Using taskflow on Hermes Agent
Install taskflow as an MCP server on Hermes Agent and orchestrate multi-phase workflows with isolated Hermes subagents.
Taskflow 1.0.0 documentation. These instructions use exact 1.0.0 package pins. Check GitHub Releases for publication and the acceptance scoreboard for tested scope and limitations. The Control Plane and WebUI are implemented; control binaries are bundled in
taskflow-mcp-core.
Hermes Agent is Nous Research's open-source coding agent (CLI, TUI, desktop, messaging gateways). Taskflow's published path is its MCP server: Hermes discovers taskflow_* tools and each agent phase spawns an isolated hermes chat -q -Q child.
The full reference lives in docs/hermes-mcp.md in the repo.
Install
Published MCP package (after npm publish)
hermes mcp add taskflow --command npx --args -y -p hermes-taskflow@1.0.0 hermes-taskflow-mcpOr paste into ~/.hermes/config.yaml / $HERMES_HOME/config.yaml:
mcp_servers:
taskflow:
command: "npx"
args: ["-y", "-p", "hermes-taskflow@1.0.0", "hermes-taskflow-mcp"]
env:
# Required for mutating agent phases (terminal / file write / coding).
# PI_TASKFLOW_HERMES_UNSAFE_YOLO: "1" # required for mutating agent phases
timeout: 600Requires Node.js ≥ 22.19.0. Restart Hermes so MCP tools are rediscovered as mcp_taskflow_*.
Local monorepo dogfood (before npm publish)
pnpm install
pnpm --filter taskflow-core build
pnpm --filter taskflow-mcp-core build
pnpm --filter taskflow-hosts build
pnpm --filter hermes-taskflow build
# Point MCP at the compiled bin:
# command: node
# args: ["/abs/path/to/packages/hermes-taskflow/dist/mcp/bin.js"]
# env: { PI_TASKFLOW_HERMES_UNSAFE_YOLO: "1" }Optional skill:
cp -R packages/hermes-taskflow/plugin/skills/taskflow ~/.hermes/skills/taskflowVerify the install
hermes mcp list
hermes mcp test taskflowIn a Hermes session, tools should appear as mcp_taskflow_list, mcp_taskflow_verify, mcp_taskflow_run, …
Run your first flow
Just ask
> List my saved taskflows.
> Verify this flow, then run it: {name:"review-changes", phases:[...]}
> Run the "review-changes" taskflow with dir set to src.Call tools directly
Prefer mcp_taskflow_plan or mcp_taskflow_verify before any spend, then mcp_taskflow_run.
Permissions & safety
| Phase kind | Behavior |
|---|---|
| Child isolation | ephemeral HERMES_HOME with non-secret model/fallback routing, a routed-provider-only inference auth.json, and provider-only dotenv; no parent MCP/skills/memory/rules; RO plugin only on RO children |
| Read-only tool whitelist | local-read → taskflow_readonly_files; else model-only; PI_TASKFLOW_HERMES_READONLY_WEB=1 → web,search (never attach writable file) |
| Mutating / default-capable | Requires PI_TASKFLOW_HERMES_UNSAFE_YOLO=1 → child --yolo |
| Budgeted flows | Rejected — Hermes quiet mode does not report token/cost |
Install footgun: write MCP env: in config.yaml. Do not pass --env after bare CLI args in a way that lands in the node argv list.
Optional env:
PI_TASKFLOW_HERMES_UNSAFE_YOLO(mutating only; leave unset for verify/script)PI_TASKFLOW_HERMES_READONLY_WEB(opt-in network on RO)PI_TASKFLOW_HERMES_MAX_TURNS(default64)PI_TASKFLOW_HERMES_BIN(defaulthermes)HERMES_HOMEas the parent source for model routing and allowlisted provider auth; every child receives a separate ephemeral home
How subagents run
Each agent phase is:
hermes chat -q "<system + task>" -Q --source tool -t <toolsets> --max-turns N [--in cwd] [-m model] [--yolo]Quiet mode prints answer text on stdout and session_id metadata on stderr. The runner preserves stdout, including blank lines and answer text that begins with session_id:.
Next
Last updated on
Was this helpful?
Help us improve the docs or ask a question in the community.