Tools for agents, inspected
ActionMesh is a workflow studio that runs entirely in your browser. Design tool calls, validate payloads against schemas, simulate failures, and inspect every effect - before a single real credential exists anywhere near the run.
How it works
Build the workflow from the tool catalog: steps, JSON inputs, and a failure mode per step - stop, continue, or retry once. No account is needed to design anything.
Contract tests throw fixtures at every tool schema, valid and invalid. Simulation runs the whole plan and records what would have happened, effect by effect.
Policies evaluate least privilege against the plan before a run is allowed. Every simulated effect lands in the run inspector with secrets scrubbed from the logs.
Inside a run
Three moments from one workflow, exactly as the studio records them: composed, simulated, and stopped at the policy gate. Nothing here is a mock of a mock - each scene mirrors a real view inside the app.
Tool catalog
A searchable catalog of tools across everyday toolkits - Gmail, Calendar, Slack, Drive, and more. Each action is labeled Read, Write, or Destructive, with its input schema attached, so the blast radius is visible before a step exists.
Filter the catalog by text or by risk level. Looking for everything destructive is one click - which is exactly the point.
Every tool ships a JSON schema for its payload. Validation happens against the schema, not against vibes.
Three levels, displayed everywhere the tool appears - catalog, builder, tests, policies, and traces.
Workflow builder
Pick tools, write their input JSON, and choose what happens when a step fails. Templates and an approvals inbox cover the common flows; the plan is data you can inspect, not a black box.
Stop the workflow, continue past the failure, or retry once. The choice is explicit and visible on every step.
Common flows like daily issue triage ship as templates you can instantiate, edit, and simulate.
Steps that write or destroy can require per-step approval. The approval inbox is part of the run, not an afterthought.
Environments and connections
Environments scope what a workflow may touch - simulation, or production-locked with an allowlist of toolkits. Connections record that an account was connected outside the studio; no token is ever seen or stored here, and a step will not simulate until its toolkit has an active, unexpired connection record.
Simulation environments run anything; production-locked environments constrain runs to an explicit toolkit list.
You record that a Gmail or Slack account was connected elsewhere. The studio sees a claim with an expiry - never a credential.
Expired connection records stop the steps that depend on them, and the blocked run leaves a trace.
Contract tests and triggers
Attach fixtures to any tool and declare the expected result - valid or invalid. Run the suite and read the failures. Triggers get the same treatment: build a local subscription for a Gmail, Calendar, or Slack event and test it against sample payloads before subscribing to anything real.
A test pairs a fixture payload with an expectation. The suite reports every mismatch with the tool and fixture named.
Filter JSON plus a sample event, evaluated locally. Nothing subscribes to a real source.
Daily, weekday, or weekly schedules re-run the current workflow. Nothing fires while the tab is closed; a due-run check catches up and writes real traces.
Run inspector
Each simulated run logs every step: the tool, the effect it would have had, approvals, retries, and failures - with inputs matching token, password, secret, or key scrubbed from the log.
Totals across runs make drift visible: how many simulations passed, which steps fail most, which tools get used.
Policy engine
A policy names the maximum risk level, the allowed toolkits, the maximum step count, and whether writes need per-step approval. Evaluate the current workflow against every policy and read exactly where it fails.
A policy capped at Read rejects any workflow containing a write or destructive step. No exceptions, no silent downgrade.
A sandbox that allows Gmail and Slack does not quietly permit Drive. The evaluation says which step violated which rule.
Cap the number of steps a workflow may have. Long chains of effects are a risk surface, and the policy says so.
Local-first
ActionMesh has no server, no account system, and no analytics. No external API call is ever made - the studio proves schemas, approvals, retries, idempotency, and logs before real credentials exist. Workspaces live in local storage, export as JSON, and delete in one click.