annotate_outputs
[SUPPORT] 9e02e448 — capture a human annotation for a path inside an outputs tree WITHOUT touching the filesystem. Upserts a row into the annotations layer of the local DuckDB outputs index for outputs_dir. Two tiers, same mechanism: Tier 1 = pass outputs_dir as path to annotate the whole tree ('...
This record as markdown: /tools/io-github-ajc3xc-meridian/annotate-outputs.md
What annotate_outputs does on Meridian
AI agents use annotate_outputs to create or update resources in Meridian, usually the action step of a workflow, after the agent has gathered context. Every call changes real data in your Meridian environment.
| Parameter | Type | Required | Description |
|---|---|---|---|
note | string | Yes | The annotation text (e.g. 'PCA on, BFS off — results from run on 2026-07-12 with lr=0.001'). |
path | string | Yes | The path to annotate — either the outputs_dir root (Tier 1, tree-level annotation) or any file/subdirectory path within the tree (Tier 2, per-run or per-file an |
run_params | object | — | Optional free-form key-value dict of run parameters to log alongside the note (e.g. {"lr": 0.001, "epochs": 100}). |
outputs_dir | string | Yes | Absolute path to the outputs directory tree root (same value you pass to search_outputs). |
Parameters from the server's own tool schema.
Why annotate_outputs is rated Medium
An AI agent can call annotate_outputs faster than any human can review: one bad instruction and it creates or modifies resources in Meridian by the hundred, each call as confident as the last.
Risk signalsAccepts file system path (path)
Attacks that exploit this kind of access
The rule that runs annotate_outputs safely
PolicyLayer is an MCP gateway: it sits between your AI agents and Meridian, and checks every tool call against a rule you set before the call runs. Nothing changes on the server itself. For annotate_outputs, this is the rule to start with:
annotate_outputs stays usable, but capped: an agent stuck in a loop can't make hundreds of changes a minute. Everything else on the server is denied unless you say otherwise.
The button opens the PolicyLayer dashboard: create your workspace, connect Meridian, apply this rule, and every annotate_outputs call is checked against it from then on.
Questions about annotate_outputs
[SUPPORT] 9e02e448 — capture a human annotation for a path inside an outputs tree WITHOUT touching the filesystem. Upserts a row into the annotations layer of the local DuckDB outputs index for outputs_dir. Two tiers, same mechanism: Tier 1 = pass outputs_dir as path to annotate the whole tree ('what this experiment tree is about'); Tier 2 = pass any sub-path (file or directory) to annotate a specific run, file, or subdirectory ('PCA on, BFS off, overwritten 5x'). run_params is an optional free-form dict of parameters logged alongside the note (e.g. {"lr": 0.001, "batch_size": 32}). Annotations are automatically surfaced in search_outputs results — any hit's path (or its nearest ancestor directory) that has an annotation will have it included in the hit's 'annotations' field without a second tool call. A MERIDIAN_NOTES.md file placed anywhere in the tree is also auto-ingested into the same table on every rebuild, keyed to its containing directory. Returns the stored annotation as a dict. Persistent-state disclosure: on hosted Meridian, supplied text and project/session metadata -- including task log entries, pinned decisions, sprint items, notes, handoff/goal state, and HITL queue items -- are sent to and stored in Meridian's service, in an isolated per-tenant Postgres database (Neon); self-hosted deployments keep the same categories in the configured local SQLite/Postgres database. This data is visible in the dashboard and API, and may resurface in later project context or handoffs. Notes and pinned decisions can be deleted individually; task log entries and sprint items can be deleted via the dashboard/API (not exposed as an agent-facing tool); HITL queue items and handoff state have no per-record delete. Full removal of any of this data is available via project or account deletion, using the documented controls. Do not include secrets. It is categorised as a Write tool in the Meridian MCP Server, which means it can create or modify data. Consider rate limits to prevent runaway writes.
annotate_outputs accepts 4 parameters: note, path, run_params, outputs_dir. Required: note, path, outputs_dir. The full parameter table on this page comes from the server's own tool schema.
Register the Meridian MCP server in PolicyLayer and add a rule for annotate_outputs: allow, deny, rate-limit, or require approval. Point your MCP client at the PolicyLayer proxy URL and the rule is enforced on every call, before it reaches Meridian. Nothing to install.
annotate_outputs is a Write tool with medium risk. Write tools should be rate-limited to prevent accidental bulk modifications.
Yes. Add a rate_limit block to the annotate_outputs rule in your PolicyLayer policy. For example, setting max: 10 and window: 60 limits the tool to 10 calls per minute. Rate limits are tracked per agent session and reset automatically.
Set action: deny in the PolicyLayer policy for annotate_outputs. The AI agent will receive a policy violation error and cannot call the tool. You can also include a reason field to explain why the tool is blocked.
annotate_outputs is provided by the Meridian MCP server (@meridianmcp/mcp). PolicyLayer sits as a proxy in front of this server to enforce policies before tool calls reach the server.
More on Meridian, and thousands of servers like it.
This server
Across the catalogue