What is Roam?

Roam is a structural intelligence engine for software. It pre-indexes your codebase into a semantic graph -- symbols, dependencies, call graphs, architecture layers, git history, and runtime traces -- stored in a local SQLite DB. Agents query it via CLI or MCP instead of repeatedly grepping files and guessing structure.

For teams running AI coding agents: Roam ships two paid layers on top of the free CLI — Roam Review (PR bot scoring AI-generated changes for structural risk) and Roam Cloud (metrics-history dashboard, no source-code bodies uploaded). Both run on the OSS engine and are licensed separately.

Unlike LSPs (editor-bound, language-specific) or Sourcegraph (hosted search), Roam provides architecture-level graph queries -- offline, cross-language, and compact. It goes beyond comprehension: Roam governs architecture through budget gates, simulates refactoring outcomes, orchestrates multi-agent swarms with zero-conflict guarantees, maps vulnerability reachability paths, and enables graph-level code editing without syntax errors.

Codebase ──> [Index] ──> Semantic Graph ──> 194 Commands ──> AI Agent
              │              │                  │
           tree-sitter    symbols            comprehend
           28 languages   + edges            govern
           git history    + metrics          refactor
           runtime traces + architecture     orchestrate

Start here — the 5 verbs that cover ~80% of agent workflows

pip install roam-code

cd your-repo/
roam understand                       # 1. landing pad — what is this codebase?
roam retrieve "where is auth?"        # 2. graph-aware retrieval for free-form tasks
roam context AuthService              # 3. exact files+lines to read before changing
roam preflight AuthService            # 4. blast radius + tests + complexity check
git diff | roam critique              # 5. patch verifier — clones-not-edited, hot-path

That's the full mental model. The other CLI surface — taint, fleet, cga, simulate, mutate, partition, attest, eval-retrieve, oracle, py-types, py-modern, dark-matter, clones, propagation, fingerprint, etc. — is advanced surface for specialised workflows; you'll never need most of them.

The problem

Coding agents explore codebases inefficiently: dozens of grep/read cycles, high token cost, no structural understanding. Roam replaces this with one graph query:

$ roam context Flask
Callers: 47  Callees: 3
Affected tests: 31

Files to read:
  src/flask/app.py:76-963              # definition
  src/flask/__init__.py:1-15           # re-export
  src/flask/testing.py:22-45           # caller: FlaskClient.__init__
  tests/test_basic.py:12-30            # caller: test_app_factory
  ...12 more files

Terminal demo

Core commands

$ roam understand              # full codebase briefing
$ roam context <name>          # files-to-read with exact line ranges
$ roam retrieve "<task>"       # graph-aware spans for free-form natural-language tasks
$ roam preflight <name>        # blast radius + tests + complexity + architecture rules
$ roam critique                # verify a patch (`git diff | roam critique`)
$ roam health                  # composite score (0-100)
$ roam diff                    # blast radius of uncommitted changes

What's New in v12

v12.1+ -- Boolean oracles, IDOR classifier, index portability + Django bridge

• roam oracle <name>: 5 boolean oracles for agents — 1-token yes/no answers (symbol-exists, route-exists, is-test-only, is-reachable-from-entry, is-clone-of). Direct counter to CKB v9.2's symbolExists pattern. MCP tools: roam_oracle_*.
• roam_taint_classify (MCP only): LLM-augmented taint classification — runs roam taint then asks the agent's own model (via MCP sampling) to label each reachable finding as IDOR/AUTHZ/SQLI/XSS/etc. with confidence + reasoning. Counter to Semgrep Multimodal — same LLM-reasoning narrative without a hosted API key. Sequential for v12.1; concurrency-bounded gather lands in v12.2.
• roam index-export / roam index-import: portable, integrity-checked tarball format with manifest sha256 round-trip + optional cosign signing. Counter to Cursor's "92% similar codebase = reuse teammate's index" without a vendor cloud. Tamper-evident (manifest verifies index.db sha256 on import).
• roam eval-retrieve --emit-format coderag|beir: bench-portable JSONL emit for public leaderboard submission. CodeRAG-Bench-compatible ctxs array + BEIR-style trec_eval run files.
• Django bridge: full implicit-relationship resolution (admin→model, serializer→model, FK transitive, signal handlers, URL configs, Celery tasks, DRF routers). Ported from @LukasBerka/roam-code — credit Lukas Berka. New schema columns: framework_type, field_type, field_metadata. Post-resolver runs after graph metrics.
• worktree_git_env() (git_utils.py): GIT_INDEX_FILE override fixes .git/index.lock contention when parallel agents run roam in sibling worktrees. Wired into discovery.py, git_stats.py, changed_files.py. Ported from @river-mounts/roam-code-sf — credit Sam Hannan.

