scry

Scry Marker Specification - This project implements a formal, generic scry marker contract defined here

See scryspec.com for an overview.

Marker-indexed SQL cache MCP server. Scry indexes in-file @scry.* markers into a SQLite database that agents query via read-only SQL, providing structured project knowledge without LLM reasoning.

Install

uv pip install scry-mcp
# or:
pip install scry-mcp

The PyPI distribution is scry-mcp (the bare name scry was already taken on PyPI). The import name and the installed console command are both scry.

Quickstart

# In your project root:
scry init                # scaffolds agent/ + driver dirs, updates .gitignore

Then add it to your MCP client config:

{
  "mcpServers": {
    "scry": {
      "command": "scry"
    }
  }
}

The MCP client inherits cwd from wherever it's launched, and scry walks up from there looking for an agent/ directory — so the same config works for any project. Bare scry (no subcommand) starts the MCP server over stdio — that's what Claude calls. Other subcommands:

Command	Purpose
scry	Run the MCP server (default).
scry init [path]	Create agent/ + agent/drivers/@local/scry/{data,runtime,scripts}/ and a local .gitignore inside the driver dir. Idempotent — safe to run inside an ACP project.
scry surface [--force]	One-shot batch reindex without booting the server.
scry version	Print the package version.

The server walks up from cwd until it finds an agent/ directory; that becomes the project root. The cache lives at agent/drivers/@<namespace>/scry/data/project.db and is gitignored.

Markers

Scry recognizes three marker kinds per scry-spec v1.2.0: @scry.entry and @scry.anchor are block markers with a YAML body between open/close tokens; @scry.bind is a line or block marker declaring cross-references.

<!-- @scry.entry
id: design.auth-flow~a1b2c3d4
kind: design
summary: >
  JWT auth middleware, token validation, refresh flow
status: active
weight: 0.85
tags: ["scope:auth", "topic:security"]
rationale: >
  Missing this causes auth bypass bugs
applies: modifying auth, adding protected endpoints
seeded_questions:
  - How does token refresh work?
extras:
  owner: auth-team
  jira: AUTH-1247
  reviewed_at: 2026-05-19
@scry.entry.end -->

<!-- @scry.anchor auth-check~f1e2d3c4
description: JWT validation point for protected routes
@scry.anchor.end -->

# @scry.bind validate-jwt~a1b2c3d4 spec.auth~xyz89012#FR3
# @scry.bind jwt-expiry~b2c3d4e5 spec.auth~xyz89012#UT1

Block markers can be embedded in any host-language comment style (HTML, Python, JS, JSDoc, Rust, bare YAML). Comment prefixes are inferred from the YAML body — there is no per-language config.

extras field (scry-spec v1.2.0, FR4.B)

@scry.entry accepts an optional extras field: a single-depth map of arbitrary string keys to scalar values (string | number | boolean | null). Use it to attach structured metadata that isn't part of the core schema — ownership, ticket IDs, cost ledgers, review timestamps, anything you want to query later.

Constraints:

• Single depth only — values must be scalars; nested maps and lists are diagnostic violations.
• Serialized size ≤ 4 KB (the spec's SHOULD-level cap).
• Oversize payloads MUST round-trip intact — the cap is a diagnostic, not a truncation gate. Empty extras: {} emits a SHOULD-level warning.

Indexed and queryable as of v0.17.0. The extras field is serialized to compact JSON-text in the scry__doc.extras column (NULL when absent), and exposed to scry_sql via SQLite JSON1. A doc with

extras:
  cost_usd: 12.5
  tier: gold
  active: true

answers queries like:

SELECT id,
       json_extract(extras, '$.cost_usd') AS cost,
       json_extract(extras, '$.tier')     AS tier
FROM scry__doc
WHERE kind = 'deliverable'
  AND json_extract(extras, '$.active') = 1
ORDER BY cost DESC;

Round-trip is byte-equivalent for the YAML scalar map: integers, floats, strings, booleans, and null are preserved.

