding without a permanent agent army, a second hosted control plane, or hidden coordination-token drift.
Related MCP server: Agent Squad Bridge
Install
Requires: Python 3.10+ and macOS or Linux for the supported installation
paths. The full install.sh workflow expects at least one host AI CLI (gh,
claude, codex, cursor-agent, junie, or opencode); the packaged MCP
server can start without one for setup and diagnostics.
Claude Code plugin marketplace (recommended) — bundles the MCP server and nine routing skills, no shell restart:
claude plugin marketplace add timjensgrossinger/threnody
claude plugin install threnody@threnodyMCP package via uvx or pip — works with any MCP-aware host:
claude mcp add threnody -- uvx threnody-mcp
pip install threnody-mcp
threnody-mcpThe package entry point is a local stdio server. The packaged server can start
for setup and diagnostics without a host CLI; THRENODY_ALLOW_NO_HOST=1 is
only used by the full install.sh workflow to bypass its host-CLI check.
Full CLI install (adds ghc/ghcs/ghce aliases, syncs routing
instructions; restart shell after):
curl -fsSL https://raw.githubusercontent.com/timjensgrossinger/threnody/main/install.sh | bash
# plugin-only (skips shell aliases): ... | bash -s -- --plugin-modeFor a configured power-user installation, run threnody settings to finish
setup. The packaged stdio entry point does not run the interactive wizard.
Provider terms: Threnody is not affiliated with or endorsed by any AI provider. Credentials stay in provider-native stores; configure auth in each host CLI. See docs/LEGAL.md.
Docs: plugin install · limitations · legal · architecture
Compliance posture (defaults)
With default config, Threnody matches Anthropic's intended MCP pattern for Claude Code:
Control | Default behavior |
Execution | Host runs work via Agent and direct edits — Threnody returns |
Router-only | Claude Code is a coordination anchor; not a subprocess backend |
Same-host | Returns |
Host→host delegation |
|
Utility delegation | Off by default; opt-in targets OpenCode, Aider, and local endpoints only |
Routing policy | Advisory by default — |
Operator opt-in risk: Enabling providers.router_only_allow_execution can subprocess Claude Code. With subscription OAuth that pattern is documented as high policy risk in docs/LEGAL.md. Verify your auth mode and provider terms before changing defaults.
How it works
Host shell (Claude / Copilot / Codex / Cursor / …)
→ start_task / route_task / plan_task returns next_action and host_spawn_waves for host-native work
→ host executes Agent or Task subagents, direct edits (spawned from host_spawn_waves)
→ utility delegation execute_subtask → OpenCode / Aider / local (opt-in)
→ swarm / learning execute_swarm (host_native default), memory_*, learning_*Call the start_task tool (or route_task/plan_task) from your MCP host shell — start_task returns a compact next_action and optional host_spawn_waves for host-native execution.
Threnody scores complexity → low / medium / high tier (no extra LLM call on the hot path). Pass
target_filestoroute_taskand the score also reflects the target code's security risk and size, rather than only the wording of the task; omit it and paths named in the task text are used. Security-sensitive work is floored tomedium, and a target holding a high-severity security defect tohigh.route_task/plan_taskreturn spawn metadata —host_spawnfor single-agent,host_spawn_wavesfor multi-step plans.The host runs the work — Claude Code uses Agent; other shells use Task.
Swarms or utility delegation —
execute_swarmreturns a host-native wave plan by default;execute_subtaskonly for utility backends when enabled.
Local-first: routing state, telemetry, and caches stay in local SQLite (~/.local/lib/threnody/); the MCP server talks to your host over stdio — no Threnody-hosted control plane. Outbound traffic comes only from the provider CLIs you invoke.