v12.0 (released 2026-05-01) -- Retrieval substrate + patch verifier

• roam retrieve "<task>": graph-aware context server. Hybrid first stage (FTS5) + structural reranker (personalised PageRank + clone-canonical signal + lexical baseline) + token-budget cap. Returns ranked spans with justification tags (pagerank=..., clone_cluster=..., fts=...) so callers can see why each span ranked. MCP tool: roam_retrieve(task, budget, k, rerank, seed_files).
• roam critique: graph-grounded patch verifier. Pipe git diff | roam critique to get findings ranked by severity. The killer signal is clones-not-edited: for every changed symbol with persisted clone siblings outside the diff, we flag the sibling as a likely missed change. Plus a blast-radius caller-count finding. Exits 5 on high severity (CI-gateable). MCP tool: roam_critique(diff_text).
• roam clones --persist: populate the clone_pairs and clone_clusters tables so downstream consumers (critique, retrieve) can query clones in O(1) instead of re-running detection.
• personalized_pagerank() in graph/pagerank.py: NetworkX personalization= wrapper with empty-seed fallback to global PR; biases ranking toward query-relevant nodes for the retrieve reranker.
• .roam/config.toml (new): zero-dep TOML loader (stdlib tomllib → tomli → in-tree subset parser). Tunable retrieve weights (alpha/beta/gamma/delta/epsilon), tokens_per_line, lexical_baseline, first_stage_token_cap, default_budget, default_k, default_rerank.
• DX corrections from dogfood pass: roam --detail <cmd> is the canonical group-level flag; misleading "use --detail" hints in 7 commands rewritten to point users at roam --detail <cmd>. --top N aliased on complexity/algo/rules (--top 0 means unlimited on rules). roam fingerprint no longer refuses graphs ≥5,000 symbols (new soft-warn threshold 20k, hard cap 100k).
• 202 CLI commands, 136 MCP tools (fleet, ask, workflow, cga, eval-retrieve remain CLI-only; v12 exposes roam_retrieve, roam_critique, roam_fleet_plan, plus 5 v12.1 boolean oracles (roam_oracle_*), roam_taint_classify, roam_pytest_fixtures, and roam_hover as MCP tools). 35-tool core preset is the default for token-budget-conscious clients.

What's New in v11

v11.2 -- AST Clone Detection + Debug Artifact Rules

• roam clones: New AST structural clone detection via subtree hashing. Finds Type-2 clones (identical control flow, different identifiers/literals) with Jaccard similarity scoring, Union-Find clustering, and automated refactoring suggestions. More precise than the metric-based duplicates command.
• 9 debug artifact rules (COR-560 through COR-568): Detect leftover print(), breakpoint(), pdb.set_trace(), console.log(), debugger, and System.out.println() in Python, JavaScript, TypeScript, and Java code. All use ast_match type with test file exemptions.
• 140 commands, 102 MCP tools (at v11.2.0 release).

v11.1.2 -- SQL + Scala Tier 1, 27 Languages

• SQL DDL promoted to Tier 1 with dedicated SqlExtractor -- tables, columns, views, functions, triggers, schemas, types (enums), sequences, ALTER TABLE ADD COLUMN. Foreign keys produce graph edges; views and triggers reference source tables. Database-schema projects now work with roam health, roam layers, roam impact, roam coupling and all graph commands.
• Scala promoted to Tier 1 with dedicated ScalaExtractor -- classes, traits, objects, case classes, sealed hierarchies, val/var properties, type aliases, imports, and inheritance. Full extends + with trait mixin resolution.
• 28 languages with 17 dedicated Tier 1 extractors.
• server.json for official MCP Registry submission.

v11.1.1 -- Command Quality Audit

• Full command audit: all 152 commands reviewed for usefulness, duplicates, and test coverage. ~20 bugs fixed, 21 new test files (700+ tests), every command docstring updated with cross-references to related commands.
• Kotlin promoted to Tier 1 via new YAML-based declarative extractor architecture. Classes, interfaces, enums, objects, functions, methods, properties, and inheritance fully extracted.
• 7 new commands: roam congestion, roam adrs, roam flag-dead, roam test-scaffold, roam sbom, roam triage, roam ci-setup.
• CI templates: roam ci-setup generates pipelines for GitHub Actions, GitLab CI, Azure Pipelines, Jenkins, and Bitbucket.
• Bug fixes: --undocumented mode in intent (wrong DB table), --changed flag in verify (was permanently dead), lazy-load violation in visualize (~500ms penalty), exit code inconsistency in rules, VERDICT-first convention enforced across all commands.
• Code quality: 15 unused variables removed, dead code swept (4 orphaned cmd files, 2 dead helper functions), algo detector false-positive rate reduced (regex-in-loop: 7 to 1, list-prepend deque suppression), 6 regex patterns pre-compiled for loop performance.

v11.0 -- MCP v2 for Agent-First Workflows

• In-process MCP execution removes per-call subprocess overhead.
• 4 compound operations (roam_explore, roam_prepare_change, roam_review_change, roam_diagnose_issue) reduce multi-step agent workflows to single calls.
• Preset-based tool surfacing (core, review, refactor, debug, architecture, full) keeps default tool choice tight for agents while retaining full depth on demand.
• MCP tools now expose structured schemas and richer annotations for safer planner behavior.
• MCP token overhead for default core context dropped from ~36K to <3K tokens (about 92% reduction).

Performance and Retrieval

• Symbol search moved to SQLite FTS5/BM25: typical search moved from seconds to tens of milliseconds on the indexed cohort (mileage varies by repo size and query selectivity — see bench/retrieve/ for the methodology).
• Incremental indexing shifted from O(N) full-edge rebuild behavior to O(changed) updates.
• DB/runtime optimizations (mmap_size, safer large-graph guards, batched writes) reduce first-run and reindex friction on larger repos.

CI, Governance, and Delivery

• GitHub Action supports quality gates, SARIF upload, sticky PR comments, and cache-aware execution.
• CI hardening includes changed-only analysis mode, trend-aware gates, and SARIF pre-upload guardrails (size/result caps + truncation signaling).
• Agent governance expanded with verification and AI-quality tooling (roam verify, roam vibe-check, roam ai-readiness, roam ai-ratio) for teams managing agent-written code.