@scry.entry kind values (v1.2.0 baseline)

kind	use for
design	architecture and design docs
pattern	canonical recipes, established patterns
spec	requirements and specifications
lesson	post-mortems, "I tried X and it failed because Y"
internal	service quirks, undocumented behaviors
task	discrete work items
milestone	phase markers, exit criteria
goal	objective satisfied by a deliverable set via the satisfies typed edge (v1.2.0)
report	wake/session reports
audit	security or integrity audits
research	research notes
code	implementation-specific docs

MCP tools

Tool	Purpose
scry_sql(query)	Read-only SQL gateway. Rejects mutator keywords. Returns {results, row_count} JSON.
scry_grep(query, kind?, status?, path_glob?, limit?)	Full-text search over indexed file bodies (scry__file_fts). Complements scry_sql when you want to search prose, not just marker fields.
scry_mint(kind, prefix)	Generate a collision-free ID and the marker schema.
scry_mint_with_check(kind, prefix)	Preferred minter — same as scry_mint plus tier-1/tier-2 collision warnings.
scry_surface(force=false, path?)	Batch reindex from disk. force=true hard-deletes records whose source file no longer exists; path scopes the walk to a single file or subdirectory.
scry_sink(then_surface=false)	Lower the index back to disk-only state — truncates all index tables atomically (schema preserved, disk markers untouched). Requires user confirmation; then_surface=true rebuilds in one call.
scry_scrub()	Create a <branch>--clean git branch with all @scry.* markers stripped from non-agent files.
scry_script(action, script?, params?)	Discover and run validation scripts from src/scry/scripts/ and agent/drivers/@<ns>/scry/scripts/.
scry_db_health()	Probe the project DB and report status ∈ {ok, corrupt, locked}, integrity_check result, and scry__doc row count. For substrate auto-restore loops that need to distinguish corruption from transient WAL write-lock contention.

Database schema

Marker-backed core tables:

Table	Holds
scry__doc	One row per @scry.entry marker — id, kind, status, weight, summary, rationale, applies, current_path, ephemeral, missing_since, content_hash, extras, timestamps.
scry__doc_tag	Tags as a join table — (doc_id, tag).
scry__doc_seeded_question	Seeded questions per doc — (doc_id, ordinal, question).
scry__anchor	One row per @scry.anchor marker — id, doc_id, description, content_hash, timestamps.
scry__anchor_seeded_question	Seeded questions per anchor — (anchor_id, ordinal, question).
scry__bind	One row per @scry.bind marker — source_doc_id, source_local_id, target_id, target_fragment, comment, content_hash, timestamps.
scry__rel	Typed edges between docs — (from_id, to_id, predicate, fragment). Predicates: depends_on, implements, supersedes. Cycle detection enforced on depends_on.
scry__file	Universal file-body index — (path, doc_id, body, content_hash, last_modified). Populated during scry_surface.
scry__warning	Lint-style warnings emitted by the parser/indexer — id, kind, marker_kind, marker_id, file_path, message, detected_at.
migration	Schema migration ledger.

FTS5 virtual tables (trigger-maintained, queryable via FTS MATCH):

FTS table	Searches
scry__doc_fts	Doc summary, rationale, applies, current_path.
scry__doc_tag_fts	Tag strings.
scry__doc_seeded_question_fts	Doc seeded questions.
scry__anchor_fts	Anchor descriptions.
scry__bind_fts	Bind source_local_id, target_id, comment.
scry__file_fts	Full file bodies. Prefer the scry_grep tool over hand-rolled queries.

The cache is fully reconstructable from disk via scry_surface. The DB is gitignored; after git pull, agents call scry_surface to rebuild.

Watcher

A daemon thread runs alongside the MCP server, watching the project tree with a 150 ms debounce window. A lock file at agent/drivers/@<ns>/scry/runtime/lock performs PID-based primary election so multiple sessions don't race writes. The primary instance runs a cold scan on startup; secondaries observe and wait.

On file deletion: docs are soft-deleted (missing_since set); anchors and binds are hard-deleted.

Tests

uv pip install -e ".[dev]"
pytest

159 tests cover the parser, SQL gateway, mint, surface, watcher plumbing, script discovery, relationship cycle detection, FR4.B extras indexing + JSON1 queryability, schema-migration backfill, and concurrent-connection retry behavior.

License

MIT — see LICENSE.