MCP tools
What your AI agent can do with Grace.
Grace's tools let your AI agent consult the right instructions, then check its work. You explain what you want to do; it uses the tools it needs.
MCP is how your AI agent connects to Grace. If the connection is not ready yet, follow Connect a project.
The tools
| When are they needed? | What the AI agent can do |
|---|---|
| Before starting a task | Consult instructions from the talents it chose for that task. |
| An instruction needs explaining | Read that instruction in more detail. |
| The work is ready for review | Get the verification rules, then apply them to its changes. |
| An architecture decision needs comparing | Present three distinct options with diagrams, the same criteria, and a recommendation that leaves the decision to you. |
| Your organization requires a record of the review | Record the results of checking Grace talents and Custom Talents. |
| You ask to create a lasting organization rule | Help you write a coherent rule, have you approve its exact contents, then save an unpublished draft. This is reserved for owners and admins. |
| You have approved new talents | Add those talents to the project. |
| A talent seems incomplete or incompatible | Identify missing prerequisites and conflicts. |
| You need to choose your first talents | Examine the project and propose talents for your approval. |
| You want to report a problem or suggest an idea | Prepare feedback limited to useful information, show you exactly what would be sent, then wait for your explicit approval. |
Remember: Grace provides the instructions. The AI agent does the work and the review. Obtaining a list of rules does not mean the changes have been checked.
The agent chooses talents relevant to your request, then receives the detailed procedure when it starts the review. This review must focus on your changes. It does not replace project tests, builds, or lint checks after the last fixes. If the agent cannot use an independent reviewer, it must disclose this limitation.
Recognize Grace in the conversation
Grace instructions ask your agent to show short messages at useful stages, without requiring you to open tool details:
| Milestone | Expected information |
|---|---|
| Grace — Preparation | Talents actually received and why they matter for your task. |
| Grace — Decision | The choice changed by a Grace rule and why, only if it really changed. |
| Grace — Review | Relevant talents and planned checks. This does not mean the review passed. |
| Grace — Blocker | Incomplete preparation, a failure or review limitation, its consequence and the next known action. |
Calls within a stage are grouped into one announcement. Messages remain understandable in plain text and in your language. Your client may also display titles such as Grace — Get the review plan. If it cannot, the agent explains this display limitation once and keeps the useful information in its messages.
For an existing connection, use Update Grace instructions with a preview and approval, then reread changed procedures in your agent. The display depends on your agent following these instructions; available tools alone do not prove that milestones were followed.
At the end of a Grace-assisted task, the sole result milestone is a "Grace contributions" block containing only a table. Each talent that actually contributed has a row, without an arbitrary limit, showing its status, concrete contribution, and observed evidence. The statuses distinguish what was improved or verified from what remains unapplied, unverified, or unresolved. Talents without a contribution are omitted. If no talent contributed, one explicit row states that no contribution was observed.
When grace_prepare_task supplied a URL, the block ends with
"View the resolution in Grace". A project member can open it to see the same
details as in the drawer: workflow, findings, fixes, and evidence. The agent omits the link if
Grace did not provide it, rather than trying to reconstruct it.
This table ends the response: no second “Grace — Result” summary repeats it.
Recognize the names shown in your AI agent
| Displayed name | What it does |
|---|---|
grace_check_installation | Check locally whether the official kit needs an update, without sending your files. |
grace_prepare_task | Prepare instructions for the task. |
grace_deepen | Read an instruction in more detail. |
grace_review | Obtain the complete review dossier, in one or a few parts prepared by Grace. Served Grace talents remain reserved for the task and its review. |
grace_record_validation | Record the review verdicts and evidence. |
grace_present_architecture_comparison | Present three comparable architecture options and a reasoned recommendation when your profile preference or the project setting enables it. |
grace_create_organization_talent | After your explicit request and approval, save an organization talent draft without activating it. |
grace_install_talents | Add talents you approved. |
grace_doctor | Find missing prerequisites and conflicts. |
grace_discover_codebase | Help choose the project's first talents. |
grace-feedback | Start guided preparation of voluntary feedback. |
grace_submit_feedback | Submit only the contents you just approved. |
Choose the review level
Use the grace-effort skill or your profile setting to choose the level for future reviews.
Invoke it as $grace-effort in Codex, /grace-effort in a slash-capable client, or through the native
skill selector. In T3, use the active provider’s syntax. Without an argument, it reads your
preference and team minimums. The existing MCP prompt also remains available. See
Choose the review level for the four modes,
synchronization and validation limits. Your AI tool chooses the model and its reasoning;
Grace supplies the expected level without requiring a particular model name.
Compare architectures
Configure this mode in Grace independently of talents. It is enabled by default under Profile > Architecture comparisons, where each user can turn it off. A project can require it for the whole team under Project > Settings > Architecture comparison. The project setting takes precedence; otherwise each member follows their profile preference. When both are off, the tool is not exposed to the AI agent.
Ask your AI agent to compare an architecture decision. After consulting the active talents and project context, it presents exactly three options, each with a diagram, main components, advantages, risks, complexity, operational considerations, and migration. A table applies the same criteria to all three options before the recommendation.
The recommendation explains how the option fits your constraints and the main trade-off. It does not approve the decision on your behalf.
In a client compatible with MCP Apps interfaces, the comparison also appears in a visual view within the conversation. In other clients, the same content is available as Markdown with Mermaid diagrams: the enriched interface is never required for the full comparison.
Send feedback from your agent
The prompt and submission tool are offered only when your agent is connected to your Grace account through OAuth. A project API key cannot discover them or send feedback.
Launch the grace-feedback prompt from your agent's MCP interface. Describe the issue
or idea; for an issue, specify the expected result, observed result, and impact.
The agent locally removes secrets and personal information, then shows the exact contents before sending. Check every field: automatic removal may miss sensitive information. An optional PNG or JPEG screenshot is limited to 2 MiB and must be cropped or redacted before approval. Also describe in text any necessary information visible in the image: the screenshot is a visual supplement.
Nothing is sent without your explicit approval. The report goes to Grace platform
administrators along with a link to your account, organization, project, and client
used. The report and screenshot are deleted after 30 days. If submission fails, the agent keeps
the draft in the conversation and can claim success only with an FB-… reference.
Pentest mode tools
Pentest mode uses a separate MCP connection for preparation, exposing two tools:
| Displayed name | What it does |
|---|---|
grace_pentest_guidance | Prepare versioned instructions. The call accepts only a phase, AI agent, and language from predefined choices. |
grace_pentest_local_toolbox | Prepare public local-toolbox files for the selected Docker architecture. This call does not run tests. |
Prepare the relevant talents before this session. Once preparation resources are saved locally,
disable all Grace servers, including grace-pentest, and remove their credentials from
campaign processes. No Grace calls take place during the test; no campaign data is sent to Grace,
even when redacted. Testing, native subagents, evidence, and the report remain your AI agent's
and client environment's responsibility.
See Connect a project
to prepare this isolation.
What you must approve
The AI agent proposes talents to install and waits for your approval before adding them. Project discovery alone does not install talents.
To create an organization talent, it first shows the exact name, instruction, application mode, and triggers. Any change to this set requires your approval again. The tool saves only a draft: publication is a separate action in the Grace dashboard.
If Grace does not respond or denies access, see Troubleshooting. To learn about the proposed expertise, read What is a talent?.
Update Grace instructions explains the check at the start of each new conversation and the Update / Later choice. An explicit request may recheck and authorizes application after backups without a second conversational confirmation.
Local review and paged reads
See the local review workflow for prerequisites, complete reads, retries and evidence limits.