Best for

• Agent-assisted coding -- structured answers that reduce token usage vs raw file exploration
• Large codebases (100+ files) -- graph queries beat linear search at scale
• Architecture governance -- health scores, CI quality gates, budget enforcement, fitness functions
• Safe refactoring -- blast radius, affected tests, pre-change safety checks, graph-level editing
• Multi-agent orchestration -- partition codebases for parallel agent work with zero-conflict guarantees
• Security analysis -- vulnerability reachability mapping, auth gaps, CVE path tracing
• Algorithm optimization -- detect O(n^2) loops, N+1 queries, and 21 other anti-patterns with suggested fixes
• Backend quality -- auth gaps, missing indexes, over-fetching models, non-idempotent migrations, orphan routes, API drift
• Runtime analysis -- overlay production trace data onto the static graph for hotspot detection
• Multi-repo projects -- cross-repo API edge detection between frontend and backend

When NOT to use Roam

• Real-time type checking -- use an LSP (pyright, gopls, tsserver). Roam is static and offline.
• Small scripts (<10 files) -- just read the files directly.
• Pure text search -- ripgrep is faster for raw string matching.

Why use Roam

Speed. One command replaces 5-10 tool calls (in typical workflows). Under 0.5s for any query.

Dependency-aware. Computes structure, not string matches. Knows Flask has 47 dependents and 31 affected tests. grep knows it appears 847 times.

LLM-optimized output. Plain ASCII, compact abbreviations (fn, cls, meth), --json envelopes. Designed for agent consumption, not human decoration.

Fully local. No API keys, telemetry, or network calls. Works in air-gapped environments.

Algorithm-aware. Built-in catalog of 23 anti-patterns. Detects suboptimal algorithms (quadratic loops, N+1 queries, unbounded recursion) and suggests fixes with Big-O improvements and confidence scores. Receiver-aware loop-invariant analysis minimizes false positives.

CI-ready. --json output, --gate quality gates, GitHub Action, SARIF 2.1.0.

Measured on a typical agent workflow in a 200-file Python project (Flask). See benchmarks for more.

Install

pip install roam-code

# Recommended: isolated environment
pipx install roam-code
# or
uv tool install roam-code

# From source
pip install git+https://github.com/Cranot/roam-code.git

Requires Python 3.9+. Works on Linux, macOS, and Windows.

Windows: If roam is not found after installing with uv, run uv tool update-shell and restart your terminal.

Docker (alpine-based)

docker build -t roam-code .
docker run --rm -v "$PWD:/workspace" roam-code index
docker run --rm -v "$PWD:/workspace" roam-code health

Quick Start

cd your-project
roam init                  # indexes codebase, creates config + CI workflow
roam understand            # full codebase briefing

First index takes ~5s for 200 files, ~15s for 1,000 files. Subsequent runs are incremental and near-instant.

Next steps:

• Set up your AI agent: roam describe --write (auto-detects CLAUDE.md, AGENTS.md, .cursor/rules, etc. — see integration instructions)
• Explore: roam health → roam weather → roam map
• Run the v2 stack on every PR: git diff | roam pr-analyze --explain (gates AI-generated risk; pair with roam pr-comment-render for sticky GitHub comments — see Roam Review)
• First-touch demo: roam dogfood (audit + pr-analyze + audit-trail + governance checks in one envelope)
• Add to CI: roam init already generated a GitHub Action
• Customer-facing artifacts: see starter rule packs at templates/rules/, the audit-report template + redacted sample at templates/audit-report/, and the security/procurement packet at templates/legal/security-procurement-packet.md.

git clone https://github.com/Cranot/roam-code.git
cd roam-code
pip install -e .
roam init
roam understand
roam health

Works With

Claude Code • Cursor • Windsurf • GitHub Copilot • Aider • Cline • Gemini CLI • OpenAI Codex CLI • MCP • GitHub Actions • GitLab CI • Azure DevOps

Commands

Lead with the 5 verbs. The 5 core commands cover ~80% of agent workflows: understand, context, retrieve, preflight, critique. The remaining ~197 commands are detail surface for specialised workflows (taint, fleet, cga, oracle, eval, …) — they're called by agents on demand, not memorised. This is intentional design; under the hood the canonical surface is 202 commands organised into 7 categories (plus 6 aliases for muscle memory: algo → math, weather → churn, digest / snapshot / trend → trends, onboard → understand), but you don't need to know that to start.

$ roam algo
VERDICT: 8 algorithmic improvements found (3 high, 4 medium, 1 low)
Ordering: highest impact first
Profile: balanced (filtered 0 low-signal findings)

Nested loop lookup (2):
  fn   resolve_permissions          src/auth/rbac.py:112     [high, impact=86.4]
        Current: Nested iteration -- O(n*m)
        Better:  Hash-map join -- O(n+m)
        Tip: Build a dict/set from one collection, iterate the other

  fn   find_matching_rule           src/rules/engine.py:67   [high, impact=78.1]
        Current: Nested iteration -- O(n*m)
        Better:  Hash-map join -- O(n+m)
        Tip: Build a dict/set from one collection, iterate the other

String building (1):
  meth build_query                  src/db/query.py:88       [high, impact=74.0]
        Current: Loop concatenation -- O(n^2)
        Better:  Join / StringBuilder -- O(n)
        Tip: Collect parts in a list, join once at the end