Features
Feature | What it does | |
🎯 | Tier routing | Heuristic complexity scoring + |
🧠 | Learning loop | Pattern tracking → draft agents → approval queue → plan-time context injection. No auto-promotion; conservative recurrence/quality/rework gates |
🐝 | Swarm orchestration |
|
⚡ | Dynamic Workflows | Opt-in (claude-code): fan-out plans emit a tier-aware Workflow script; recurring shapes export to permanent |
🧾 | Receipts and run cards |
|
🧩 | Task packs and blueprints | Curated packs ( |
🔬 | Static pre-scan | Deterministic AST + pattern scan (zero tokens) sharpens review tiering, injects verified-or-refuted leads, and grades reviewers against ground truth |
♻️ | Prior-review memory | An unchanged |
🪄 | Hybrid diagnose→implement | One read-only high-tier diagnosis produces a change-spec, cheaper implementers execute it. Emits tiers only; the discount is learned per work profile, never hardcoded |
✅ | Baseline-diff verify gate | Lint/type/test graded against the merge base, so pre-existing red never blocks and only real regressions get a fix pass (in-process, zero tokens on the host path) |
📚 | Repo beliefs | Each run leaves a free |
🪜 | Graded task ladder |
|
💾 | Cross-session memory |
|
🔌 | MCP-native | 53 published tools over stdio JSON-RPC; works with any MCP-compatible host |
📈 | Adaptive thresholds | EMA threshold learning from |
🛡️ | Write safety | Path validation, outside-workspace grant model + audit trail |
🔒 | Guarded routing | Optional coordination gate + Claude PreToolUse hooks ( |
Cross-CLI memory: all hosts share one SQLite store at ~/.local/lib/threnody/cache.db. Use global (no project_id), project (pass a stable absolute path, not "."), or task (explicit task_id) scopes. Do not store secrets — any connected host can read keys.
Adaptive routing: route_task returns a task_id; after work, call record_outcome(task_id=…, outcome=accepted|revised|rejected|reworked, actual_model=…) (also recorded as a proxy outcome score in the quality ledger). Enable per project with threnody tune set learning_enabled true --project ..
Project skills
Nine repo-local skills under skills/ guide MCP workflows from any host. install.sh installs them into provider-native roots (directory-style for Claude Code / Cursor / Codex; flat markdown for Copilot CLI / OpenCode).
install.sh also writes the five review-dimension agent definitions into each host's own definition directory, so a review fan-out does not repeat the same checklist in every agent prompt (prompt_economy.externalize_boilerplate). A definition you already have under the same name is never overwritten — your tuned reviewer is the better version of that file.
Skill | Use when |
Plan-only or plan-then-execute; waves vs swarm | |
| |
| |
| |
Default broad review swarm — one read-only agent per file plus synthesis | |
Deep review swarm — file × dimension fanout, optional | |
Claude Code tier-aware Dynamic Workflows; save | |
Contract-first parallel frontend + backend + API | |
Monitor opt-in utility |
Supported providers
Provider | Binary | Role |
Claude Code |
| Host (router-only) — executes via Agent / direct edits |
GitHub Copilot |
| Host — executes via Task |
OpenAI Codex |
| Host — Task execution |
Cursor |
| Host — Task execution |
OpenCode |
| Host + utility delegation target |
JetBrains Junie |
| Host / legacy paths |
Aider |
| Utility (opt-in delegation) |
Amazon Q / Kiro · Mistral Vibe · Blackbox |
| Secondary adapters / detect |
Windsurf |
| Detect only — never executes |
Live matrix: threnody inspect status --project . --details. Full table: docs/PROVIDER_COMPATIBILITY.md
See it in action
plan_task("add JWT auth with tests")
→ host_spawn_waves: [
{ "wave": 1, "agents": [{ "tool": "Agent", "tier": "medium", "target_files": ["src/auth.py"] }] },
{ "wave": 2, "agents": [{ "tool": "Agent", "tier": "low", "target_files": ["tests/test_auth.py"] }] }
]📋 Wave 1 — spawn Agent (host model) → src/auth.py
📋 Wave 2 — spawn Agent (host model) → tests/test_auth.pyOptional utility delegation (opt-in via providers.delegation_utilities_enabled):
execute_subtask(prompt="…", tier="low", provider_id="opencode") → OpenCode utility backend
execute_subtask(prompt="…", tier="medium") → HostNativeRequired + spawn payload
execute_subtask(provider_id="codex") → HostDelegationBlockedShell commands
ghc agent "implement JWT auth for the user service" # multi-agent waves
ghcs "how to list files recursively in python" # quick routed call
threnody inspect status --project . --details # provider readiness
threnody quality --since 7d # model quality ledger
threnody ladder run --tier low,medium,high # graded ground truth (spends tokens)
threnody ladder run --provider codex # grade another host's tier models; in Claude Code use /threnody-ladder
threnody ladder run --stale # re-grade only tiers whose model changed
threnody db learn status # durable learning journal + table counts
threnody db learn rebuild # rebuild the DB's learning tables from the journal
threnody-watch # live TUI monitorFull reference: docs/CLI.md
Documentation
Doc | Contents |
uvx, plugin marketplace, | |
All 53 MCP tool surfaces | |
Shell aliases and operator commands | |
Trust boundaries and local-first design | |
Safe starting config | |
Eval methodology and accuracy | |
Claude PreToolUse guard script | |
Host-native vs utility delegation | |
Threnody vs heavy swarm platforms | |
Beta scope, privacy, roadmap | |
Operator responsibilities | |
Common fixes |
Beta status
Public alpha v0.3.0-alpha.3 — MCP tool schemas may change between releases; pin a git tag for stability. macOS and Linux (zsh/bash); Windows not supported by the installer. See CHANGELOG.md.
