ForgeSAFE
Turn Claude Code into a plan-execute-validate loop with parallel work, intelligent retry, and memory
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
[](https://github.com/TT-Wang/forge/actions/workflows/ci.yml) [](./LICENSE) [](https://nodejs.org/) [](https://glama.ai/mcp/servers/TT-Wang/forge)
Turn Claude Code into a structured delivery loop: plan the work, run modules in parallel, validate deeply, retry intelligently, and carry forward what worked.
Why Forge
Single-agent Claude Code drifts past ~5 steps. The failure mode isn't code quality — it's silent state corruption: parallel workers branching from stale HEAD, modules quietly clobbering each other's changes, "DONE" status that hides broken integration. Forge externalizes the plan → execute → validate loop so the same task that would silently break at 7 steps cleanly delivers at 30.
Core mechanic:
- DAG plans, not linear chains — workers run in parallel where dependencies allow
- Worktree isolation + auto-WIP commits between tiers — every tier writes to disk before the next branches off (this exists because we shipped memem v0.10.0 once with this exact failure: 20 min of recovery work)
- Per-module verify commands & acceptance criteria — defined at plan time so workers can't quietly lower the bar
- Multi-lens review — a separate reviewer agent + 3× self-consistency catches cross-module bugs single-module reviewers miss
- Structured failure ledger — failure modes captured as JSONL, recalled by pattern ID in the next plan
Track record: shipped memem v1.7 → v1.8.3 (7 releases) in a single day, including a 9-file anti-recursion safety fix and a persistent slice daemon — with human in the loop only at yes/modify/abort gates.
Install
Copy-paste:
claude plugin marketplace add TT-Wang/forge claude plugin install fo
60533a72e299OBSERVED · 2026-10-08Connect
Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control.
claude mcp add forge-mcp-server -- npx -y @tt-wang/[email protected]
Exposed tools (7)
3 read · 3 write · 1 destructive. Blast radius: 1 tool can delete or overwrite — an agent that can be talked into calling a tool can be talked into calling this one.
| Tool | Risk | Description |
|---|---|---|
forge_logs | read | Query the structured JSONL event stream that forge writes on every tool call throughout a run. Filter by \ |
iteration_state | destructive | Read, update, or reset the per-module retry state for a forge run. Tracks attempt count, score history, last status, last root cause from the debugger, and a stagnation flag. In v0.4.0+ state is scoped per run via \ |
memory_recall | read | Search forge |
memory_save | write | Persist a learned pattern to forge |
session_state | write | Save, load, or list orchestrator session snapshots for resumability. Lets a \ |
validate | write | Run full verification for a forge module against a specific working directory. Executes the module |
validate_plan | read | Structurally validate a forge plan JSON file before any worker spawns. Checks: required-field schema (\ |
Trust audit
SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | PASS |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | WARN |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- declared (1 observation(s))
- Network
- declared (2 observation(s))
- Shell
- declared (2 observation(s))
- Dependencies
- not all pinned
- Secrets in source
- none-found
Findings (8)
iteration_state
.prettierignore
.prettierrc.json
runId: "../../etc/passwd",
const out = handleForgeLogs({ runId: "../../etc/passwd" });runId: "../../etc/passwd",
@modelcontextprotocol/sdk, eslint, prettier
- If user says **abort** → stop entirely, do not execute
Gates applied: no_behavioural_pass.
60533a72e299full audit observations/trust-audit/mcp-server/tt-wang__forge-12.json · Report an issue / request a re-scanAudit history
Every audit this server has had. A grade with a past is a grade somebody is still checking.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-08 | 60533a72e299 | SAFE | B | 89 | first audit |
Questions
What is the Forge MCP server?
Turn Claude Code into a plan-execute-validate loop with parallel work, intelligent retry, and memory
What tools does Forge expose?
7 in total: 3 read-only, 3 that write, and 1 that can delete or overwrite (iteration_state). Every one is listed on this page with its risk.
Is Forge safe to connect to an agent?
The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean server reads B. Separately from the audit: 1 of its tools can destroy data, so scope the token you give it to what you actually need.
What credentials does Forge need?
No credential environment variables were found in its source, so it appears to need none.
How does Forge run?
It speaks stdio, so it runs as a local process your client starts. It is published on npm as @tt-wang/forge-mcp-server at 0.7.0.
How current is this page?
The grade is for one exact copy of the source (60533a72e299), read on 2026-10-08. The repository is watched and re-audited when it changes.