Branching recursion without memoization (1):
  fn   compute_cost                 src/pricing/calc.py:34   [medium, impact=49.5]
        Current: Naive branching recursion -- O(2^n)
        Better:  Memoized / iterative DP -- O(n)
        Tip: Add @cache / @lru_cache, or convert to iterative with a table

roam algo --task nested-lookup       # one pattern type only
roam algo --confidence high          # high-confidence findings only
roam algo --profile strict           # precision-first filtering
roam algo --task io-in-loop -n 5    # top 5 N+1 query sites
roam --json algo                     # machine-readable output
roam --sarif algo > roam-algo.sarif  # SARIF with fingerprints + fixes

$ roam minimap
<!-- roam:minimap generated=2026-02-25 -->
**Stack:** Python · JavaScript · YAML

**Workflow:**

```bash
roam minimap                    # print to stdout
roam minimap --update           # replace sentinel block in CLAUDE.md in-place
roam minimap -o docs/AGENTS.md  # target a different file
roam minimap --init-notes       # scaffold .roam/minimap-notes.md for project gotchas

Walkthrough: Investigating a Codebase

$ roam init
Created .roam/fitness.yaml (6 starter rules)
Created .github/workflows/roam.yml
Done. 226 files, 1132 symbols, 233 edges.
Health: 78/100

$ roam understand
Tech stack: Python (flask, jinja2, werkzeug)
Architecture: Monolithic — 3 layers, 5 clusters
Key abstractions: Flask, Blueprint, Request, Response
Health: 78/100 — 1 god component (Flask)
Entry points: src/flask/__init__.py, src/flask/cli.py
Conventions: snake_case functions, PascalCase classes, relative imports
Complexity: avg 4.2, 3 high (>15), 0 critical (>25)

$ roam file src/flask/app.py
src/flask/app.py  (python, 963 lines)

  cls  Flask(App)                                   :76-963
    meth  __init__(self, import_name, ...)           :152
    meth  route(self, rule, **options)               :411
    meth  register_blueprint(self, blueprint, ...)   :580
    meth  make_response(self, rv)                    :742
    ...12 more methods

$ roam deps src/flask/app.py
Imported by:
file                        symbols
--------------------------  -------
src/flask/__init__.py       3
src/flask/testing.py        2
tests/test_basic.py         1
...18 files total

$ roam weather
=== Hotspots (churn x complexity) ===
Score  Churn  Complexity  Path                    Lang
-----  -----  ----------  ----------------------  ------
18420  460    40.0        src/flask/app.py        python
12180  348    35.0        src/flask/blueprints.py python

$ roam health
Health: 78/100
  Tangle: 0.0% (0/1132 symbols in cycles)
  1 god component (Flask, degree 47, actionable)
  0 bottlenecks, 0 layer violations

=== God Components (degree > 20) ===
Sev      Name   Kind  Degree  Cat  File
-------  -----  ----  ------  ---  ------------------
WARNING  Flask  cls   47      act  src/flask/app.py

$ roam context Flask
Files to read:
  src/flask/app.py:76-963              # definition
  src/flask/__init__.py:1-15           # re-export
  src/flask/testing.py:22-45           # caller: FlaskClient.__init__
  tests/test_basic.py:12-30            # caller: test_app_factory
  ...12 more files

Callers: 47  Callees: 3

$ roam preflight Flask
=== Preflight: Flask ===
Blast radius: 47 callers, 89 transitive
Affected tests: 31 (DIRECT: 12, TRANSITIVE: 19)
Complexity: cc=40 (critical), nesting=6
Coupling: 3 hidden co-change partners
Fitness: 1 violation (max-complexity exceeded)
Verdict: HIGH RISK — consider splitting before modifying

$ roam split src/flask/app.py
=== Split analysis: src/flask/app.py ===
  87 symbols, 42 internal edges, 95 external edges
  Cross-group coupling: 18%

  Group 1 (routing) — 12 symbols, isolation: 83% [extractable]
    meth  route              L411  PR=0.0088
    meth  add_url_rule       L450  PR=0.0045
    ...

=== Extraction Suggestions ===
  Extract 'routing' group: route, add_url_rule, endpoint (+9 more)
    83% isolated, only 3 edges to other groups

$ roam why Flask url_for Blueprint
Symbol     Role          Fan         Reach     Risk      Verdict
---------  ------------  ----------  --------  --------  --------------------------------------------------
Flask      Hub           fan-in:47   reach:89  CRITICAL  God symbol (47 in, 12 out). Consider splitting.
url_for    Core utility  fan-in:31   reach:45  HIGH      Widely used utility (31 callers). Stable interface.
Blueprint  Bridge        fan-in:18   reach:34  moderate  Coupling point between clusters.

$ roam describe --write
Wrote CLAUDE.md (98 lines)  # auto-detects: CLAUDE.md, AGENTS.md, .cursor/rules, etc.

$ roam health --gate
Health: 78/100 — PASS

Integration with AI Coding Tools

Roam is designed to be called by coding agents via shell commands. Instead of repeatedly grepping and reading files, the agent runs one roam command and gets structured output.

Decision order for agents:

Fastest setup:

roam describe --write               # auto-detects your agent's config file
roam describe --write -o AGENTS.md  # or specify an explicit path
roam describe --agent-prompt        # compact ~500-token prompt (append to any config)
roam minimap --update               # inject/refresh annotated codebase minimap in CLAUDE.md

Agent not using Roam correctly? If your agent is ignoring Roam and falling back to grep/read exploration, it likely doesn't have the instructions. Run:

roam describe --write          # writes instructions to your agent's config (CLAUDE.md, AGENTS.md, etc.)

If you already have a config file and don't want to overwrite it:

roam describe --agent-prompt   # prints a compact prompt — copy-paste into your existing config
roam minimap --update          # injects an annotated codebase snapshot into CLAUDE.md (won't touch other content)

This teaches the agent which Roam command to use for each situation (e.g., roam preflight before changes, roam context for files to read, roam diagnose for debugging).

## Codebase navigation

This project uses `roam` for codebase comprehension. Always prefer roam over Glob/Grep/Read exploration.

Before modifying any code:
1. First time in the repo: `roam understand` then `roam tour`
2. Find a symbol: `roam search <pattern>`
3. Before changing a symbol: `roam preflight <name>` (blast radius + tests + fitness)
4. Need files to read: `roam context <name>` (files + line ranges, prioritized)
5. Debugging a failure: `roam diagnose <name>` (root cause ranking)
6. After making changes: `roam diff` (blast radius of uncommitted changes)

Additional: `roam health` (0-100 score), `roam impact <name>` (what breaks),
`roam pr-risk` (PR risk), `roam file <path>` (file skeleton).

Run `roam --help` for all commands. Use `roam --json <cmd>` for structured output.

MCP Server

Roam includes a Model Context Protocol server for direct integration with tools that support MCP.

pip install "roam-code[mcp]"
roam mcp

136 tools, 10 resources, and 5 prompts are available in the full preset. Most tools are read-only index queries; side-effect tools are explicitly annotated.

MCP v2 highlights (v11):

• In-process MCP execution (no subprocess shell-out per call)
• Preset-based tool surfacing (core, review, refactor, debug, architecture, full)
• Compound tools that collapse multi-step exploration/review flows into one call
• Structured output schemas + tool annotations for safer planner behavior

MCP-native enhancements (v12):

• Sampling-driven compression -- pass summarize=True to roam_explore, roam_understand, roam_health, or roam_repo_map. The server asks the client's own LLM (no API keys) to compress the full envelope into a short briefing, dropping output from ~50 KB JSON to ~1-2 KB prose. Falls back gracefully when the client doesn't support sampling.
• Server-side session memory -- roam_context, roam_explore, and roam_retrieve now remember symbols you've touched in the current session and auto-bias ranking without you threading recent_symbols through every call. Explicit args still win.
• Phase-aware progress -- roam_init, roam_reindex, and roam_orchestrate stream real discover -> parse -> extract -> resolve -> graph -> metrics progress to the client, replacing the old 5/100 placeholders.
• Symbol & path completions -- new roam_complete(prefix, kind, limit) tool returns just names from the FTS5 index (cheaper than roam_search_symbol). A protocol-level handler is also installed for clients that support completion/complete.
• Reactive resource invalidation (opt-in) -- set ROAM_MCP_WATCH=1 and the server watches the working tree, runs incremental reindex on file changes, and emits notifications/resources/updated for roam://health, roam://summary, etc., so subscribed clients see fresh data without polling.

Default preset: core (25 tools: 24 core + roam_expand_toolset meta-tool).

# Default
roam mcp

# Full toolset
ROAM_MCP_PRESET=full roam mcp

# Legacy compatibility (same as full preset)
ROAM_MCP_LITE=0 roam mcp

Core preset tools: roam_affected_tests, roam_batch_get, roam_batch_search, roam_complete, roam_complexity_report, roam_context, roam_dead_code, roam_deps, roam_diagnose, roam_diagnose_issue, roam_diff, roam_expand_toolset, roam_explore, roam_file_info, roam_health, roam_impact, roam_pr_risk, roam_preflight, roam_prepare_change, roam_review_change, roam_search_symbol, roam_syntax_check, roam_trace, roam_understand, roam_uses.

claude mcp add roam-code -- roam mcp

{
  "mcpServers": {
    "roam-code": {
      "command": "roam",
      "args": ["mcp"]
    }
  }
}

{
  "mcpServers": {
    "roam-code": {
      "command": "roam",
      "args": ["mcp"],
      "cwd": "/path/to/your/project"
    }
  }
}

{
  "mcpServers": {
    "roam-code": {
      "command": "roam",
      "args": ["mcp"]
    }
  }
}

{
  "servers": {
    "roam-code": {
      "type": "stdio",
      "command": "roam",
      "args": ["mcp"]
    }
  }
}

Roam Review (PR bot for AI-generated changes)

Most "AI safety net" tools — CodeRabbit, Greptile, Qodo — review PR semantics (does the diff look right?). They don't read your graph: who calls the changed symbol, which layer it belongs to, whether the AI just touched a god-component with 47 callers. Roam Review fills that gap.

roam pr-analyze is the CLI engine: pipe a unified diff in, get back a verdict (INTENTIONAL / SAFE / REVIEW / BLOCK) plus AI-likelihood score, blast radius, rule violations, suggested reviewers, and a governance audit-trail record.

git diff main..HEAD | roam pr-analyze --explain --with-reviewers --audit-trail
roam pr-analyze main..HEAD --gate                      # exit 5 on BLOCK (CI gate)
roam --json pr-analyze --input pr.diff | roam pr-comment-render   # ready-to-post markdown
roam pr-analyze --batch ./diffs/ --cache --parallel 4  # 24-55x speedup on incremental re-runs
roam audit-trail-verify                                # check SHA-256 chain integrity
roam audit-trail-conformance-check --gate              # governance-evidence score (CI)