Running tests
THRENODY_TEST_MODE=1 python3 -m pytest tests/ -q
THRENODY_TEST_MODE=1 python3 -m shared.routing_eval
python3 scripts/check_release_archive.pyUninstall
~/.local/lib/threnody/uninstall.sh [--purge-data]Legal
Threnody is an independent open-source project, not affiliated with or endorsed by Anthropic, OpenAI, GitHub, Google, Cursor, JetBrains, or any other provider named here. Provided "AS IS" under the Apache License 2.0 (no warranty). You are solely responsible for determining whether your routing patterns comply with each provider's current terms.
Operator responsibilities: docs/LEGAL.md · Third-party attributions: NOTICE
Built by @timjensgrossinger.
Available Tools
55 toolsagent_queue_approveC
Approve one pending approval-queue item for one project.
Inputs: queue_id (required), project_id (required), operator_id (required).
agent_queue_listC
List pending approval-queue items for one project.
Inputs: project_id (required), limit.
agent_queue_mergeC
Merge one pending approval-queue item into a canonical agent.
Inputs: queue_id (required), project_id (required), operator_id (required), canonical_agent_id (required), +1 more.
agent_queue_rejectC
Reject one pending approval-queue item for one project.
Inputs: queue_id (required), project_id (required), operator_id (required), reason.
apply_previewB
Approve or deny a pending outside-workspace file write preview created by execute_subtask.
Inputs: approve (required), preview_token (required).
approval_queue_approveD
Compatibility alias for agent_queue_approve.
Inputs: queue_id (required), project_id (required), operator_id (required).
approval_queue_listD
Compatibility alias for agent_queue_list.
Inputs: project_id (required), limit.
approval_queue_mergeD
Compatibility alias for agent_queue_merge.
Inputs: queue_id (required), project_id (required), operator_id (required), canonical_agent_id (required), +1 more.
approval_queue_rejectD
Compatibility alias for agent_queue_reject.
Inputs: queue_id (required), project_id (required), operator_id (required), reason.
cache_getA
Look up a cached result for a task.
Inputs: task (required).
cache_putC
Store a completed task result in the cache.
Inputs: task (required), model (required), result (required).
cache_statsB
Return cache statistics: total entries and breakdown by model.
Inputs: no named parameters.
check_providersA
List detected AI CLI providers with routeability, detection reason, model summary, and health status. Output is compact and secret-safe.
Inputs: no named parameters.
decompose_taskA
Alias for plan_task. Preferred entry point for multi-file or multi-concern tasks.
Calls the planner to decompose a task into independent subtasks with model tier assignments and dependency waves.
Returns:
analysis: planner reasoning
subtasks: list with id, description, tier, model, depends_on
waves: parallel execution groups
strategy: parallel | sequential | dag
Use this instead of route_task whenever the task spans more than one file, module, or concern. Spawn one agent per subtask; run waves in order.
Inputs: task (required), cwd.
execute_subtaskA
Execute a prompt via the cheapest available AI CLI provider.
