Agent tools (Mezair MCP)#

Every saved local chat gets a set of Mezair tools over the Model Context Protocol: ask you a structured question, publish a review into the thread, read a dependency audit, start a workflow, and more. The same tools work across providers, so a review published by Codex and one published by Devin look and behave the same.

How the endpoint works#

• The main process hosts a loopback Streamable HTTP MCP endpoint. It only listens on 127.0.0.1.
• Each session gets its own bearer token. A token is scoped to that session and dies with it.
• Codex-family sessions (Codex, OpenAI, Azure, LLMGateway) receive the endpoint through per-launch configuration overrides. Cursor and Devin receive it through their ACP mcpServers setting, over HTTP or a small stdio shim.
• Repository tools check that the repository belongs to the session's workspace before they do anything. An agent can't point a Mezair tool at a checkout outside its scope.

Your permission mode still applies. Tools that change state go through the same approvals as the rest of the session.

Tool groups#

Group	Tools	What it's for
Questions	question_ask	Ask you a structured question with suggested answers. The agent waits for your answer and receives it as data. Works for every provider and permission mode.
Reviews	review_context, review_publish, review_patch	Gather the diff and publish an inline review: chapters, risks and findings, updated in place as the review progresses.
Findings	findings_publish, findings_patch	Publish a findings list with severity, file and line, ready to hand back as tasks.
Security	security_audit_run, security_findings	Run the repository's security audit and read its findings.
Dependencies	deps_list, deps_audit, deps_licenses, deps_sbom	List packages, check advisories and licences, and produce an SBOM.
CI/CD	cicd_analyze	Read a pipeline and publish it as a graph with slow or failing legs called out.
Data	data_show	Show a table or chart inline.
Diagrams	diagram_show, diagram_list, diagram_read, diagram_update	Draw and revise diagrams in the thread.
Artifacts	artifact_create, artifact_list, artifact_read, artifact_update	Create durable documents. A newer version supersedes the older one in place.
Sketches	sketch_publish, sketch_show, sketch_update, sketch_patch, sketch_snapshot, sketch_keep	Quick sandboxed visual sketches you can keep as artifacts.
Surfaces	surface_read, surface_action, surface_comments, surface_complete	Read a published surface, respond to your comments and mark it complete.
Delegation	delegation_discover, delegation_propose, delegation_read, delegation_stop	Propose work for helper agents, follow it and stop it. Delegated results come back for review before they touch your checkout.
Workflows	workflow_list, workflow_read, workflow_start	Find and start a reusable workflow.
Factories	factory_list, factory_read, factory_draft, factory_preview, factory_save, factory_activate, factory_control, factory_runs, factory_run_read, factory_run_control	Draft and operate software factories. Activation still needs your confirmation.
Browser	browser_navigate, browser_click, browser_type, browser_get_text, browser_get_html, browser_evaluate, browser_screenshot, browser_status	Drive Mezair's built-in browser for local previews.
Skills	mezair_skill_get	Load a Mezair skill's instructions on demand.

The catalogue is defined in the app's source under mezair-app/src/main/agent-tools/ and mezair-app/src/main/surfaces/kinds/. To add a surface kind, see Extending Mezair.

What agents can't do with these tools#

• Reach another workspace's repositories.
• Read your provider credentials, Sync tokens or share URLs. Share links are bearer credentials and never enter model context.
• Approve their own requests. Approvals come from you, in the thread or from a companion device.