Nine weighted heuristic signals score the diff for AI-likelihood: add/remove ratio, comment density, test coverage, function-size variance, generic naming, orphan imports, placeholder density (TODO/FIXME/NotImplementedError stubs), LLM-phrase density ("we use this approach because…"), suspicious imports (numbered modules, mass typing imports). Each carries language-aware weights across Python / TypeScript / JavaScript / Go / Rust / Java / Kotlin. Starter rule packs ship for Python, TypeScript, Go, and Java at templates/rules/ — drop one at .roam/rules.yml to enable. Custom rules look like:

rules:
  - id: no-frontend-db-import
    description: Frontend modules must not import from db/ directly
    pattern: import_from              # supported: import_from, function_call, class_inherit, decorator_use
    source_glob: "frontend/**/*.{ts,tsx}"
    forbidden_target_glob: "lib/db/**"
    severity: BLOCK
  - id: no-eval
    pattern: function_call
    source_glob: "src/**/*.py"
    forbidden_target_glob: "eval"
    severity: BLOCK

A drift baseline (--save-baseline / --baseline FILE) compares the current PR's signals against the previous analysis and auto-escalates the verdict on regression — the GitHub App reads this to render (+5 vs prev) / (-22 vs prev) arrows on every push.

The pr-analyze --audit-trail flag appends a SHA-256-chained record to .roam/audit-trail.jsonl for each analysis (actor, repo, git SHA, diff hash, verdict, blast radius, AI-likelihood, intent marker). roam audit-trail-verify walks the chain and surfaces tampered records; roam audit-trail-export --format md|csv|json produces procurement-friendly reports. This is best framed as SOC 2 CC8.1, ISO 42001, and internal AI-governance evidence. Article 12 is only relevant for actual Annex III high-risk AI-system buyers, and Roam's records are supporting evidence rather than complete runtime inference logs.

A hosted GitHub App is in development on top of this CLI engine. Until it ships, the CLI is usable today as a free CI gate (roam pr-analyze --gate exits 5 on BLOCK).

Roam Cloud (metrics history, no source upload)

roam metrics-push sends a summary-only payload from roam audit --json to a Roam Cloud endpoint — numerical metrics, file paths (or SHA-256 hashes when --anonymize), and identifier names only. No source-code bodies are transmitted, ever. Inspect the exact payload locally with --dry-run before any token is set.

roam metrics-push --dry-run                            # local-only inspection
roam metrics-push --token $ROAM_CLOUD_TOKEN --anonymize
roam metrics-push --no-hotspots --json                 # minimal payload

The hosted dashboard at roam.cloud (in development) renders trend charts of health-score, debt, dead-code count, danger-zone count, and bus-factor concentration over time. The schema (roam-metrics-v1) is allow-listed: any payload key outside the allow-list is rejected by the receiving API.

Both products are paid layers on top of the free CLI; the CLI itself stays Apache 2.0, zero-API-key, fully local, forever.

CI/CD Integration

All you need is Python 3.9+ and pip install roam-code.

GitHub Actions

# .github/workflows/roam.yml
name: Roam Analysis
on: [pull_request]

jobs:
  roam:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - uses: Cranot/roam-code@main
        with:
          commands: health
          gate: "score>=70"
          sarif: true
          comment: true

Use roam init to auto-generate this workflow.

roam-analysis:
  stage: test
  image: python:3.12-slim
  before_script:
    - pip install roam-code
  script:
    - roam index
    - roam health --gate
    - roam --json pr-risk origin/main..HEAD > roam-report.json
  artifacts:
    paths:
      - roam-report.json
  rules:
    - if: $CI_MERGE_REQUEST_IID

pip install roam-code
roam index
roam health --gate               # exit 5 on failure (reads .roam-gates.yml)
roam --json health > report.json

SARIF Output

Roam exports analysis results in SARIF 2.1.0 format for GitHub Code Scanning.

from roam.output.sarif import health_to_sarif, write_sarif

sarif = health_to_sarif(health_data)
write_sarif(sarif, "roam-health.sarif")

- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: roam-health.sarif

For Teams

Zero infrastructure, zero vendor lock-in, zero data leaving your network.

Language Support

Tier 1 -- Full extraction (dedicated parsers)

| Ruby | .rb | classes, modules, methods, singleton methods, constants | require, require_relative, include/extend, calls, ClassName.new | class inheritance | | Kotlin | .kt .kts | classes, interfaces, enums, objects, functions, methods, properties | imports, calls, type refs | extends, implements | | Scala | .scala .sc | classes, traits, objects, case classes, functions, methods, val/var, type aliases | imports, calls, new | extends, with (trait mixins) | | SQL (DDL) | .sql | tables, columns, views, functions, triggers, schemas, types (enums), sequences | foreign keys, view table deps, trigger table/function refs | -- | | Swift | .swift | classes, structs, enums, protocols, functions, methods, properties | imports, calls, type refs | extends, conforms | | JSONC | .jsonc | via JSON grammar | -- | -- | | MDX | .mdx | via Markdown grammar | -- | -- |

Performance

Agent-efficiency benchmarks: see the benchmarks/ directory for harness, repos, and results.

How It Works

Codebase
    |
[1] Discovery ──── git ls-files (respects .gitignore + .roamignore)
    |
[2] Parse ──────── tree-sitter AST per file (28 languages)
    |
[3] Extract ────── symbols + references (calls, imports, inheritance)
    |
[4] Resolve ────── match references to definitions → edges
    |
[5] Metrics ────── adaptive PageRank, betweenness, cognitive complexity, Halstead