Routes to the cheapest model for the given tier across all installed CLI tools (GitHub Copilot, Codex, Cursor, and others). Falls back to next cheapest on failure.
When target_file is provided, writes the result directly to that path and returns file metadata. This is the preferred way to create files for low-tier subtasks — saves tokens by avoiding round-trip through the main agent.
Surgical edit modes (set mode=): rewrite (full-file injection + length-ratio guard), blocks (Aider-style SEARCH/REPLACE, token-efficient), patch (unified diff).
Returns:
result: the model's response text
provider: which CLI tool was used
model: which model handled it
tier: the tier that was requested
fallback_used: whether a fallback provider was needed
file_written: path written to (when target_file is set)
lines_written: line count of written file
diff: unified diff showing changes (when target_file is set)
change_type: 'created', 'modified', or 'unchanged'
lines_added: number of lines added
lines_removed: number of lines removed
Inputs: prompt (required), mode, tier, effort, +8 more.
execute_swarmC
Start a swarm run; returns an immediate run contract (swarm_id, wave summary, cost estimate) or a budget preview.
Inputs: task (required), topology, task_spec, task_text, +6 more.
expand_host_planC
Expand a host-native swarm/plan run with additional file-scoped subtasks. Returns pending host_spawn_waves for discovered files not yet assigned.
Inputs: discovered_files (required), reason, run_id, swarm_id, +3 more.
fleet_planA
Plan a task AND format it for /fleet execution.
Calls the LLM planner to decompose the task, then produces ready-to-run /fleet command strings — one per wave, respecting dependency order.
Returns:
plan: full plan (same as plan_task)
fleet_waves: list of wave objects, each with: wave_number: int command: '/fleet "[tier] subtask1" "[tier] subtask2"' agents: list of {tier, model, prompt}
execution_note: how to run the waves
cache_hit: bool
Use this when you want copilot-router's model intelligence combined with /fleet's true parallel execution. Run wave 1 command, wait, run wave 2, etc.
Inputs: task (required).
inspect_qualityA
Return the granular model quality ledger: per-(model x effort x dimension x sub_dimension) average score (0-10), sample count, findings/judge breakdown, and approximate escalation rate for a time window. e.g. opus | high | security/sql-injection.
Inputs: since.
inspect_run_receiptC
Return an operator run receipt by run_id/swarm_id, as JSON, Markdown, or local HTML.
Inputs: format, run_id, swarm_id.
inspect_spendB
Return aggregated spend and savings telemetry from delegated subtasks (est_cost_usd vs counterfactual) and persisted cost receipts for a time window.
Inputs: since.
inspect_statusB
Return a compact readiness/status snapshot for one project, including enabled features, current limits, and pending approvals.
Inputs: project_id.
inspect_swarmB
Return compact swarm run summary from SQLite (status, progress, host metadata).
Inputs: swarm_id (required), run_id.
inspect_taskA
Return structured provider/model/tier telemetry and fallback/speculation flags for a previously routed task.
Inputs: task_id (required).
inspect_write_auditB
Return recent out-of-workspace write audit log entries.
Inputs: limit.
learning_agent_summaryA
Get summary of all learned agents by status (active, pending, rejected).
Returns compact agent list with description, lane, pattern hash, and status. Sensitive data (tokens, secrets) is filtered out.
Inputs: no named parameters.
learning_audit_logB
Get audit trail for agent creation, approval, and registration events.
Returns event stream with timestamps and operator identity. Sensitive fields (tokens, API keys) are filtered out.
Inputs: limit, agent_id.
learning_outcome_statsA
Get outcome distribution snapshot over 1-hour recent window grouped by tier and model.
Returns: outcome counts (accepted, revised, rejected, reworked) per tier:model combination, coverage percentage, and window timestamps. Aggregates are computed in background and retrieved from memory. Use for operator observability into routing quality by model.
Inputs: no named parameters.
learning_pattern_healthA
Get health metrics for the pattern tracking system.
Reveals: total patterns tracked, mature patterns ready for drafting, patterns awaiting proof, draft proposals in approval queue, and active agent count. Use to monitor learning loop maturity.
Inputs: project_id.
list_subtasksA
Return structured status of currently running and recently completed execute_subtask calls.
active: tasks currently executing (show elapsed time, model, prompt excerpt, target file). recent: last 10 completed or failed tasks this session.
Use this to monitor parallel execute_subtask calls — similar to /tasks for background agents.
Inputs: no named parameters.
list_task_packsA
List curated task packs for cheap repeatable planning.
Inputs: no named parameters.
memory_deleteC
Hard-delete one memory value from an explicit scope.
Inputs: key (required), scope (required), task_id, project_id.
memory_getC
Fetch one full memory envelope from an explicit scope.
Inputs: key (required), scope (required), task_id, project_id.
memory_listB
List keys and lightweight metadata for one explicit memory scope.
Inputs: scope (required), task_id, project_id.
memory_searchA
Search memory values via local FTS5 (no embeddings).
Inputs: query (required), limit, scope, project_id.
memory_setC
Store or overwrite one memory value in an explicit scope.
Inputs: key (required), scope (required), value (required), task_id, +1 more.
plan_taskA
Ask the planner to analyse a coding task, decompose it into subtasks, and assign each subtask a model tier.
Returns an execution plan with:
analysis: the planner's reasoning
subtasks: list with id, description, tier, model, depends_on
waves: groups of subtask IDs that run in parallel
strategy: parallel | sequential | dag
Spawn one agent per subtask. Run waves in order — all subtasks within a wave run in parallel.
Inputs: task (required), cwd.
plan_task_packC
Plan a task using a curated task-pack preset.
Inputs: pack (required), task (required), cwd, max_agents.
record_outcomeC
Record an explicit routed-task outcome and persist the latest task snapshot. operator_id must match the authenticated caller when provided; omitted values are stored as anonymous.
Inputs: outcome (required), task_id (required), note, operator_id.
report_host_swarm_completeC
Terminal shortcut for host-native runs — equivalent to report_host_wave with terminal=true. Requires outcome.
Inputs: wave (required), agents (required), outcome (required), run_id, +2 more.
report_host_waveA
Report completion of one host-native swarm/plan wave. Call after spawning and finishing each host_spawn_waves wave. Pass workspace_root from the handoff (learning_report_contract) and per-agent output_excerpt for learning quality. Set terminal=true on the last wave with outcome=accepted|revised|reworked|rejected.
Inputs: wave (required), agents (required), run_id, outcome, +6 more.
report_workflow_resultA
Report the result of a Threnody-emitted Dynamic Workflow run (workflow_emit). Pass the workflow_name from the handoff and the agents[] array the workflow returned (each entry: id/label, tier, model, result{summary,findings,success}). Threnody records per-agent learning telemetry and, once the orchestration shape recurs across successful runs, enqueues it as an approval-gated draft you can approve to save as a permanent /workflow command.
Inputs: agents (required), workflow_name (required), run_id, consensus, +1 more.
resume_subtaskA
Send SIGCONT to a stopped subtask, resuming its execution. Use after stop_subtask. macOS/Linux only.
Inputs: task_id (required).
resume_swarm_confirmC
Resume a failed swarm from a chosen coordinator checkpoint using a new swarm_id.
Inputs: failed_swarm_id (required), checkpoint_index (required), plan_revision.
resume_swarm_inspectB
List compact coordinator checkpoints available for resuming a failed swarm.
Inputs: failed_swarm_id (required), plan_revision.
route_taskA
Quick heuristic classification of a task — no LLM call. Returns model, score, reason, agents. Use for simple tasks or when speed matters more than accuracy.
Inputs: task (required), cwd.
routing_exception_addA
Add a routing bypass rule so that matching tasks skip validate_routing_guard.
exception_type must be one of: skill, filetype, project, command, caller, path. pattern supports glob wildcards (e.g. 'tgsd-*', '.md', '/home/user/notes').
Examples: routing_exception_add(exception_type='skill', pattern='auto-time') routing_exception_add(exception_type='skill', pattern='tgsd-*') routing_exception_add(exception_type='filetype', pattern='.md') routing_exception_add(exception_type='project', pattern='/home/me/notes') routing_exception_add(exception_type='command', pattern='Write') routing_exception_add(exception_type='caller', pattern='github-copilot') routing_exception_add(exception_type='path', pattern='/tmp/')
Inputs: pattern (required), exception_type (required), note.
routing_exception_listA
List all active routing bypass rules (from the DB; static config.yaml entries are separate).