    |
[6] Algorithms ── 23-pattern anti-pattern catalog (O(n^2) loops, N+1, recursion)
    |
[7] Git ────────── churn, co-change matrix, authorship, Renyi entropy
    |
[8] Clusters ───── Louvain community detection
    |
[9] Health ─────── per-file scores (7-factor) + composite score (0-100)
    |
[10] Store ─────── .roam/index.db (SQLite, WAL mode)

After the first full index, roam index only re-processes changed files (mtime + SHA-256 hash). Incremental updates are near-instant.

.roamignore

Create a .roamignore file in your project root to exclude files from indexing. It uses full gitignore syntax:

Key rules: * matches within a single path segment (not across /). ** matches across / boundaries. Last matching pattern wins (for negation). Patterns containing / are anchored to the repo root.

# .roamignore example
*_pb2.py
*_pb2_grpc.py
vendor/
node_modules/
*.generated.*
/build/
!build/keep/

You can also exclude patterns via roam config --exclude "*.proto" (stored in .roam/config.json) or inspect active patterns with roam config --show.

How Roam Compares

roam-code is the only tool that combines graph algorithms (PageRank, Tarjan SCC, Louvain clustering), git archaeology, architecture simulation, and multi-agent partitioning in a single local CLI with zero API keys.

Documentation (local HTML in docs/site/, CI-deployed via .github/workflows/pages.yml):

• docs/site/getting-started.html — tutorial
• docs/site/command-reference.html — examples
• docs/site/architecture.html — diagram + internals
• docs/site/landscape.html — competitor matrix

Key Differentiators

• vs AI IDEs (Cursor, Windsurf, Augment): roam-code provides deterministic structural analysis. AI IDEs use probabilistic embeddings that can't guarantee reproducible results.
• vs AI Agents (Claude Code, Codex CLI, Gemini CLI): These agents read files one at a time. roam-code pre-computes relationships so agents get instant answers about architecture, blast radius, and dependencies.
• vs SAST Tools (SonarQube, CodeQL, Semgrep): SAST tools find bugs and vulnerabilities. roam-code understands architecture -- how code is structured, where it's coupled, and what breaks when you change it. Complementary, not competitive.
• vs Code Search (Sourcegraph/Amp, Greptile): Text search finds where code is. roam-code understands why code matters -- which functions are central, which modules are tangled, which files are high-risk.

FAQ

Does Roam send any data externally? No. Zero network calls. No telemetry, no analytics, no update checks.

Can Roam run in air-gapped environments? Yes. Once installed, no internet access is required.

Does Roam modify my source code? Read-only by default. Creates .roam/ with an index database. The roam mutate command can apply code changes (move/rename/extract) but defaults to --dry-run mode — you must explicitly pass --apply to write changes.

How does Roam handle monorepos? Indexes from the root. Batched SQL handles 100k+ symbols. Incremental updates stay fast.

How does Roam handle multi-repo projects (e.g., frontend + backend)? Use roam ws init <repo1> <repo2> to create a workspace. Each repo keeps its own index; a workspace overlay DB stores cross-repo API edges. roam ws resolve scans for REST endpoints and matches frontend calls to backend routes. Then roam ws context, roam ws trace, etc. work across repos.

Is Roam compatible with SonarQube / CodeScene? Yes. Roam complements existing tools. Both can run in the same CI pipeline. SARIF output integrates with GitHub Code Scanning.

Limitations

Static analysis trade-offs:

• Static analysis primarily -- can't trace dynamic dispatch, reflection, or eval'd code. Runtime trace ingestion (roam ingest-trace) adds production data but requires external trace export
• Import resolution is heuristic -- complex re-exports or conditional imports may not resolve
• Limited cross-language edges -- Salesforce, Protobuf, REST API, and multi-repo edges are supported, but not arbitrary FFI
• Tier 2 languages get basic symbol extraction only via generic tree-sitter walker
• Large monorepos (100k+ files) may have slow initial indexing

Troubleshooting

Update / Uninstall

# Update
pipx upgrade roam-code
uv tool upgrade roam-code
pip install --upgrade roam-code

# Uninstall
pipx uninstall roam-code
uv tool uninstall roam-code
pip uninstall roam-code

Delete .roam/ from your project root to clean up local data.

Development

git clone https://github.com/Cranot/roam-code.git
cd roam-code
pip install -e ".[dev]"   # includes pytest, ruff
pytest tests/              # ~5500 tests, Python 3.9-3.13

# Or use Make targets:
make dev      # install with dev extras
make test     # run tests
make lint     # ruff check

roam-code/
├── pyproject.toml
├── action.yml                         # Reusable GitHub Action
├── src/roam/
│   ├── __init__.py                    # Version (from pyproject.toml)
│   ├── cli.py                         # Click CLI (194 canonical + 7 aliases)
│   ├── mcp_server.py                  # MCP server (136 tools, 10 resources, 5 prompts)
│   ├── db/
│   │   ├── connection.py              # SQLite (WAL, pragmas, batched IN)
│   │   ├── schema.py                  # Tables, indexes, migrations
│   │   └── queries.py                 # Named SQL constants
│   ├── index/
│   │   ├── indexer.py                 # Orchestrates full pipeline
│   │   ├── discovery.py               # git ls-files, .gitignore
│   │   ├── parser.py                  # Tree-sitter parsing
│   │   ├── symbols.py                 # Symbol + reference extraction