Inputs: no named parameters.
routing_exception_removeA
Remove a routing bypass rule by type and pattern.
Inputs: pattern (required), exception_type (required).
start_taskA
Start a host-native task handoff with modes implement, review, investigate.
implement: use host-native heuristics for simple tasks and return host_spawn_waves for the host to execute; for complex tasks return a planner handoff (no provider subprocesses launched). review: reuse read-only REVIEW fanout and return host_spawn_waves suitable for review (read-only). investigate: profile repository and return a machine-readable ProjectProfile; read-only.
Returns machine-readable next_action, profile, warnings, selected tier/model/provider metadata, and host_native handoff when applicable. Does not launch provider subprocesses.
Inputs: task (required), cwd, mode, workspace_root.
stop_subtaskA
Send SIGSTOP to a running subtask, pausing its execution. Use list_subtasks to find the task_id. Resume with resume_subtask. macOS/Linux only.
Inputs: task_id (required).
tune_showC
Show persisted operator-facing tuning controls for one project.
Inputs: project_id (required), key.
validate_routing_guardA
Validate whether a direct Edit or Write tool call is allowed for the current routed task context. Intended for Claude Code PreToolUse hooks.
Inputs: cwd, skill, tool_name, target_file.
workflow_blueprint_exportB
Export a successful host-native run receipt into a replayable workflow blueprint.
Inputs: name, run_id, swarm_id.
workflow_blueprint_runB
Replay a saved workflow blueprint with optional string replacements; no planner call.
Inputs: name (required), inputs.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 55 tools
Multiple tools are direct aliases (e.g., approval_queue_list vs agent_queue_list), and several planning tools (plan_task, decompose_task, fleet_plan) have overlapping purposes with only subtle differences, creating confusion for agent selection.
Most tools use snake_case but follow inconsistent patterns: some are verb_noun (inspect_status, list_subtasks), others are noun_verb (tune_show, memory_list), and some are noun_noun_verb (routing_exception_add). No single convention dominates.
With 55 tools, the server is overly large. Many tools are redundant (aliases) or highly specific (e.g., resume_swarm_inspect, apply_preview), which could be consolidated. A more focused set would improve usability.
The server covers a wide range of agent management tasks: planning, execution, learning, caching, routing, workflow, and inspection. Minor gaps may exist (e.g., no generic project CRUD), but core operations are well represented.
Maintenance
Related MCP Connectors
AI routing, memory, guardrails, and governance. Routes across Claude, GPT, Gemini.
Multiplayer for the coding agents your team already uses: shared tasks, decisions, skills and tools.
AI work orchestration for plans, tasks, teams, and coding-agent dispatch.
Command center for developers coordinating AI coding agents across projects.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceRoutes coding tasks to local and remote LLMs with intelligent cost-quality optimization, supports benchmarking and code search.43ISC
- AlicenseNot gradedqualityCmaintenanceEnables multiple AI coding CLIs (Claude Code, Gemini/Antigravity, Codex, and OpenCode) to collaborate as a coordinated team by routing cross-agent prompts, sharing messages and review tickets, tracking tasks on a shared store, and isolating each agent in its own Git worktree with turn-budget safeguards—all inspectable and steerable from a local web dashboard.MIT
- FlicenseNot gradedqualityAmaintenanceEnables a primary AI coding assistant to delegate tasks to multiple specialized CLI agents (such as Gemini, Claude, Codex, and Cursor) in parallel from a single prompt, sharing project context across all of them. It aggregates their outputs into one synthesized response while relying on the user's existing CLI tools and subscriptions.-
- FlicenseNot gradedqualityAmaintenanceEnables users to delegate tasks from a single prompt in their IDE to multiple specialized AI coding assistants such as Gemini, Claude, Codex, and Cursor, which run in parallel with shared project context. Aggregates and synthesizes their responses into one unified answer, using existing CLI tools and API subscriptions without extra cost.-