│   │   ├── relations.py               # Reference resolution -> edges
│   │   ├── complexity.py              # Cognitive complexity (SonarSource) + Halstead metrics
│   │   ├── git_stats.py               # Churn, co-change, blame, Renyi entropy
│   │   ├── incremental.py             # mtime + hash change detection
│   │   ├── file_roles.py              # Smart file role classifier
│   │   └── test_conventions.py        # Pluggable test naming adapters
│   ├── languages/
│   │   ├── base.py                    # Abstract LanguageExtractor
│   │   ├── registry.py                # Language detection + aliasing
│   │   ├── *_lang.py                  # One file per language (21 dedicated + generic)
│   │   └── generic_lang.py            # Tier 2 fallback
│   ├── bridges/
│   │   ├── base.py, registry.py       # Cross-language bridge framework
│   │   ├── bridge_salesforce.py       # Apex <-> Aura/LWC/Visualforce
│   │   └── bridge_protobuf.py         # .proto -> Go/Java/Python stubs
│   ├── catalog/
│   │   ├── tasks.py                  # Universal algorithm catalog (23 patterns)
│   │   └── detectors.py              # Anti-pattern detectors with confidence calibration
│   ├── workspace/
│   │   ├── config.py                  # .roam-workspace.json
│   │   ├── db.py                      # Workspace overlay DB
│   │   ├── api_scanner.py             # REST API endpoint detection
│   │   └── aggregator.py              # Cross-repo aggregation
│   ├── graph/
│   │   ├── builder.py, pagerank.py    # DB -> NetworkX, PageRank
│   │   ├── cycles.py, clusters.py     # Tarjan SCC, propagation cost, Louvain, modularity Q
│   │   ├── layers.py, pathfinding.py  # Topo layers, k-shortest paths
│   │   ├── simulate.py, spectral.py   # Architecture simulation, Fiedler bisection
│   │   ├── partition.py, fingerprint.py # Multi-agent partitioning, topology fingerprints
│   │   └── anomaly.py                 # Statistical anomaly detection
│   ├── commands/
│   │   ├── resolve.py                 # Shared symbol resolution
│   │   ├── graph_helpers.py           # Shared graph utilities (adj builders, BFS)
│   │   ├── context_helpers.py         # Data-gathering helpers for context command
│   │   ├── gate_presets.py            # Framework-specific gate rules
│   │   └── cmd_*.py                   # One module per command
│   ├── analysis/
│   │   ├── effects.py                 # Side-effect classification engine
│   │   └── taint.py                   # Taint analysis
│   ├── refactor/
│   │   ├── codegen.py                 # Import generation (Python/JS/Go)
│   │   └── transforms.py             # move/rename/add-call/extract transforms
│   ├── rules/
│   │   ├── engine.py                  # YAML rule parser + graph query evaluator
│   │   ├── builtin.py                 # 10 built-in governance rules
│   │   ├── ast_match.py               # AST pattern matching with $METAVAR captures
│   │   └── dataflow.py                # Intra-procedural dataflow analysis
│   ├── runtime/
│   │   ├── trace_ingest.py            # OpenTelemetry/Jaeger/Zipkin ingestion
│   │   └── hotspots.py                # Runtime hotspot analysis
│   ├── search/
│   │   ├── tfidf.py                   # TF-IDF semantic search engine
│   │   ├── index_embeddings.py        # Embedding index builder
│   │   └── onnx_embeddings.py         # Optional local ONNX semantic backend
│   ├── security/
│   │   ├── vuln_store.py              # CVE/vulnerability storage
│   │   └── vuln_reach.py              # Vulnerability reachability paths
│   └── output/
│       ├── formatter.py               # Token-efficient formatting
│       ├── sarif.py                   # SARIF 2.1.0 output
│       └── schema_registry.py         # JSON envelope schema versioning
└── tests/                             # ~5500 tests across 186 test files

Dependencies

Optional: fastmcp >= 2.0 (MCP server — install with pip install "roam-code[mcp]")

Optional: Local semantic ONNX stack (numpy, onnxruntime, tokenizers) via pip install "roam-code[semantic]"; verify activation with roam config --semantic-status.

Roadmap

Shipped

• MCP v2 agent surface: in-process execution, compound operations, presets, schemas, annotations, and compatibility profiles.
• Full command and MCP inventory parity in docs: 197 canonical CLI commands and 136 MCP tools.
• CI hardening: composite action, changed-only mode, trend-aware gates, sticky PR updater, and SARIF guardrails.
• Performance foundation: FTS5/BM25 search, O(changed) incremental indexing, DB/index optimizations.
• Agent governance suite: vibe-check, ai-readiness, verify, ai-ratio, duplicates, advanced algo scoring/SARIF.
• Ownership/review intelligence: codeowners, drift, simulate-departure, suggest-reviewers, api-changes, test-gaps, semantic-diff, secrets.
• Multi-agent operations: partition, affected, syntax-check, workspace-aware context and traces.
• Budget-aware context delivery: --budget (partial rollout), PageRank-weighted truncation, conversation-aware ranking.

Next

• Terminal demo GIF in README.
• GitHub repo topics.
• GitHub Discussions enabled.
• MCP directory + awesome-list submissions.

Contributing

git clone https://github.com/Cranot/roam-code.git
cd roam-code
pip install -e .
pytest tests/   # all ~5500 tests must pass

Good first contributions: add a Tier 1 language (see go_lang.py or php_lang.py as templates), improve reference resolution, add benchmark repos, extend SARIF converters, add MCP tools.

Please open an issue first to discuss larger changes.

License

Apache 2.0